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_expression_flattening.h"
38 #include "glsl_types.h"
39 #include "glsl_parser_extras.h"
40 #include "../glsl/program.h"
41 #include "ir_optimization.h"
44 #include "main/mtypes.h"
45 #include "main/shaderobj.h"
46 #include "main/uniforms.h"
47 #include "program/hash_table.h"
50 #include "main/shaderapi.h"
51 #include "program/prog_instruction.h"
52 #include "program/prog_optimize.h"
53 #include "program/prog_print.h"
54 #include "program/program.h"
55 #include "program/prog_parameter.h"
56 #include "program/sampler.h"
58 #include "pipe/p_compiler.h"
59 #include "pipe/p_context.h"
60 #include "pipe/p_screen.h"
61 #include "pipe/p_shader_tokens.h"
62 #include "pipe/p_state.h"
63 #include "util/u_math.h"
64 #include "tgsi/tgsi_ureg.h"
65 #include "tgsi/tgsi_info.h"
66 #include "st_context.h"
67 #include "st_program.h"
68 #include "st_glsl_to_tgsi.h"
69 #include "st_mesa_to_tgsi.h"
72 #define PROGRAM_IMMEDIATE PROGRAM_FILE_MAX
73 #define PROGRAM_ANY_CONST ((1 << PROGRAM_STATE_VAR) | \
74 (1 << PROGRAM_CONSTANT) | \
75 (1 << PROGRAM_UNIFORM))
78 * Maximum number of temporary registers.
80 * It is too big for stack allocated arrays -- it will cause stack overflow on
81 * Windows and likely Mac OS X.
83 #define MAX_TEMPS 4096
86 * Maximum number of arrays
88 #define MAX_ARRAYS 256
90 #define MAX_GLSL_TEXTURE_OFFSET 4
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
;
112 this->type
= type
? type
->base_type
: GLSL_TYPE_ERROR
;
113 this->reladdr
= NULL
;
114 this->reladdr2
= NULL
;
115 this->has_index2
= false;
118 st_src_reg(gl_register_file file
, int index
, int type
)
124 this->swizzle
= SWIZZLE_XYZW
;
126 this->reladdr
= NULL
;
127 this->reladdr2
= NULL
;
128 this->has_index2
= false;
131 st_src_reg(gl_register_file file
, int index
, int type
, int index2D
)
136 this->index2D
= index2D
;
137 this->swizzle
= SWIZZLE_XYZW
;
139 this->reladdr
= NULL
;
140 this->reladdr2
= NULL
;
141 this->has_index2
= false;
146 this->type
= GLSL_TYPE_ERROR
;
147 this->file
= PROGRAM_UNDEFINED
;
152 this->reladdr
= NULL
;
153 this->reladdr2
= NULL
;
154 this->has_index2
= false;
157 explicit st_src_reg(st_dst_reg reg
);
159 gl_register_file file
; /**< PROGRAM_* from Mesa */
160 int index
; /**< temporary index, VERT_ATTRIB_*, VARYING_SLOT_*, etc. */
162 GLuint swizzle
; /**< SWIZZLE_XYZWONEZERO swizzles from Mesa. */
163 int negate
; /**< NEGATE_XYZW mask from mesa */
164 int type
; /** GLSL_TYPE_* from GLSL IR (enum glsl_base_type) */
165 /** Register index should be offset by the integer in this reg. */
167 st_src_reg
*reladdr2
;
173 st_dst_reg(gl_register_file file
, int writemask
, int type
, int index
)
177 this->writemask
= writemask
;
178 this->cond_mask
= COND_TR
;
179 this->reladdr
= NULL
;
183 st_dst_reg(gl_register_file file
, int writemask
, int type
)
187 this->writemask
= writemask
;
188 this->cond_mask
= COND_TR
;
189 this->reladdr
= NULL
;
195 this->type
= GLSL_TYPE_ERROR
;
196 this->file
= PROGRAM_UNDEFINED
;
199 this->cond_mask
= COND_TR
;
200 this->reladdr
= NULL
;
203 explicit st_dst_reg(st_src_reg reg
);
205 gl_register_file file
; /**< PROGRAM_* from Mesa */
206 int index
; /**< temporary index, VERT_ATTRIB_*, VARYING_SLOT_*, etc. */
207 int writemask
; /**< Bitfield of WRITEMASK_[XYZW] */
209 int type
; /** GLSL_TYPE_* from GLSL IR (enum glsl_base_type) */
210 /** Register index should be offset by the integer in this reg. */
214 st_src_reg::st_src_reg(st_dst_reg reg
)
216 this->type
= reg
.type
;
217 this->file
= reg
.file
;
218 this->index
= reg
.index
;
219 this->swizzle
= SWIZZLE_XYZW
;
221 this->reladdr
= reg
.reladdr
;
223 this->reladdr2
= NULL
;
224 this->has_index2
= false;
227 st_dst_reg::st_dst_reg(st_src_reg reg
)
229 this->type
= reg
.type
;
230 this->file
= reg
.file
;
231 this->index
= reg
.index
;
232 this->writemask
= WRITEMASK_XYZW
;
233 this->cond_mask
= COND_TR
;
234 this->reladdr
= reg
.reladdr
;
237 class glsl_to_tgsi_instruction
: public exec_node
{
239 DECLARE_RALLOC_CXX_OPERATORS(glsl_to_tgsi_instruction
)
244 /** Pointer to the ir source this tree came from for debugging */
246 GLboolean cond_update
;
248 int sampler
; /**< sampler index */
249 int tex_target
; /**< One of TEXTURE_*_INDEX */
250 GLboolean tex_shadow
;
252 st_src_reg tex_offsets
[MAX_GLSL_TEXTURE_OFFSET
];
253 unsigned tex_offset_num_offset
;
254 int dead_mask
; /**< Used in dead code elimination */
256 class function_entry
*function
; /* Set on TGSI_OPCODE_CAL or TGSI_OPCODE_BGNSUB */
259 class variable_storage
: public exec_node
{
261 variable_storage(ir_variable
*var
, gl_register_file file
, int index
)
262 : file(file
), index(index
), var(var
)
267 gl_register_file file
;
269 ir_variable
*var
; /* variable that maps to this, if any */
272 class immediate_storage
: public exec_node
{
274 immediate_storage(gl_constant_value
*values
, int size
, int type
)
276 memcpy(this->values
, values
, size
* sizeof(gl_constant_value
));
281 gl_constant_value values
[4];
282 int size
; /**< Number of components (1-4) */
283 int type
; /**< GL_FLOAT, GL_INT, GL_BOOL, or GL_UNSIGNED_INT */
286 class function_entry
: public exec_node
{
288 ir_function_signature
*sig
;
291 * identifier of this function signature used by the program.
293 * At the point that TGSI instructions for function calls are
294 * generated, we don't know the address of the first instruction of
295 * the function body. So we make the BranchTarget that is called a
296 * small integer and rewrite them during set_branchtargets().
301 * Pointer to first instruction of the function body.
303 * Set during function body emits after main() is processed.
305 glsl_to_tgsi_instruction
*bgn_inst
;
308 * Index of the first instruction of the function body in actual TGSI.
310 * Set after conversion from glsl_to_tgsi_instruction to TGSI.
314 /** Storage for the return value. */
315 st_src_reg return_reg
;
318 struct glsl_to_tgsi_visitor
: public ir_visitor
{
320 glsl_to_tgsi_visitor();
321 ~glsl_to_tgsi_visitor();
323 function_entry
*current_function
;
325 struct gl_context
*ctx
;
326 struct gl_program
*prog
;
327 struct gl_shader_program
*shader_program
;
328 struct gl_shader
*shader
;
329 struct gl_shader_compiler_options
*options
;
333 unsigned array_sizes
[MAX_ARRAYS
];
336 int num_address_regs
;
338 bool indirect_addr_consts
;
341 bool native_integers
;
344 variable_storage
*find_variable_storage(ir_variable
*var
);
346 int add_constant(gl_register_file file
, gl_constant_value values
[4],
347 int size
, int datatype
, GLuint
*swizzle_out
);
349 function_entry
*get_function_signature(ir_function_signature
*sig
);
351 st_src_reg
get_temp(const glsl_type
*type
);
352 void reladdr_to_temp(ir_instruction
*ir
, st_src_reg
*reg
, int *num_reladdr
);
354 st_src_reg
st_src_reg_for_float(float val
);
355 st_src_reg
st_src_reg_for_int(int val
);
356 st_src_reg
st_src_reg_for_type(int type
, int val
);
359 * \name Visit methods
361 * As typical for the visitor pattern, there must be one \c visit method for
362 * each concrete subclass of \c ir_instruction. Virtual base classes within
363 * the hierarchy should not have \c visit methods.
366 virtual void visit(ir_variable
*);
367 virtual void visit(ir_loop
*);
368 virtual void visit(ir_loop_jump
*);
369 virtual void visit(ir_function_signature
*);
370 virtual void visit(ir_function
*);
371 virtual void visit(ir_expression
*);
372 virtual void visit(ir_swizzle
*);
373 virtual void visit(ir_dereference_variable
*);
374 virtual void visit(ir_dereference_array
*);
375 virtual void visit(ir_dereference_record
*);
376 virtual void visit(ir_assignment
*);
377 virtual void visit(ir_constant
*);
378 virtual void visit(ir_call
*);
379 virtual void visit(ir_return
*);
380 virtual void visit(ir_discard
*);
381 virtual void visit(ir_texture
*);
382 virtual void visit(ir_if
*);
383 virtual void visit(ir_emit_vertex
*);
384 virtual void visit(ir_end_primitive
*);
389 /** List of variable_storage */
392 /** List of immediate_storage */
393 exec_list immediates
;
394 unsigned num_immediates
;
396 /** List of function_entry */
397 exec_list function_signatures
;
398 int next_signature_id
;
400 /** List of glsl_to_tgsi_instruction */
401 exec_list instructions
;
403 glsl_to_tgsi_instruction
*emit(ir_instruction
*ir
, unsigned op
);
405 glsl_to_tgsi_instruction
*emit(ir_instruction
*ir
, unsigned op
,
406 st_dst_reg dst
, st_src_reg src0
);
408 glsl_to_tgsi_instruction
*emit(ir_instruction
*ir
, unsigned op
,
409 st_dst_reg dst
, st_src_reg src0
, st_src_reg src1
);
411 glsl_to_tgsi_instruction
*emit(ir_instruction
*ir
, unsigned op
,
413 st_src_reg src0
, st_src_reg src1
, st_src_reg src2
);
415 glsl_to_tgsi_instruction
*emit(ir_instruction
*ir
, unsigned op
,
417 st_src_reg src0
, st_src_reg src1
,
418 st_src_reg src2
, st_src_reg src3
);
420 unsigned get_opcode(ir_instruction
*ir
, unsigned op
,
422 st_src_reg src0
, st_src_reg src1
);
425 * Emit the correct dot-product instruction for the type of arguments
427 glsl_to_tgsi_instruction
*emit_dp(ir_instruction
*ir
,
433 void emit_scalar(ir_instruction
*ir
, unsigned op
,
434 st_dst_reg dst
, st_src_reg src0
);
436 void emit_scalar(ir_instruction
*ir
, unsigned op
,
437 st_dst_reg dst
, st_src_reg src0
, st_src_reg src1
);
439 void emit_arl(ir_instruction
*ir
, st_dst_reg dst
, st_src_reg src0
);
441 void emit_scs(ir_instruction
*ir
, unsigned op
,
442 st_dst_reg dst
, const st_src_reg
&src
);
444 bool try_emit_mad(ir_expression
*ir
,
446 bool try_emit_mad_for_and_not(ir_expression
*ir
,
448 bool try_emit_sat(ir_expression
*ir
);
450 void emit_swz(ir_expression
*ir
);
452 bool process_move_condition(ir_rvalue
*ir
);
454 void simplify_cmp(void);
456 void rename_temp_register(int index
, int new_index
);
457 int get_first_temp_read(int index
);
458 int get_first_temp_write(int index
);
459 int get_last_temp_read(int index
);
460 int get_last_temp_write(int index
);
462 void copy_propagate(void);
463 int eliminate_dead_code(void);
464 void merge_registers(void);
465 void renumber_registers(void);
467 void emit_block_mov(ir_assignment
*ir
, const struct glsl_type
*type
,
468 st_dst_reg
*l
, st_src_reg
*r
);
473 static st_src_reg undef_src
= st_src_reg(PROGRAM_UNDEFINED
, 0, GLSL_TYPE_ERROR
);
475 static st_dst_reg undef_dst
= st_dst_reg(PROGRAM_UNDEFINED
, SWIZZLE_NOOP
, GLSL_TYPE_ERROR
);
477 static st_dst_reg address_reg
= st_dst_reg(PROGRAM_ADDRESS
, WRITEMASK_X
, GLSL_TYPE_FLOAT
, 0);
478 static st_dst_reg address_reg2
= st_dst_reg(PROGRAM_ADDRESS
, WRITEMASK_X
, GLSL_TYPE_FLOAT
, 1);
481 fail_link(struct gl_shader_program
*prog
, const char *fmt
, ...) PRINTFLIKE(2, 3);
484 fail_link(struct gl_shader_program
*prog
, const char *fmt
, ...)
488 ralloc_vasprintf_append(&prog
->InfoLog
, fmt
, args
);
491 prog
->LinkStatus
= GL_FALSE
;
495 swizzle_for_size(int size
)
497 int size_swizzles
[4] = {
498 MAKE_SWIZZLE4(SWIZZLE_X
, SWIZZLE_X
, SWIZZLE_X
, SWIZZLE_X
),
499 MAKE_SWIZZLE4(SWIZZLE_X
, SWIZZLE_Y
, SWIZZLE_Y
, SWIZZLE_Y
),
500 MAKE_SWIZZLE4(SWIZZLE_X
, SWIZZLE_Y
, SWIZZLE_Z
, SWIZZLE_Z
),
501 MAKE_SWIZZLE4(SWIZZLE_X
, SWIZZLE_Y
, SWIZZLE_Z
, SWIZZLE_W
),
504 assert((size
>= 1) && (size
<= 4));
505 return size_swizzles
[size
- 1];
509 is_tex_instruction(unsigned opcode
)
511 const tgsi_opcode_info
* info
= tgsi_get_opcode_info(opcode
);
516 num_inst_dst_regs(unsigned opcode
)
518 const tgsi_opcode_info
* info
= tgsi_get_opcode_info(opcode
);
519 return info
->num_dst
;
523 num_inst_src_regs(unsigned opcode
)
525 const tgsi_opcode_info
* info
= tgsi_get_opcode_info(opcode
);
526 return info
->is_tex
? info
->num_src
- 1 : info
->num_src
;
529 glsl_to_tgsi_instruction
*
530 glsl_to_tgsi_visitor::emit(ir_instruction
*ir
, unsigned op
,
532 st_src_reg src0
, st_src_reg src1
,
533 st_src_reg src2
, st_src_reg src3
)
535 glsl_to_tgsi_instruction
*inst
= new(mem_ctx
) glsl_to_tgsi_instruction();
536 int num_reladdr
= 0, i
;
538 op
= get_opcode(ir
, op
, dst
, src0
, src1
);
540 /* If we have to do relative addressing, we want to load the ARL
541 * reg directly for one of the regs, and preload the other reladdr
542 * sources into temps.
544 num_reladdr
+= dst
.reladdr
!= NULL
;
545 num_reladdr
+= src0
.reladdr
!= NULL
|| src0
.reladdr2
!= NULL
;
546 num_reladdr
+= src1
.reladdr
!= NULL
|| src1
.reladdr2
!= NULL
;
547 num_reladdr
+= src2
.reladdr
!= NULL
|| src2
.reladdr2
!= NULL
;
548 num_reladdr
+= src3
.reladdr
!= NULL
|| src3
.reladdr2
!= NULL
;
550 reladdr_to_temp(ir
, &src3
, &num_reladdr
);
551 reladdr_to_temp(ir
, &src2
, &num_reladdr
);
552 reladdr_to_temp(ir
, &src1
, &num_reladdr
);
553 reladdr_to_temp(ir
, &src0
, &num_reladdr
);
556 emit_arl(ir
, address_reg
, *dst
.reladdr
);
559 assert(num_reladdr
== 0);
570 inst
->function
= NULL
;
572 /* Update indirect addressing status used by TGSI */
575 case PROGRAM_STATE_VAR
:
576 case PROGRAM_CONSTANT
:
577 case PROGRAM_UNIFORM
:
578 this->indirect_addr_consts
= true;
580 case PROGRAM_IMMEDIATE
:
581 assert(!"immediates should not have indirect addressing");
588 for (i
=0; i
<4; i
++) {
589 if(inst
->src
[i
].reladdr
) {
590 switch(inst
->src
[i
].file
) {
591 case PROGRAM_STATE_VAR
:
592 case PROGRAM_CONSTANT
:
593 case PROGRAM_UNIFORM
:
594 this->indirect_addr_consts
= true;
596 case PROGRAM_IMMEDIATE
:
597 assert(!"immediates should not have indirect addressing");
606 this->instructions
.push_tail(inst
);
611 glsl_to_tgsi_instruction
*
612 glsl_to_tgsi_visitor::emit(ir_instruction
*ir
, unsigned op
,
613 st_dst_reg dst
, st_src_reg src0
,
614 st_src_reg src1
, st_src_reg src2
)
616 return emit(ir
, op
, dst
, src0
, src1
, src2
, undef_src
);
619 glsl_to_tgsi_instruction
*
620 glsl_to_tgsi_visitor::emit(ir_instruction
*ir
, unsigned op
,
621 st_dst_reg dst
, st_src_reg src0
, st_src_reg src1
)
623 return emit(ir
, op
, dst
, src0
, src1
, undef_src
, undef_src
);
626 glsl_to_tgsi_instruction
*
627 glsl_to_tgsi_visitor::emit(ir_instruction
*ir
, unsigned op
,
628 st_dst_reg dst
, st_src_reg src0
)
630 assert(dst
.writemask
!= 0);
631 return emit(ir
, op
, dst
, src0
, undef_src
, undef_src
, undef_src
);
634 glsl_to_tgsi_instruction
*
635 glsl_to_tgsi_visitor::emit(ir_instruction
*ir
, unsigned op
)
637 return emit(ir
, op
, undef_dst
, undef_src
, undef_src
, undef_src
, undef_src
);
641 * Determines whether to use an integer, unsigned integer, or float opcode
642 * based on the operands and input opcode, then emits the result.
645 glsl_to_tgsi_visitor::get_opcode(ir_instruction
*ir
, unsigned op
,
647 st_src_reg src0
, st_src_reg src1
)
649 int type
= GLSL_TYPE_FLOAT
;
651 if (op
== TGSI_OPCODE_MOV
)
654 assert(src0
.type
!= GLSL_TYPE_ARRAY
);
655 assert(src0
.type
!= GLSL_TYPE_STRUCT
);
656 assert(src1
.type
!= GLSL_TYPE_ARRAY
);
657 assert(src1
.type
!= GLSL_TYPE_STRUCT
);
659 if (src0
.type
== GLSL_TYPE_FLOAT
|| src1
.type
== GLSL_TYPE_FLOAT
)
660 type
= GLSL_TYPE_FLOAT
;
661 else if (native_integers
)
662 type
= src0
.type
== GLSL_TYPE_BOOL
? GLSL_TYPE_INT
: src0
.type
;
664 #define case4(c, f, i, u) \
665 case TGSI_OPCODE_##c: \
666 if (type == GLSL_TYPE_INT) \
667 op = TGSI_OPCODE_##i; \
668 else if (type == GLSL_TYPE_UINT) \
669 op = TGSI_OPCODE_##u; \
671 op = TGSI_OPCODE_##f; \
674 #define case3(f, i, u) case4(f, f, i, u)
675 #define case2fi(f, i) case4(f, f, i, i)
676 #define case2iu(i, u) case4(i, LAST, i, u)
678 #define casecomp(c, f, i, u) \
679 case TGSI_OPCODE_##c: \
680 if (type == GLSL_TYPE_INT) \
681 op = TGSI_OPCODE_##i; \
682 else if (type == GLSL_TYPE_UINT) \
683 op = TGSI_OPCODE_##u; \
684 else if (native_integers) \
685 op = TGSI_OPCODE_##f; \
687 op = TGSI_OPCODE_##c; \
694 case3(DIV
, IDIV
, UDIV
);
695 case3(MAX
, IMAX
, UMAX
);
696 case3(MIN
, IMIN
, UMIN
);
699 casecomp(SEQ
, FSEQ
, USEQ
, USEQ
);
700 casecomp(SNE
, FSNE
, USNE
, USNE
);
701 casecomp(SGE
, FSGE
, ISGE
, USGE
);
702 casecomp(SLT
, FSLT
, ISLT
, USLT
);
707 case3(ABS
, IABS
, IABS
);
711 case2iu(IMUL_HI
, UMUL_HI
);
715 assert(op
!= TGSI_OPCODE_LAST
);
719 glsl_to_tgsi_instruction
*
720 glsl_to_tgsi_visitor::emit_dp(ir_instruction
*ir
,
721 st_dst_reg dst
, st_src_reg src0
, st_src_reg src1
,
724 static const unsigned dot_opcodes
[] = {
725 TGSI_OPCODE_DP2
, TGSI_OPCODE_DP3
, TGSI_OPCODE_DP4
728 return emit(ir
, dot_opcodes
[elements
- 2], dst
, src0
, src1
);
732 * Emits TGSI scalar opcodes to produce unique answers across channels.
734 * Some TGSI opcodes are scalar-only, like ARB_fp/vp. The src X
735 * channel determines the result across all channels. So to do a vec4
736 * of this operation, we want to emit a scalar per source channel used
737 * to produce dest channels.
740 glsl_to_tgsi_visitor::emit_scalar(ir_instruction
*ir
, unsigned op
,
742 st_src_reg orig_src0
, st_src_reg orig_src1
)
745 int done_mask
= ~dst
.writemask
;
747 /* TGSI RCP is a scalar operation splatting results to all channels,
748 * like ARB_fp/vp. So emit as many RCPs as necessary to cover our
751 for (i
= 0; i
< 4; i
++) {
752 GLuint this_mask
= (1 << i
);
753 glsl_to_tgsi_instruction
*inst
;
754 st_src_reg src0
= orig_src0
;
755 st_src_reg src1
= orig_src1
;
757 if (done_mask
& this_mask
)
760 GLuint src0_swiz
= GET_SWZ(src0
.swizzle
, i
);
761 GLuint src1_swiz
= GET_SWZ(src1
.swizzle
, i
);
762 for (j
= i
+ 1; j
< 4; j
++) {
763 /* If there is another enabled component in the destination that is
764 * derived from the same inputs, generate its value on this pass as
767 if (!(done_mask
& (1 << j
)) &&
768 GET_SWZ(src0
.swizzle
, j
) == src0_swiz
&&
769 GET_SWZ(src1
.swizzle
, j
) == src1_swiz
) {
770 this_mask
|= (1 << j
);
773 src0
.swizzle
= MAKE_SWIZZLE4(src0_swiz
, src0_swiz
,
774 src0_swiz
, src0_swiz
);
775 src1
.swizzle
= MAKE_SWIZZLE4(src1_swiz
, src1_swiz
,
776 src1_swiz
, src1_swiz
);
778 inst
= emit(ir
, op
, dst
, src0
, src1
);
779 inst
->dst
.writemask
= this_mask
;
780 done_mask
|= this_mask
;
785 glsl_to_tgsi_visitor::emit_scalar(ir_instruction
*ir
, unsigned op
,
786 st_dst_reg dst
, st_src_reg src0
)
788 st_src_reg undef
= undef_src
;
790 undef
.swizzle
= SWIZZLE_XXXX
;
792 emit_scalar(ir
, op
, dst
, src0
, undef
);
796 glsl_to_tgsi_visitor::emit_arl(ir_instruction
*ir
,
797 st_dst_reg dst
, st_src_reg src0
)
799 int op
= TGSI_OPCODE_ARL
;
801 if (src0
.type
== GLSL_TYPE_INT
|| src0
.type
== GLSL_TYPE_UINT
)
802 op
= TGSI_OPCODE_UARL
;
804 assert(dst
.file
== PROGRAM_ADDRESS
);
805 if (dst
.index
>= this->num_address_regs
)
806 this->num_address_regs
= dst
.index
+ 1;
808 emit(NULL
, op
, dst
, src0
);
812 * Emit an TGSI_OPCODE_SCS instruction
814 * The \c SCS opcode functions a bit differently than the other TGSI opcodes.
815 * Instead of splatting its result across all four components of the
816 * destination, it writes one value to the \c x component and another value to
817 * the \c y component.
819 * \param ir IR instruction being processed
820 * \param op Either \c TGSI_OPCODE_SIN or \c TGSI_OPCODE_COS depending
821 * on which value is desired.
822 * \param dst Destination register
823 * \param src Source register
826 glsl_to_tgsi_visitor::emit_scs(ir_instruction
*ir
, unsigned op
,
828 const st_src_reg
&src
)
830 /* Vertex programs cannot use the SCS opcode.
832 if (this->prog
->Target
== GL_VERTEX_PROGRAM_ARB
) {
833 emit_scalar(ir
, op
, dst
, src
);
837 const unsigned component
= (op
== TGSI_OPCODE_SIN
) ? 0 : 1;
838 const unsigned scs_mask
= (1U << component
);
839 int done_mask
= ~dst
.writemask
;
842 assert(op
== TGSI_OPCODE_SIN
|| op
== TGSI_OPCODE_COS
);
844 /* If there are compnents in the destination that differ from the component
845 * that will be written by the SCS instrution, we'll need a temporary.
847 if (scs_mask
!= unsigned(dst
.writemask
)) {
848 tmp
= get_temp(glsl_type::vec4_type
);
851 for (unsigned i
= 0; i
< 4; i
++) {
852 unsigned this_mask
= (1U << i
);
853 st_src_reg src0
= src
;
855 if ((done_mask
& this_mask
) != 0)
858 /* The source swizzle specified which component of the source generates
859 * sine / cosine for the current component in the destination. The SCS
860 * instruction requires that this value be swizzle to the X component.
861 * Replace the current swizzle with a swizzle that puts the source in
864 unsigned src0_swiz
= GET_SWZ(src
.swizzle
, i
);
866 src0
.swizzle
= MAKE_SWIZZLE4(src0_swiz
, src0_swiz
,
867 src0_swiz
, src0_swiz
);
868 for (unsigned j
= i
+ 1; j
< 4; j
++) {
869 /* If there is another enabled component in the destination that is
870 * derived from the same inputs, generate its value on this pass as
873 if (!(done_mask
& (1 << j
)) &&
874 GET_SWZ(src0
.swizzle
, j
) == src0_swiz
) {
875 this_mask
|= (1 << j
);
879 if (this_mask
!= scs_mask
) {
880 glsl_to_tgsi_instruction
*inst
;
881 st_dst_reg tmp_dst
= st_dst_reg(tmp
);
883 /* Emit the SCS instruction.
885 inst
= emit(ir
, TGSI_OPCODE_SCS
, tmp_dst
, src0
);
886 inst
->dst
.writemask
= scs_mask
;
888 /* Move the result of the SCS instruction to the desired location in
891 tmp
.swizzle
= MAKE_SWIZZLE4(component
, component
,
892 component
, component
);
893 inst
= emit(ir
, TGSI_OPCODE_SCS
, dst
, tmp
);
894 inst
->dst
.writemask
= this_mask
;
896 /* Emit the SCS instruction to write directly to the destination.
898 glsl_to_tgsi_instruction
*inst
= emit(ir
, TGSI_OPCODE_SCS
, dst
, src0
);
899 inst
->dst
.writemask
= scs_mask
;
902 done_mask
|= this_mask
;
907 glsl_to_tgsi_visitor::add_constant(gl_register_file file
,
908 gl_constant_value values
[4], int size
, int datatype
,
911 if (file
== PROGRAM_CONSTANT
) {
912 return _mesa_add_typed_unnamed_constant(this->prog
->Parameters
, values
,
913 size
, datatype
, swizzle_out
);
916 immediate_storage
*entry
;
917 assert(file
== PROGRAM_IMMEDIATE
);
919 /* Search immediate storage to see if we already have an identical
920 * immediate that we can use instead of adding a duplicate entry.
922 foreach_in_list(immediate_storage
, entry
, &this->immediates
) {
923 if (entry
->size
== size
&&
924 entry
->type
== datatype
&&
925 !memcmp(entry
->values
, values
, size
* sizeof(gl_constant_value
))) {
931 /* Add this immediate to the list. */
932 entry
= new(mem_ctx
) immediate_storage(values
, size
, datatype
);
933 this->immediates
.push_tail(entry
);
934 this->num_immediates
++;
940 glsl_to_tgsi_visitor::st_src_reg_for_float(float val
)
942 st_src_reg
src(PROGRAM_IMMEDIATE
, -1, GLSL_TYPE_FLOAT
);
943 union gl_constant_value uval
;
946 src
.index
= add_constant(src
.file
, &uval
, 1, GL_FLOAT
, &src
.swizzle
);
952 glsl_to_tgsi_visitor::st_src_reg_for_int(int val
)
954 st_src_reg
src(PROGRAM_IMMEDIATE
, -1, GLSL_TYPE_INT
);
955 union gl_constant_value uval
;
957 assert(native_integers
);
960 src
.index
= add_constant(src
.file
, &uval
, 1, GL_INT
, &src
.swizzle
);
966 glsl_to_tgsi_visitor::st_src_reg_for_type(int type
, int val
)
969 return type
== GLSL_TYPE_FLOAT
? st_src_reg_for_float(val
) :
970 st_src_reg_for_int(val
);
972 return st_src_reg_for_float(val
);
976 type_size(const struct glsl_type
*type
)
981 switch (type
->base_type
) {
984 case GLSL_TYPE_FLOAT
:
986 if (type
->is_matrix()) {
987 return type
->matrix_columns
;
989 /* Regardless of size of vector, it gets a vec4. This is bad
990 * packing for things like floats, but otherwise arrays become a
991 * mess. Hopefully a later pass over the code can pack scalars
992 * down if appropriate.
996 case GLSL_TYPE_ARRAY
:
997 assert(type
->length
> 0);
998 return type_size(type
->fields
.array
) * type
->length
;
999 case GLSL_TYPE_STRUCT
:
1001 for (i
= 0; i
< type
->length
; i
++) {
1002 size
+= type_size(type
->fields
.structure
[i
].type
);
1005 case GLSL_TYPE_SAMPLER
:
1006 case GLSL_TYPE_IMAGE
:
1007 /* Samplers take up one slot in UNIFORMS[], but they're baked in
1011 case GLSL_TYPE_ATOMIC_UINT
:
1012 case GLSL_TYPE_INTERFACE
:
1013 case GLSL_TYPE_VOID
:
1014 case GLSL_TYPE_ERROR
:
1015 assert(!"Invalid type in type_size");
1022 * In the initial pass of codegen, we assign temporary numbers to
1023 * intermediate results. (not SSA -- variable assignments will reuse
1027 glsl_to_tgsi_visitor::get_temp(const glsl_type
*type
)
1031 src
.type
= native_integers
? type
->base_type
: GLSL_TYPE_FLOAT
;
1035 if (!options
->EmitNoIndirectTemp
&&
1036 (type
->is_array() || type
->is_matrix())) {
1038 src
.file
= PROGRAM_ARRAY
;
1039 src
.index
= next_array
<< 16 | 0x8000;
1040 array_sizes
[next_array
] = type_size(type
);
1044 src
.file
= PROGRAM_TEMPORARY
;
1045 src
.index
= next_temp
;
1046 next_temp
+= type_size(type
);
1049 if (type
->is_array() || type
->is_record()) {
1050 src
.swizzle
= SWIZZLE_NOOP
;
1052 src
.swizzle
= swizzle_for_size(type
->vector_elements
);
1059 glsl_to_tgsi_visitor::find_variable_storage(ir_variable
*var
)
1062 foreach_in_list(variable_storage
, entry
, &this->variables
) {
1063 if (entry
->var
== var
)
1071 glsl_to_tgsi_visitor::visit(ir_variable
*ir
)
1073 if (strcmp(ir
->name
, "gl_FragCoord") == 0) {
1074 struct gl_fragment_program
*fp
= (struct gl_fragment_program
*)this->prog
;
1076 fp
->OriginUpperLeft
= ir
->data
.origin_upper_left
;
1077 fp
->PixelCenterInteger
= ir
->data
.pixel_center_integer
;
1080 if (ir
->data
.mode
== ir_var_uniform
&& strncmp(ir
->name
, "gl_", 3) == 0) {
1082 const ir_state_slot
*const slots
= ir
->state_slots
;
1083 assert(ir
->state_slots
!= NULL
);
1085 /* Check if this statevar's setup in the STATE file exactly
1086 * matches how we'll want to reference it as a
1087 * struct/array/whatever. If not, then we need to move it into
1088 * temporary storage and hope that it'll get copy-propagated
1091 for (i
= 0; i
< ir
->num_state_slots
; i
++) {
1092 if (slots
[i
].swizzle
!= SWIZZLE_XYZW
) {
1097 variable_storage
*storage
;
1099 if (i
== ir
->num_state_slots
) {
1100 /* We'll set the index later. */
1101 storage
= new(mem_ctx
) variable_storage(ir
, PROGRAM_STATE_VAR
, -1);
1102 this->variables
.push_tail(storage
);
1106 /* The variable_storage constructor allocates slots based on the size
1107 * of the type. However, this had better match the number of state
1108 * elements that we're going to copy into the new temporary.
1110 assert((int) ir
->num_state_slots
== type_size(ir
->type
));
1112 dst
= st_dst_reg(get_temp(ir
->type
));
1114 storage
= new(mem_ctx
) variable_storage(ir
, dst
.file
, dst
.index
);
1116 this->variables
.push_tail(storage
);
1120 for (unsigned int i
= 0; i
< ir
->num_state_slots
; i
++) {
1121 int index
= _mesa_add_state_reference(this->prog
->Parameters
,
1122 (gl_state_index
*)slots
[i
].tokens
);
1124 if (storage
->file
== PROGRAM_STATE_VAR
) {
1125 if (storage
->index
== -1) {
1126 storage
->index
= index
;
1128 assert(index
== storage
->index
+ (int)i
);
1131 /* We use GLSL_TYPE_FLOAT here regardless of the actual type of
1132 * the data being moved since MOV does not care about the type of
1133 * data it is moving, and we don't want to declare registers with
1134 * array or struct types.
1136 st_src_reg
src(PROGRAM_STATE_VAR
, index
, GLSL_TYPE_FLOAT
);
1137 src
.swizzle
= slots
[i
].swizzle
;
1138 emit(ir
, TGSI_OPCODE_MOV
, dst
, src
);
1139 /* even a float takes up a whole vec4 reg in a struct/array. */
1144 if (storage
->file
== PROGRAM_TEMPORARY
&&
1145 dst
.index
!= storage
->index
+ (int) ir
->num_state_slots
) {
1146 fail_link(this->shader_program
,
1147 "failed to load builtin uniform `%s' (%d/%d regs loaded)\n",
1148 ir
->name
, dst
.index
- storage
->index
,
1149 type_size(ir
->type
));
1155 glsl_to_tgsi_visitor::visit(ir_loop
*ir
)
1157 emit(NULL
, TGSI_OPCODE_BGNLOOP
);
1159 visit_exec_list(&ir
->body_instructions
, this);
1161 emit(NULL
, TGSI_OPCODE_ENDLOOP
);
1165 glsl_to_tgsi_visitor::visit(ir_loop_jump
*ir
)
1168 case ir_loop_jump::jump_break
:
1169 emit(NULL
, TGSI_OPCODE_BRK
);
1171 case ir_loop_jump::jump_continue
:
1172 emit(NULL
, TGSI_OPCODE_CONT
);
1179 glsl_to_tgsi_visitor::visit(ir_function_signature
*ir
)
1186 glsl_to_tgsi_visitor::visit(ir_function
*ir
)
1188 /* Ignore function bodies other than main() -- we shouldn't see calls to
1189 * them since they should all be inlined before we get to glsl_to_tgsi.
1191 if (strcmp(ir
->name
, "main") == 0) {
1192 const ir_function_signature
*sig
;
1195 sig
= ir
->matching_signature(NULL
, &empty
);
1199 foreach_in_list(ir_instruction
, ir
, &sig
->body
) {
1206 glsl_to_tgsi_visitor::try_emit_mad(ir_expression
*ir
, int mul_operand
)
1208 int nonmul_operand
= 1 - mul_operand
;
1210 st_dst_reg result_dst
;
1212 ir_expression
*expr
= ir
->operands
[mul_operand
]->as_expression();
1213 if (!expr
|| expr
->operation
!= ir_binop_mul
)
1216 expr
->operands
[0]->accept(this);
1218 expr
->operands
[1]->accept(this);
1220 ir
->operands
[nonmul_operand
]->accept(this);
1223 this->result
= get_temp(ir
->type
);
1224 result_dst
= st_dst_reg(this->result
);
1225 result_dst
.writemask
= (1 << ir
->type
->vector_elements
) - 1;
1226 emit(ir
, TGSI_OPCODE_MAD
, result_dst
, a
, b
, c
);
1232 * Emit MAD(a, -b, a) instead of AND(a, NOT(b))
1234 * The logic values are 1.0 for true and 0.0 for false. Logical-and is
1235 * implemented using multiplication, and logical-or is implemented using
1236 * addition. Logical-not can be implemented as (true - x), or (1.0 - x).
1237 * As result, the logical expression (a & !b) can be rewritten as:
1241 * - (a * 1) - (a * b)
1245 * This final expression can be implemented as a single MAD(a, -b, a)
1249 glsl_to_tgsi_visitor::try_emit_mad_for_and_not(ir_expression
*ir
, int try_operand
)
1251 const int other_operand
= 1 - try_operand
;
1254 ir_expression
*expr
= ir
->operands
[try_operand
]->as_expression();
1255 if (!expr
|| expr
->operation
!= ir_unop_logic_not
)
1258 ir
->operands
[other_operand
]->accept(this);
1260 expr
->operands
[0]->accept(this);
1263 b
.negate
= ~b
.negate
;
1265 this->result
= get_temp(ir
->type
);
1266 emit(ir
, TGSI_OPCODE_MAD
, st_dst_reg(this->result
), a
, b
, a
);
1272 glsl_to_tgsi_visitor::try_emit_sat(ir_expression
*ir
)
1274 /* Emit saturates in the vertex shader only if SM 3.0 is supported.
1276 if (this->prog
->Target
== GL_VERTEX_PROGRAM_ARB
&&
1277 !st_context(this->ctx
)->has_shader_model3
) {
1281 ir_rvalue
*sat_src
= ir
->as_rvalue_to_saturate();
1285 sat_src
->accept(this);
1286 st_src_reg src
= this->result
;
1288 /* If we generated an expression instruction into a temporary in
1289 * processing the saturate's operand, apply the saturate to that
1290 * instruction. Otherwise, generate a MOV to do the saturate.
1292 * Note that we have to be careful to only do this optimization if
1293 * the instruction in question was what generated src->result. For
1294 * example, ir_dereference_array might generate a MUL instruction
1295 * to create the reladdr, and return us a src reg using that
1296 * reladdr. That MUL result is not the value we're trying to
1299 ir_expression
*sat_src_expr
= sat_src
->as_expression();
1300 if (sat_src_expr
&& (sat_src_expr
->operation
== ir_binop_mul
||
1301 sat_src_expr
->operation
== ir_binop_add
||
1302 sat_src_expr
->operation
== ir_binop_dot
)) {
1303 glsl_to_tgsi_instruction
*new_inst
;
1304 new_inst
= (glsl_to_tgsi_instruction
*)this->instructions
.get_tail();
1305 new_inst
->saturate
= true;
1307 this->result
= get_temp(ir
->type
);
1308 st_dst_reg result_dst
= st_dst_reg(this->result
);
1309 result_dst
.writemask
= (1 << ir
->type
->vector_elements
) - 1;
1310 glsl_to_tgsi_instruction
*inst
;
1311 inst
= emit(ir
, TGSI_OPCODE_MOV
, result_dst
, src
);
1312 inst
->saturate
= true;
1319 glsl_to_tgsi_visitor::reladdr_to_temp(ir_instruction
*ir
,
1320 st_src_reg
*reg
, int *num_reladdr
)
1322 if (!reg
->reladdr
&& !reg
->reladdr2
)
1325 if (reg
->reladdr
) emit_arl(ir
, address_reg
, *reg
->reladdr
);
1326 if (reg
->reladdr2
) emit_arl(ir
, address_reg2
, *reg
->reladdr2
);
1328 if (*num_reladdr
!= 1) {
1329 st_src_reg temp
= get_temp(glsl_type::vec4_type
);
1331 emit(ir
, TGSI_OPCODE_MOV
, st_dst_reg(temp
), *reg
);
1339 glsl_to_tgsi_visitor::visit(ir_expression
*ir
)
1341 unsigned int operand
;
1342 st_src_reg op
[Elements(ir
->operands
)];
1343 st_src_reg result_src
;
1344 st_dst_reg result_dst
;
1346 /* Quick peephole: Emit MAD(a, b, c) instead of ADD(MUL(a, b), c)
1348 if (ir
->operation
== ir_binop_add
) {
1349 if (try_emit_mad(ir
, 1))
1351 if (try_emit_mad(ir
, 0))
1355 /* Quick peephole: Emit OPCODE_MAD(-a, -b, a) instead of AND(a, NOT(b))
1357 if (!native_integers
&& ir
->operation
== ir_binop_logic_and
) {
1358 if (try_emit_mad_for_and_not(ir
, 1))
1360 if (try_emit_mad_for_and_not(ir
, 0))
1364 if (try_emit_sat(ir
))
1367 if (ir
->operation
== ir_quadop_vector
)
1368 assert(!"ir_quadop_vector should have been lowered");
1370 for (operand
= 0; operand
< ir
->get_num_operands(); operand
++) {
1371 this->result
.file
= PROGRAM_UNDEFINED
;
1372 ir
->operands
[operand
]->accept(this);
1373 if (this->result
.file
== PROGRAM_UNDEFINED
) {
1374 printf("Failed to get tree for expression operand:\n");
1375 ir
->operands
[operand
]->print();
1379 op
[operand
] = this->result
;
1381 /* Matrix expression operands should have been broken down to vector
1382 * operations already.
1384 assert(!ir
->operands
[operand
]->type
->is_matrix());
1387 int vector_elements
= ir
->operands
[0]->type
->vector_elements
;
1388 if (ir
->operands
[1]) {
1389 vector_elements
= MAX2(vector_elements
,
1390 ir
->operands
[1]->type
->vector_elements
);
1393 this->result
.file
= PROGRAM_UNDEFINED
;
1395 /* Storage for our result. Ideally for an assignment we'd be using
1396 * the actual storage for the result here, instead.
1398 result_src
= get_temp(ir
->type
);
1399 /* convenience for the emit functions below. */
1400 result_dst
= st_dst_reg(result_src
);
1401 /* Limit writes to the channels that will be used by result_src later.
1402 * This does limit this temp's use as a temporary for multi-instruction
1405 result_dst
.writemask
= (1 << ir
->type
->vector_elements
) - 1;
1407 switch (ir
->operation
) {
1408 case ir_unop_logic_not
:
1409 if (result_dst
.type
!= GLSL_TYPE_FLOAT
)
1410 emit(ir
, TGSI_OPCODE_NOT
, result_dst
, op
[0]);
1412 /* Previously 'SEQ dst, src, 0.0' was used for this. However, many
1413 * older GPUs implement SEQ using multiple instructions (i915 uses two
1414 * SGE instructions and a MUL instruction). Since our logic values are
1415 * 0.0 and 1.0, 1-x also implements !x.
1417 op
[0].negate
= ~op
[0].negate
;
1418 emit(ir
, TGSI_OPCODE_ADD
, result_dst
, op
[0], st_src_reg_for_float(1.0));
1422 if (result_dst
.type
== GLSL_TYPE_INT
|| result_dst
.type
== GLSL_TYPE_UINT
)
1423 emit(ir
, TGSI_OPCODE_INEG
, result_dst
, op
[0]);
1425 op
[0].negate
= ~op
[0].negate
;
1430 emit(ir
, TGSI_OPCODE_ABS
, result_dst
, op
[0]);
1433 emit(ir
, TGSI_OPCODE_SSG
, result_dst
, op
[0]);
1436 emit_scalar(ir
, TGSI_OPCODE_RCP
, result_dst
, op
[0]);
1440 emit_scalar(ir
, TGSI_OPCODE_EX2
, result_dst
, op
[0]);
1444 assert(!"not reached: should be handled by ir_explog_to_explog2");
1447 emit_scalar(ir
, TGSI_OPCODE_LG2
, result_dst
, op
[0]);
1450 emit_scalar(ir
, TGSI_OPCODE_SIN
, result_dst
, op
[0]);
1453 emit_scalar(ir
, TGSI_OPCODE_COS
, result_dst
, op
[0]);
1455 case ir_unop_sin_reduced
:
1456 emit_scs(ir
, TGSI_OPCODE_SIN
, result_dst
, op
[0]);
1458 case ir_unop_cos_reduced
:
1459 emit_scs(ir
, TGSI_OPCODE_COS
, result_dst
, op
[0]);
1463 emit(ir
, TGSI_OPCODE_DDX
, result_dst
, op
[0]);
1467 /* The X component contains 1 or -1 depending on whether the framebuffer
1468 * is a FBO or the window system buffer, respectively.
1469 * It is then multiplied with the source operand of DDY.
1471 static const gl_state_index transform_y_state
[STATE_LENGTH
]
1472 = { STATE_INTERNAL
, STATE_FB_WPOS_Y_TRANSFORM
};
1474 unsigned transform_y_index
=
1475 _mesa_add_state_reference(this->prog
->Parameters
,
1478 st_src_reg transform_y
= st_src_reg(PROGRAM_STATE_VAR
,
1480 glsl_type::vec4_type
);
1481 transform_y
.swizzle
= SWIZZLE_XXXX
;
1483 st_src_reg temp
= get_temp(glsl_type::vec4_type
);
1485 emit(ir
, TGSI_OPCODE_MUL
, st_dst_reg(temp
), transform_y
, op
[0]);
1486 emit(ir
, TGSI_OPCODE_DDY
, result_dst
, temp
);
1490 case ir_unop_noise
: {
1491 /* At some point, a motivated person could add a better
1492 * implementation of noise. Currently not even the nvidia
1493 * binary drivers do anything more than this. In any case, the
1494 * place to do this is in the GL state tracker, not the poor
1497 emit(ir
, TGSI_OPCODE_MOV
, result_dst
, st_src_reg_for_float(0.5));
1502 emit(ir
, TGSI_OPCODE_ADD
, result_dst
, op
[0], op
[1]);
1505 emit(ir
, TGSI_OPCODE_SUB
, result_dst
, op
[0], op
[1]);
1509 emit(ir
, TGSI_OPCODE_MUL
, result_dst
, op
[0], op
[1]);
1512 if (result_dst
.type
== GLSL_TYPE_FLOAT
)
1513 assert(!"not reached: should be handled by ir_div_to_mul_rcp");
1515 emit(ir
, TGSI_OPCODE_DIV
, result_dst
, op
[0], op
[1]);
1518 if (result_dst
.type
== GLSL_TYPE_FLOAT
)
1519 assert(!"ir_binop_mod should have been converted to b * fract(a/b)");
1521 emit(ir
, TGSI_OPCODE_MOD
, result_dst
, op
[0], op
[1]);
1525 emit(ir
, TGSI_OPCODE_SLT
, result_dst
, op
[0], op
[1]);
1527 case ir_binop_greater
:
1528 emit(ir
, TGSI_OPCODE_SLT
, result_dst
, op
[1], op
[0]);
1530 case ir_binop_lequal
:
1531 emit(ir
, TGSI_OPCODE_SGE
, result_dst
, op
[1], op
[0]);
1533 case ir_binop_gequal
:
1534 emit(ir
, TGSI_OPCODE_SGE
, result_dst
, op
[0], op
[1]);
1536 case ir_binop_equal
:
1537 emit(ir
, TGSI_OPCODE_SEQ
, result_dst
, op
[0], op
[1]);
1539 case ir_binop_nequal
:
1540 emit(ir
, TGSI_OPCODE_SNE
, result_dst
, op
[0], op
[1]);
1542 case ir_binop_all_equal
:
1543 /* "==" operator producing a scalar boolean. */
1544 if (ir
->operands
[0]->type
->is_vector() ||
1545 ir
->operands
[1]->type
->is_vector()) {
1546 st_src_reg temp
= get_temp(native_integers
?
1547 glsl_type::get_instance(ir
->operands
[0]->type
->base_type
, 4, 1) :
1548 glsl_type::vec4_type
);
1550 if (native_integers
) {
1551 st_dst_reg temp_dst
= st_dst_reg(temp
);
1552 st_src_reg temp1
= st_src_reg(temp
), temp2
= st_src_reg(temp
);
1554 emit(ir
, TGSI_OPCODE_SEQ
, st_dst_reg(temp
), op
[0], op
[1]);
1556 /* Emit 1-3 AND operations to combine the SEQ results. */
1557 switch (ir
->operands
[0]->type
->vector_elements
) {
1561 temp_dst
.writemask
= WRITEMASK_Y
;
1562 temp1
.swizzle
= SWIZZLE_YYYY
;
1563 temp2
.swizzle
= SWIZZLE_ZZZZ
;
1564 emit(ir
, TGSI_OPCODE_AND
, temp_dst
, temp1
, temp2
);
1567 temp_dst
.writemask
= WRITEMASK_X
;
1568 temp1
.swizzle
= SWIZZLE_XXXX
;
1569 temp2
.swizzle
= SWIZZLE_YYYY
;
1570 emit(ir
, TGSI_OPCODE_AND
, temp_dst
, temp1
, temp2
);
1571 temp_dst
.writemask
= WRITEMASK_Y
;
1572 temp1
.swizzle
= SWIZZLE_ZZZZ
;
1573 temp2
.swizzle
= SWIZZLE_WWWW
;
1574 emit(ir
, TGSI_OPCODE_AND
, temp_dst
, temp1
, temp2
);
1577 temp1
.swizzle
= SWIZZLE_XXXX
;
1578 temp2
.swizzle
= SWIZZLE_YYYY
;
1579 emit(ir
, TGSI_OPCODE_AND
, result_dst
, temp1
, temp2
);
1581 emit(ir
, TGSI_OPCODE_SNE
, st_dst_reg(temp
), op
[0], op
[1]);
1583 /* After the dot-product, the value will be an integer on the
1584 * range [0,4]. Zero becomes 1.0, and positive values become zero.
1586 emit_dp(ir
, result_dst
, temp
, temp
, vector_elements
);
1588 /* Negating the result of the dot-product gives values on the range
1589 * [-4, 0]. Zero becomes 1.0, and negative values become zero.
1590 * This is achieved using SGE.
1592 st_src_reg sge_src
= result_src
;
1593 sge_src
.negate
= ~sge_src
.negate
;
1594 emit(ir
, TGSI_OPCODE_SGE
, result_dst
, sge_src
, st_src_reg_for_float(0.0));
1597 emit(ir
, TGSI_OPCODE_SEQ
, result_dst
, op
[0], op
[1]);
1600 case ir_binop_any_nequal
:
1601 /* "!=" operator producing a scalar boolean. */
1602 if (ir
->operands
[0]->type
->is_vector() ||
1603 ir
->operands
[1]->type
->is_vector()) {
1604 st_src_reg temp
= get_temp(native_integers
?
1605 glsl_type::get_instance(ir
->operands
[0]->type
->base_type
, 4, 1) :
1606 glsl_type::vec4_type
);
1607 emit(ir
, TGSI_OPCODE_SNE
, st_dst_reg(temp
), op
[0], op
[1]);
1609 if (native_integers
) {
1610 st_dst_reg temp_dst
= st_dst_reg(temp
);
1611 st_src_reg temp1
= st_src_reg(temp
), temp2
= st_src_reg(temp
);
1613 /* Emit 1-3 OR operations to combine the SNE results. */
1614 switch (ir
->operands
[0]->type
->vector_elements
) {
1618 temp_dst
.writemask
= WRITEMASK_Y
;
1619 temp1
.swizzle
= SWIZZLE_YYYY
;
1620 temp2
.swizzle
= SWIZZLE_ZZZZ
;
1621 emit(ir
, TGSI_OPCODE_OR
, temp_dst
, temp1
, temp2
);
1624 temp_dst
.writemask
= WRITEMASK_X
;
1625 temp1
.swizzle
= SWIZZLE_XXXX
;
1626 temp2
.swizzle
= SWIZZLE_YYYY
;
1627 emit(ir
, TGSI_OPCODE_OR
, temp_dst
, temp1
, temp2
);
1628 temp_dst
.writemask
= WRITEMASK_Y
;
1629 temp1
.swizzle
= SWIZZLE_ZZZZ
;
1630 temp2
.swizzle
= SWIZZLE_WWWW
;
1631 emit(ir
, TGSI_OPCODE_OR
, temp_dst
, temp1
, temp2
);
1634 temp1
.swizzle
= SWIZZLE_XXXX
;
1635 temp2
.swizzle
= SWIZZLE_YYYY
;
1636 emit(ir
, TGSI_OPCODE_OR
, result_dst
, temp1
, temp2
);
1638 /* After the dot-product, the value will be an integer on the
1639 * range [0,4]. Zero stays zero, and positive values become 1.0.
1641 glsl_to_tgsi_instruction
*const dp
=
1642 emit_dp(ir
, result_dst
, temp
, temp
, vector_elements
);
1643 if (this->prog
->Target
== GL_FRAGMENT_PROGRAM_ARB
) {
1644 /* The clamping to [0,1] can be done for free in the fragment
1645 * shader with a saturate.
1647 dp
->saturate
= true;
1649 /* Negating the result of the dot-product gives values on the range
1650 * [-4, 0]. Zero stays zero, and negative values become 1.0. This
1651 * achieved using SLT.
1653 st_src_reg slt_src
= result_src
;
1654 slt_src
.negate
= ~slt_src
.negate
;
1655 emit(ir
, TGSI_OPCODE_SLT
, result_dst
, slt_src
, st_src_reg_for_float(0.0));
1659 emit(ir
, TGSI_OPCODE_SNE
, result_dst
, op
[0], op
[1]);
1664 assert(ir
->operands
[0]->type
->is_vector());
1666 if (native_integers
) {
1667 int dst_swizzle
= 0, op0_swizzle
, i
;
1668 st_src_reg accum
= op
[0];
1670 op0_swizzle
= op
[0].swizzle
;
1671 accum
.swizzle
= MAKE_SWIZZLE4(GET_SWZ(op0_swizzle
, 0),
1672 GET_SWZ(op0_swizzle
, 0),
1673 GET_SWZ(op0_swizzle
, 0),
1674 GET_SWZ(op0_swizzle
, 0));
1675 for (i
= 0; i
< 4; i
++) {
1676 if (result_dst
.writemask
& (1 << i
)) {
1677 dst_swizzle
= MAKE_SWIZZLE4(i
, i
, i
, i
);
1682 assert(ir
->operands
[0]->type
->is_boolean());
1684 /* OR all the components together, since they should be either 0 or ~0
1686 switch (ir
->operands
[0]->type
->vector_elements
) {
1688 op
[0].swizzle
= MAKE_SWIZZLE4(GET_SWZ(op0_swizzle
, 3),
1689 GET_SWZ(op0_swizzle
, 3),
1690 GET_SWZ(op0_swizzle
, 3),
1691 GET_SWZ(op0_swizzle
, 3));
1692 emit(ir
, TGSI_OPCODE_OR
, result_dst
, accum
, op
[0]);
1693 accum
= st_src_reg(result_dst
);
1694 accum
.swizzle
= dst_swizzle
;
1697 op
[0].swizzle
= MAKE_SWIZZLE4(GET_SWZ(op0_swizzle
, 2),
1698 GET_SWZ(op0_swizzle
, 2),
1699 GET_SWZ(op0_swizzle
, 2),
1700 GET_SWZ(op0_swizzle
, 2));
1701 emit(ir
, TGSI_OPCODE_OR
, result_dst
, accum
, op
[0]);
1702 accum
= st_src_reg(result_dst
);
1703 accum
.swizzle
= dst_swizzle
;
1706 op
[0].swizzle
= MAKE_SWIZZLE4(GET_SWZ(op0_swizzle
, 1),
1707 GET_SWZ(op0_swizzle
, 1),
1708 GET_SWZ(op0_swizzle
, 1),
1709 GET_SWZ(op0_swizzle
, 1));
1710 emit(ir
, TGSI_OPCODE_OR
, result_dst
, accum
, op
[0]);
1713 assert(!"Unexpected vector size");
1717 /* After the dot-product, the value will be an integer on the
1718 * range [0,4]. Zero stays zero, and positive values become 1.0.
1720 glsl_to_tgsi_instruction
*const dp
=
1721 emit_dp(ir
, result_dst
, op
[0], op
[0],
1722 ir
->operands
[0]->type
->vector_elements
);
1723 if (this->prog
->Target
== GL_FRAGMENT_PROGRAM_ARB
&&
1724 result_dst
.type
== GLSL_TYPE_FLOAT
) {
1725 /* The clamping to [0,1] can be done for free in the fragment
1726 * shader with a saturate.
1728 dp
->saturate
= true;
1729 } else if (result_dst
.type
== GLSL_TYPE_FLOAT
) {
1730 /* Negating the result of the dot-product gives values on the range
1731 * [-4, 0]. Zero stays zero, and negative values become 1.0. This
1732 * is achieved using SLT.
1734 st_src_reg slt_src
= result_src
;
1735 slt_src
.negate
= ~slt_src
.negate
;
1736 emit(ir
, TGSI_OPCODE_SLT
, result_dst
, slt_src
, st_src_reg_for_float(0.0));
1739 /* Use SNE 0 if integers are being used as boolean values. */
1740 emit(ir
, TGSI_OPCODE_SNE
, result_dst
, result_src
, st_src_reg_for_int(0));
1746 case ir_binop_logic_xor
:
1747 if (native_integers
)
1748 emit(ir
, TGSI_OPCODE_XOR
, result_dst
, op
[0], op
[1]);
1750 emit(ir
, TGSI_OPCODE_SNE
, result_dst
, op
[0], op
[1]);
1753 case ir_binop_logic_or
: {
1754 if (native_integers
) {
1755 /* If integers are used as booleans, we can use an actual "or"
1758 assert(native_integers
);
1759 emit(ir
, TGSI_OPCODE_OR
, result_dst
, op
[0], op
[1]);
1761 /* After the addition, the value will be an integer on the
1762 * range [0,2]. Zero stays zero, and positive values become 1.0.
1764 glsl_to_tgsi_instruction
*add
=
1765 emit(ir
, TGSI_OPCODE_ADD
, result_dst
, op
[0], op
[1]);
1766 if (this->prog
->Target
== GL_FRAGMENT_PROGRAM_ARB
) {
1767 /* The clamping to [0,1] can be done for free in the fragment
1768 * shader with a saturate if floats are being used as boolean values.
1770 add
->saturate
= true;
1772 /* Negating the result of the addition gives values on the range
1773 * [-2, 0]. Zero stays zero, and negative values become 1.0. This
1774 * is achieved using SLT.
1776 st_src_reg slt_src
= result_src
;
1777 slt_src
.negate
= ~slt_src
.negate
;
1778 emit(ir
, TGSI_OPCODE_SLT
, result_dst
, slt_src
, st_src_reg_for_float(0.0));
1784 case ir_binop_logic_and
:
1785 /* If native integers are disabled, the bool args are stored as float 0.0
1786 * or 1.0, so "mul" gives us "and". If they're enabled, just use the
1787 * actual AND opcode.
1789 if (native_integers
)
1790 emit(ir
, TGSI_OPCODE_AND
, result_dst
, op
[0], op
[1]);
1792 emit(ir
, TGSI_OPCODE_MUL
, result_dst
, op
[0], op
[1]);
1796 assert(ir
->operands
[0]->type
->is_vector());
1797 assert(ir
->operands
[0]->type
== ir
->operands
[1]->type
);
1798 emit_dp(ir
, result_dst
, op
[0], op
[1],
1799 ir
->operands
[0]->type
->vector_elements
);
1804 emit_scalar(ir
, TGSI_OPCODE_SQRT
, result_dst
, op
[0]);
1807 /* sqrt(x) = x * rsq(x). */
1808 emit_scalar(ir
, TGSI_OPCODE_RSQ
, result_dst
, op
[0]);
1809 emit(ir
, TGSI_OPCODE_MUL
, result_dst
, result_src
, op
[0]);
1810 /* For incoming channels <= 0, set the result to 0. */
1811 op
[0].negate
= ~op
[0].negate
;
1812 emit(ir
, TGSI_OPCODE_CMP
, result_dst
,
1813 op
[0], result_src
, st_src_reg_for_float(0.0));
1817 emit_scalar(ir
, TGSI_OPCODE_RSQ
, result_dst
, op
[0]);
1820 if (native_integers
) {
1821 emit(ir
, TGSI_OPCODE_I2F
, result_dst
, op
[0]);
1824 /* fallthrough to next case otherwise */
1826 if (native_integers
) {
1827 emit(ir
, TGSI_OPCODE_AND
, result_dst
, op
[0], st_src_reg_for_float(1.0));
1830 /* fallthrough to next case otherwise */
1833 /* Converting between signed and unsigned integers is a no-op. */
1837 if (native_integers
) {
1838 /* Booleans are stored as integers using ~0 for true and 0 for false.
1839 * GLSL requires that int(bool) return 1 for true and 0 for false.
1840 * This conversion is done with AND, but it could be done with NEG.
1842 emit(ir
, TGSI_OPCODE_AND
, result_dst
, op
[0], st_src_reg_for_int(1));
1844 /* Booleans and integers are both stored as floats when native
1845 * integers are disabled.
1851 if (native_integers
)
1852 emit(ir
, TGSI_OPCODE_F2I
, result_dst
, op
[0]);
1854 emit(ir
, TGSI_OPCODE_TRUNC
, result_dst
, op
[0]);
1857 if (native_integers
)
1858 emit(ir
, TGSI_OPCODE_F2U
, result_dst
, op
[0]);
1860 emit(ir
, TGSI_OPCODE_TRUNC
, result_dst
, op
[0]);
1862 case ir_unop_bitcast_f2i
:
1864 result_src
.type
= GLSL_TYPE_INT
;
1866 case ir_unop_bitcast_f2u
:
1868 result_src
.type
= GLSL_TYPE_UINT
;
1870 case ir_unop_bitcast_i2f
:
1871 case ir_unop_bitcast_u2f
:
1873 result_src
.type
= GLSL_TYPE_FLOAT
;
1876 emit(ir
, TGSI_OPCODE_SNE
, result_dst
, op
[0], st_src_reg_for_float(0.0));
1879 if (native_integers
)
1880 emit(ir
, TGSI_OPCODE_INEG
, result_dst
, op
[0]);
1882 emit(ir
, TGSI_OPCODE_SNE
, result_dst
, op
[0], st_src_reg_for_float(0.0));
1885 emit(ir
, TGSI_OPCODE_TRUNC
, result_dst
, op
[0]);
1888 emit(ir
, TGSI_OPCODE_CEIL
, result_dst
, op
[0]);
1891 emit(ir
, TGSI_OPCODE_FLR
, result_dst
, op
[0]);
1893 case ir_unop_round_even
:
1894 emit(ir
, TGSI_OPCODE_ROUND
, result_dst
, op
[0]);
1897 emit(ir
, TGSI_OPCODE_FRC
, result_dst
, op
[0]);
1901 emit(ir
, TGSI_OPCODE_MIN
, result_dst
, op
[0], op
[1]);
1904 emit(ir
, TGSI_OPCODE_MAX
, result_dst
, op
[0], op
[1]);
1907 emit_scalar(ir
, TGSI_OPCODE_POW
, result_dst
, op
[0], op
[1]);
1910 case ir_unop_bit_not
:
1911 if (native_integers
) {
1912 emit(ir
, TGSI_OPCODE_NOT
, result_dst
, op
[0]);
1916 if (native_integers
) {
1917 emit(ir
, TGSI_OPCODE_U2F
, result_dst
, op
[0]);
1920 case ir_binop_lshift
:
1921 if (native_integers
) {
1922 emit(ir
, TGSI_OPCODE_SHL
, result_dst
, op
[0], op
[1]);
1925 case ir_binop_rshift
:
1926 if (native_integers
) {
1927 emit(ir
, TGSI_OPCODE_ISHR
, result_dst
, op
[0], op
[1]);
1930 case ir_binop_bit_and
:
1931 if (native_integers
) {
1932 emit(ir
, TGSI_OPCODE_AND
, result_dst
, op
[0], op
[1]);
1935 case ir_binop_bit_xor
:
1936 if (native_integers
) {
1937 emit(ir
, TGSI_OPCODE_XOR
, result_dst
, op
[0], op
[1]);
1940 case ir_binop_bit_or
:
1941 if (native_integers
) {
1942 emit(ir
, TGSI_OPCODE_OR
, result_dst
, op
[0], op
[1]);
1946 assert(!"GLSL 1.30 features unsupported");
1949 case ir_binop_ubo_load
: {
1950 ir_constant
*const_uniform_block
= ir
->operands
[0]->as_constant();
1951 ir_constant
*const_offset_ir
= ir
->operands
[1]->as_constant();
1952 unsigned const_offset
= const_offset_ir
? const_offset_ir
->value
.u
[0] : 0;
1953 unsigned const_block
= const_uniform_block
? const_uniform_block
->value
.u
[0] + 1 : 0;
1954 st_src_reg index_reg
= get_temp(glsl_type::uint_type
);
1957 cbuf
.type
= glsl_type::vec4_type
->base_type
;
1958 cbuf
.file
= PROGRAM_CONSTANT
;
1960 cbuf
.reladdr
= NULL
;
1963 assert(ir
->type
->is_vector() || ir
->type
->is_scalar());
1965 if (const_offset_ir
) {
1966 /* Constant index into constant buffer */
1967 cbuf
.reladdr
= NULL
;
1968 cbuf
.index
= const_offset
/ 16;
1971 /* Relative/variable index into constant buffer */
1972 emit(ir
, TGSI_OPCODE_USHR
, st_dst_reg(index_reg
), op
[1],
1973 st_src_reg_for_int(4));
1974 cbuf
.reladdr
= ralloc(mem_ctx
, st_src_reg
);
1975 memcpy(cbuf
.reladdr
, &index_reg
, sizeof(index_reg
));
1978 if (const_uniform_block
) {
1979 /* Constant constant buffer */
1980 cbuf
.reladdr2
= NULL
;
1981 cbuf
.index2D
= const_block
;
1982 cbuf
.has_index2
= true;
1985 /* Relative/variable constant buffer */
1986 cbuf
.reladdr2
= ralloc(mem_ctx
, st_src_reg
);
1988 memcpy(cbuf
.reladdr2
, &op
[0], sizeof(st_src_reg
));
1989 cbuf
.has_index2
= true;
1992 cbuf
.swizzle
= swizzle_for_size(ir
->type
->vector_elements
);
1993 cbuf
.swizzle
+= MAKE_SWIZZLE4(const_offset
% 16 / 4,
1994 const_offset
% 16 / 4,
1995 const_offset
% 16 / 4,
1996 const_offset
% 16 / 4);
1998 if (ir
->type
->base_type
== GLSL_TYPE_BOOL
) {
1999 emit(ir
, TGSI_OPCODE_USNE
, result_dst
, cbuf
, st_src_reg_for_int(0));
2001 emit(ir
, TGSI_OPCODE_MOV
, result_dst
, cbuf
);
2006 /* note: we have to reorder the three args here */
2007 emit(ir
, TGSI_OPCODE_LRP
, result_dst
, op
[2], op
[1], op
[0]);
2010 if (this->ctx
->Const
.NativeIntegers
)
2011 emit(ir
, TGSI_OPCODE_UCMP
, result_dst
, op
[0], op
[1], op
[2]);
2013 op
[0].negate
= ~op
[0].negate
;
2014 emit(ir
, TGSI_OPCODE_CMP
, result_dst
, op
[0], op
[1], op
[2]);
2017 case ir_triop_bitfield_extract
:
2018 emit(ir
, TGSI_OPCODE_IBFE
, result_dst
, op
[0], op
[1], op
[2]);
2020 case ir_quadop_bitfield_insert
:
2021 emit(ir
, TGSI_OPCODE_BFI
, result_dst
, op
[0], op
[1], op
[2], op
[3]);
2023 case ir_unop_bitfield_reverse
:
2024 emit(ir
, TGSI_OPCODE_BREV
, result_dst
, op
[0]);
2026 case ir_unop_bit_count
:
2027 emit(ir
, TGSI_OPCODE_POPC
, result_dst
, op
[0]);
2029 case ir_unop_find_msb
:
2030 emit(ir
, TGSI_OPCODE_IMSB
, result_dst
, op
[0]);
2032 case ir_unop_find_lsb
:
2033 emit(ir
, TGSI_OPCODE_LSB
, result_dst
, op
[0]);
2035 case ir_binop_imul_high
:
2036 emit(ir
, TGSI_OPCODE_IMUL_HI
, result_dst
, op
[0], op
[1]);
2039 /* NOTE: Perhaps there should be a special opcode that enforces fused
2040 * mul-add. Just use MAD for now.
2042 emit(ir
, TGSI_OPCODE_MAD
, result_dst
, op
[0], op
[1], op
[2]);
2044 case ir_unop_interpolate_at_centroid
:
2045 emit(ir
, TGSI_OPCODE_INTERP_CENTROID
, result_dst
, op
[0]);
2047 case ir_binop_interpolate_at_offset
:
2048 emit(ir
, TGSI_OPCODE_INTERP_OFFSET
, result_dst
, op
[0], op
[1]);
2050 case ir_binop_interpolate_at_sample
:
2051 emit(ir
, TGSI_OPCODE_INTERP_SAMPLE
, result_dst
, op
[0], op
[1]);
2053 case ir_unop_pack_snorm_2x16
:
2054 case ir_unop_pack_unorm_2x16
:
2055 case ir_unop_pack_half_2x16
:
2056 case ir_unop_pack_snorm_4x8
:
2057 case ir_unop_pack_unorm_4x8
:
2058 case ir_unop_unpack_snorm_2x16
:
2059 case ir_unop_unpack_unorm_2x16
:
2060 case ir_unop_unpack_half_2x16
:
2061 case ir_unop_unpack_half_2x16_split_x
:
2062 case ir_unop_unpack_half_2x16_split_y
:
2063 case ir_unop_unpack_snorm_4x8
:
2064 case ir_unop_unpack_unorm_4x8
:
2065 case ir_binop_pack_half_2x16_split
:
2068 case ir_quadop_vector
:
2069 case ir_binop_vector_extract
:
2070 case ir_triop_vector_insert
:
2071 case ir_binop_ldexp
:
2072 case ir_binop_carry
:
2073 case ir_binop_borrow
:
2074 /* This operation is not supported, or should have already been handled.
2076 assert(!"Invalid ir opcode in glsl_to_tgsi_visitor::visit()");
2080 this->result
= result_src
;
2085 glsl_to_tgsi_visitor::visit(ir_swizzle
*ir
)
2091 /* Note that this is only swizzles in expressions, not those on the left
2092 * hand side of an assignment, which do write masking. See ir_assignment
2096 ir
->val
->accept(this);
2098 assert(src
.file
!= PROGRAM_UNDEFINED
);
2100 for (i
= 0; i
< 4; i
++) {
2101 if (i
< ir
->type
->vector_elements
) {
2104 swizzle
[i
] = GET_SWZ(src
.swizzle
, ir
->mask
.x
);
2107 swizzle
[i
] = GET_SWZ(src
.swizzle
, ir
->mask
.y
);
2110 swizzle
[i
] = GET_SWZ(src
.swizzle
, ir
->mask
.z
);
2113 swizzle
[i
] = GET_SWZ(src
.swizzle
, ir
->mask
.w
);
2117 /* If the type is smaller than a vec4, replicate the last
2120 swizzle
[i
] = swizzle
[ir
->type
->vector_elements
- 1];
2124 src
.swizzle
= MAKE_SWIZZLE4(swizzle
[0], swizzle
[1], swizzle
[2], swizzle
[3]);
2130 glsl_to_tgsi_visitor::visit(ir_dereference_variable
*ir
)
2132 variable_storage
*entry
= find_variable_storage(ir
->var
);
2133 ir_variable
*var
= ir
->var
;
2136 switch (var
->data
.mode
) {
2137 case ir_var_uniform
:
2138 entry
= new(mem_ctx
) variable_storage(var
, PROGRAM_UNIFORM
,
2139 var
->data
.location
);
2140 this->variables
.push_tail(entry
);
2142 case ir_var_shader_in
:
2143 /* The linker assigns locations for varyings and attributes,
2144 * including deprecated builtins (like gl_Color), user-assign
2145 * generic attributes (glBindVertexLocation), and
2146 * user-defined varyings.
2148 assert(var
->data
.location
!= -1);
2149 entry
= new(mem_ctx
) variable_storage(var
,
2151 var
->data
.location
);
2153 case ir_var_shader_out
:
2154 assert(var
->data
.location
!= -1);
2155 entry
= new(mem_ctx
) variable_storage(var
,
2160 case ir_var_system_value
:
2161 entry
= new(mem_ctx
) variable_storage(var
,
2162 PROGRAM_SYSTEM_VALUE
,
2163 var
->data
.location
);
2166 case ir_var_temporary
:
2167 st_src_reg src
= get_temp(var
->type
);
2169 entry
= new(mem_ctx
) variable_storage(var
, src
.file
, src
.index
);
2170 this->variables
.push_tail(entry
);
2176 printf("Failed to make storage for %s\n", var
->name
);
2181 this->result
= st_src_reg(entry
->file
, entry
->index
, var
->type
);
2182 if (!native_integers
)
2183 this->result
.type
= GLSL_TYPE_FLOAT
;
2187 glsl_to_tgsi_visitor::visit(ir_dereference_array
*ir
)
2191 int element_size
= type_size(ir
->type
);
2194 index
= ir
->array_index
->constant_expression_value();
2196 ir
->array
->accept(this);
2199 is_2D_input
= this->prog
->Target
== GL_GEOMETRY_PROGRAM_NV
&&
2200 src
.file
== PROGRAM_INPUT
&&
2201 ir
->array
->ir_type
!= ir_type_dereference_array
;
2208 src
.index2D
= index
->value
.i
[0];
2209 src
.has_index2
= true;
2211 src
.index
+= index
->value
.i
[0] * element_size
;
2213 /* Variable index array dereference. It eats the "vec4" of the
2214 * base of the array and an index that offsets the TGSI register
2217 ir
->array_index
->accept(this);
2219 st_src_reg index_reg
;
2221 if (element_size
== 1) {
2222 index_reg
= this->result
;
2224 index_reg
= get_temp(native_integers
?
2225 glsl_type::int_type
: glsl_type::float_type
);
2227 emit(ir
, TGSI_OPCODE_MUL
, st_dst_reg(index_reg
),
2228 this->result
, st_src_reg_for_type(index_reg
.type
, element_size
));
2231 /* If there was already a relative address register involved, add the
2232 * new and the old together to get the new offset.
2234 if (!is_2D_input
&& src
.reladdr
!= NULL
) {
2235 st_src_reg accum_reg
= get_temp(native_integers
?
2236 glsl_type::int_type
: glsl_type::float_type
);
2238 emit(ir
, TGSI_OPCODE_ADD
, st_dst_reg(accum_reg
),
2239 index_reg
, *src
.reladdr
);
2241 index_reg
= accum_reg
;
2245 src
.reladdr2
= ralloc(mem_ctx
, st_src_reg
);
2246 memcpy(src
.reladdr2
, &index_reg
, sizeof(index_reg
));
2248 src
.has_index2
= true;
2250 src
.reladdr
= ralloc(mem_ctx
, st_src_reg
);
2251 memcpy(src
.reladdr
, &index_reg
, sizeof(index_reg
));
2255 /* If the type is smaller than a vec4, replicate the last channel out. */
2256 if (ir
->type
->is_scalar() || ir
->type
->is_vector())
2257 src
.swizzle
= swizzle_for_size(ir
->type
->vector_elements
);
2259 src
.swizzle
= SWIZZLE_NOOP
;
2261 /* Change the register type to the element type of the array. */
2262 src
.type
= ir
->type
->base_type
;
2268 glsl_to_tgsi_visitor::visit(ir_dereference_record
*ir
)
2271 const glsl_type
*struct_type
= ir
->record
->type
;
2274 ir
->record
->accept(this);
2276 for (i
= 0; i
< struct_type
->length
; i
++) {
2277 if (strcmp(struct_type
->fields
.structure
[i
].name
, ir
->field
) == 0)
2279 offset
+= type_size(struct_type
->fields
.structure
[i
].type
);
2282 /* If the type is smaller than a vec4, replicate the last channel out. */
2283 if (ir
->type
->is_scalar() || ir
->type
->is_vector())
2284 this->result
.swizzle
= swizzle_for_size(ir
->type
->vector_elements
);
2286 this->result
.swizzle
= SWIZZLE_NOOP
;
2288 this->result
.index
+= offset
;
2289 this->result
.type
= ir
->type
->base_type
;
2293 * We want to be careful in assignment setup to hit the actual storage
2294 * instead of potentially using a temporary like we might with the
2295 * ir_dereference handler.
2298 get_assignment_lhs(ir_dereference
*ir
, glsl_to_tgsi_visitor
*v
)
2300 /* The LHS must be a dereference. If the LHS is a variable indexed array
2301 * access of a vector, it must be separated into a series conditional moves
2302 * before reaching this point (see ir_vec_index_to_cond_assign).
2304 assert(ir
->as_dereference());
2305 ir_dereference_array
*deref_array
= ir
->as_dereference_array();
2307 assert(!deref_array
->array
->type
->is_vector());
2310 /* Use the rvalue deref handler for the most part. We'll ignore
2311 * swizzles in it and write swizzles using writemask, though.
2314 return st_dst_reg(v
->result
);
2318 * Process the condition of a conditional assignment
2320 * Examines the condition of a conditional assignment to generate the optimal
2321 * first operand of a \c CMP instruction. If the condition is a relational
2322 * operator with 0 (e.g., \c ir_binop_less), the value being compared will be
2323 * used as the source for the \c CMP instruction. Otherwise the comparison
2324 * is processed to a boolean result, and the boolean result is used as the
2325 * operand to the CMP instruction.
2328 glsl_to_tgsi_visitor::process_move_condition(ir_rvalue
*ir
)
2330 ir_rvalue
*src_ir
= ir
;
2332 bool switch_order
= false;
2334 ir_expression
*const expr
= ir
->as_expression();
2335 if ((expr
!= NULL
) && (expr
->get_num_operands() == 2)) {
2336 bool zero_on_left
= false;
2338 if (expr
->operands
[0]->is_zero()) {
2339 src_ir
= expr
->operands
[1];
2340 zero_on_left
= true;
2341 } else if (expr
->operands
[1]->is_zero()) {
2342 src_ir
= expr
->operands
[0];
2343 zero_on_left
= false;
2347 * (a < 0) T F F ( a < 0) T F F
2348 * (0 < a) F F T (-a < 0) F F T
2349 * (a <= 0) T T F (-a < 0) F F T (swap order of other operands)
2350 * (0 <= a) F T T ( a < 0) T F F (swap order of other operands)
2351 * (a > 0) F F T (-a < 0) F F T
2352 * (0 > a) T F F ( a < 0) T F F
2353 * (a >= 0) F T T ( a < 0) T F F (swap order of other operands)
2354 * (0 >= a) T T F (-a < 0) F F T (swap order of other operands)
2356 * Note that exchanging the order of 0 and 'a' in the comparison simply
2357 * means that the value of 'a' should be negated.
2360 switch (expr
->operation
) {
2362 switch_order
= false;
2363 negate
= zero_on_left
;
2366 case ir_binop_greater
:
2367 switch_order
= false;
2368 negate
= !zero_on_left
;
2371 case ir_binop_lequal
:
2372 switch_order
= true;
2373 negate
= !zero_on_left
;
2376 case ir_binop_gequal
:
2377 switch_order
= true;
2378 negate
= zero_on_left
;
2382 /* This isn't the right kind of comparison afterall, so make sure
2383 * the whole condition is visited.
2391 src_ir
->accept(this);
2393 /* We use the TGSI_OPCODE_CMP (a < 0 ? b : c) for conditional moves, and the
2394 * condition we produced is 0.0 or 1.0. By flipping the sign, we can
2395 * choose which value TGSI_OPCODE_CMP produces without an extra instruction
2396 * computing the condition.
2399 this->result
.negate
= ~this->result
.negate
;
2401 return switch_order
;
2405 glsl_to_tgsi_visitor::emit_block_mov(ir_assignment
*ir
, const struct glsl_type
*type
,
2406 st_dst_reg
*l
, st_src_reg
*r
)
2408 if (type
->base_type
== GLSL_TYPE_STRUCT
) {
2409 for (unsigned int i
= 0; i
< type
->length
; i
++) {
2410 emit_block_mov(ir
, type
->fields
.structure
[i
].type
, l
, r
);
2415 if (type
->is_array()) {
2416 for (unsigned int i
= 0; i
< type
->length
; i
++) {
2417 emit_block_mov(ir
, type
->fields
.array
, l
, r
);
2422 if (type
->is_matrix()) {
2423 const struct glsl_type
*vec_type
;
2425 vec_type
= glsl_type::get_instance(GLSL_TYPE_FLOAT
,
2426 type
->vector_elements
, 1);
2428 for (int i
= 0; i
< type
->matrix_columns
; i
++) {
2429 emit_block_mov(ir
, vec_type
, l
, r
);
2434 assert(type
->is_scalar() || type
->is_vector());
2436 r
->type
= type
->base_type
;
2437 emit(ir
, TGSI_OPCODE_MOV
, *l
, *r
);
2443 glsl_to_tgsi_visitor::visit(ir_assignment
*ir
)
2449 ir
->rhs
->accept(this);
2452 l
= get_assignment_lhs(ir
->lhs
, this);
2454 /* FINISHME: This should really set to the correct maximal writemask for each
2455 * FINISHME: component written (in the loops below). This case can only
2456 * FINISHME: occur for matrices, arrays, and structures.
2458 if (ir
->write_mask
== 0) {
2459 assert(!ir
->lhs
->type
->is_scalar() && !ir
->lhs
->type
->is_vector());
2460 l
.writemask
= WRITEMASK_XYZW
;
2461 } else if (ir
->lhs
->type
->is_scalar() &&
2462 ir
->lhs
->variable_referenced()->data
.mode
== ir_var_shader_out
) {
2463 /* FINISHME: This hack makes writing to gl_FragDepth, which lives in the
2464 * FINISHME: W component of fragment shader output zero, work correctly.
2466 l
.writemask
= WRITEMASK_XYZW
;
2469 int first_enabled_chan
= 0;
2472 l
.writemask
= ir
->write_mask
;
2474 for (int i
= 0; i
< 4; i
++) {
2475 if (l
.writemask
& (1 << i
)) {
2476 first_enabled_chan
= GET_SWZ(r
.swizzle
, i
);
2481 /* Swizzle a small RHS vector into the channels being written.
2483 * glsl ir treats write_mask as dictating how many channels are
2484 * present on the RHS while TGSI treats write_mask as just
2485 * showing which channels of the vec4 RHS get written.
2487 for (int i
= 0; i
< 4; i
++) {
2488 if (l
.writemask
& (1 << i
))
2489 swizzles
[i
] = GET_SWZ(r
.swizzle
, rhs_chan
++);
2491 swizzles
[i
] = first_enabled_chan
;
2493 r
.swizzle
= MAKE_SWIZZLE4(swizzles
[0], swizzles
[1],
2494 swizzles
[2], swizzles
[3]);
2497 assert(l
.file
!= PROGRAM_UNDEFINED
);
2498 assert(r
.file
!= PROGRAM_UNDEFINED
);
2500 if (ir
->condition
) {
2501 const bool switch_order
= this->process_move_condition(ir
->condition
);
2502 st_src_reg condition
= this->result
;
2504 for (i
= 0; i
< type_size(ir
->lhs
->type
); i
++) {
2505 st_src_reg l_src
= st_src_reg(l
);
2506 st_src_reg condition_temp
= condition
;
2507 l_src
.swizzle
= swizzle_for_size(ir
->lhs
->type
->vector_elements
);
2509 if (native_integers
) {
2510 /* This is necessary because TGSI's CMP instruction expects the
2511 * condition to be a float, and we store booleans as integers.
2512 * TODO: really want to avoid i2f path and use UCMP. Requires
2513 * changes to process_move_condition though too.
2515 condition_temp
= get_temp(glsl_type::vec4_type
);
2516 condition
.negate
= 0;
2517 emit(ir
, TGSI_OPCODE_I2F
, st_dst_reg(condition_temp
), condition
);
2518 condition_temp
.swizzle
= condition
.swizzle
;
2522 emit(ir
, TGSI_OPCODE_CMP
, l
, condition_temp
, l_src
, r
);
2524 emit(ir
, TGSI_OPCODE_CMP
, l
, condition_temp
, r
, l_src
);
2530 } else if (ir
->rhs
->as_expression() &&
2531 this->instructions
.get_tail() &&
2532 ir
->rhs
== ((glsl_to_tgsi_instruction
*)this->instructions
.get_tail())->ir
&&
2533 type_size(ir
->lhs
->type
) == 1 &&
2534 l
.writemask
== ((glsl_to_tgsi_instruction
*)this->instructions
.get_tail())->dst
.writemask
) {
2535 /* To avoid emitting an extra MOV when assigning an expression to a
2536 * variable, emit the last instruction of the expression again, but
2537 * replace the destination register with the target of the assignment.
2538 * Dead code elimination will remove the original instruction.
2540 glsl_to_tgsi_instruction
*inst
, *new_inst
;
2541 inst
= (glsl_to_tgsi_instruction
*)this->instructions
.get_tail();
2542 new_inst
= emit(ir
, inst
->op
, l
, inst
->src
[0], inst
->src
[1], inst
->src
[2]);
2543 new_inst
->saturate
= inst
->saturate
;
2544 inst
->dead_mask
= inst
->dst
.writemask
;
2546 emit_block_mov(ir
, ir
->rhs
->type
, &l
, &r
);
2552 glsl_to_tgsi_visitor::visit(ir_constant
*ir
)
2555 GLfloat stack_vals
[4] = { 0 };
2556 gl_constant_value
*values
= (gl_constant_value
*) stack_vals
;
2557 GLenum gl_type
= GL_NONE
;
2559 static int in_array
= 0;
2560 gl_register_file file
= in_array
? PROGRAM_CONSTANT
: PROGRAM_IMMEDIATE
;
2562 /* Unfortunately, 4 floats is all we can get into
2563 * _mesa_add_typed_unnamed_constant. So, make a temp to store an
2564 * aggregate constant and move each constant value into it. If we
2565 * get lucky, copy propagation will eliminate the extra moves.
2567 if (ir
->type
->base_type
== GLSL_TYPE_STRUCT
) {
2568 st_src_reg temp_base
= get_temp(ir
->type
);
2569 st_dst_reg temp
= st_dst_reg(temp_base
);
2571 foreach_in_list(ir_constant
, field_value
, &ir
->components
) {
2572 int size
= type_size(field_value
->type
);
2576 field_value
->accept(this);
2579 for (i
= 0; i
< (unsigned int)size
; i
++) {
2580 emit(ir
, TGSI_OPCODE_MOV
, temp
, src
);
2586 this->result
= temp_base
;
2590 if (ir
->type
->is_array()) {
2591 st_src_reg temp_base
= get_temp(ir
->type
);
2592 st_dst_reg temp
= st_dst_reg(temp_base
);
2593 int size
= type_size(ir
->type
->fields
.array
);
2598 for (i
= 0; i
< ir
->type
->length
; i
++) {
2599 ir
->array_elements
[i
]->accept(this);
2601 for (int j
= 0; j
< size
; j
++) {
2602 emit(ir
, TGSI_OPCODE_MOV
, temp
, src
);
2608 this->result
= temp_base
;
2613 if (ir
->type
->is_matrix()) {
2614 st_src_reg mat
= get_temp(ir
->type
);
2615 st_dst_reg mat_column
= st_dst_reg(mat
);
2617 for (i
= 0; i
< ir
->type
->matrix_columns
; i
++) {
2618 assert(ir
->type
->base_type
== GLSL_TYPE_FLOAT
);
2619 values
= (gl_constant_value
*) &ir
->value
.f
[i
* ir
->type
->vector_elements
];
2621 src
= st_src_reg(file
, -1, ir
->type
->base_type
);
2622 src
.index
= add_constant(file
,
2624 ir
->type
->vector_elements
,
2627 emit(ir
, TGSI_OPCODE_MOV
, mat_column
, src
);
2636 switch (ir
->type
->base_type
) {
2637 case GLSL_TYPE_FLOAT
:
2639 for (i
= 0; i
< ir
->type
->vector_elements
; i
++) {
2640 values
[i
].f
= ir
->value
.f
[i
];
2643 case GLSL_TYPE_UINT
:
2644 gl_type
= native_integers
? GL_UNSIGNED_INT
: GL_FLOAT
;
2645 for (i
= 0; i
< ir
->type
->vector_elements
; i
++) {
2646 if (native_integers
)
2647 values
[i
].u
= ir
->value
.u
[i
];
2649 values
[i
].f
= ir
->value
.u
[i
];
2653 gl_type
= native_integers
? GL_INT
: GL_FLOAT
;
2654 for (i
= 0; i
< ir
->type
->vector_elements
; i
++) {
2655 if (native_integers
)
2656 values
[i
].i
= ir
->value
.i
[i
];
2658 values
[i
].f
= ir
->value
.i
[i
];
2661 case GLSL_TYPE_BOOL
:
2662 gl_type
= native_integers
? GL_BOOL
: GL_FLOAT
;
2663 for (i
= 0; i
< ir
->type
->vector_elements
; i
++) {
2664 if (native_integers
)
2665 values
[i
].u
= ir
->value
.b
[i
] ? ~0 : 0;
2667 values
[i
].f
= ir
->value
.b
[i
];
2671 assert(!"Non-float/uint/int/bool constant");
2674 this->result
= st_src_reg(file
, -1, ir
->type
);
2675 this->result
.index
= add_constant(file
,
2677 ir
->type
->vector_elements
,
2679 &this->result
.swizzle
);
2683 glsl_to_tgsi_visitor::get_function_signature(ir_function_signature
*sig
)
2685 foreach_in_list_use_after(function_entry
, entry
, &this->function_signatures
) {
2686 if (entry
->sig
== sig
)
2690 entry
= ralloc(mem_ctx
, function_entry
);
2692 entry
->sig_id
= this->next_signature_id
++;
2693 entry
->bgn_inst
= NULL
;
2695 /* Allocate storage for all the parameters. */
2696 foreach_in_list(ir_variable
, param
, &sig
->parameters
) {
2697 variable_storage
*storage
;
2699 storage
= find_variable_storage(param
);
2702 st_src_reg src
= get_temp(param
->type
);
2704 storage
= new(mem_ctx
) variable_storage(param
, src
.file
, src
.index
);
2705 this->variables
.push_tail(storage
);
2708 if (!sig
->return_type
->is_void()) {
2709 entry
->return_reg
= get_temp(sig
->return_type
);
2711 entry
->return_reg
= undef_src
;
2714 this->function_signatures
.push_tail(entry
);
2719 glsl_to_tgsi_visitor::visit(ir_call
*ir
)
2721 glsl_to_tgsi_instruction
*call_inst
;
2722 ir_function_signature
*sig
= ir
->callee
;
2723 function_entry
*entry
= get_function_signature(sig
);
2726 /* Process in parameters. */
2727 foreach_two_lists(formal_node
, &sig
->parameters
,
2728 actual_node
, &ir
->actual_parameters
) {
2729 ir_rvalue
*param_rval
= (ir_rvalue
*) actual_node
;
2730 ir_variable
*param
= (ir_variable
*) formal_node
;
2732 if (param
->data
.mode
== ir_var_function_in
||
2733 param
->data
.mode
== ir_var_function_inout
) {
2734 variable_storage
*storage
= find_variable_storage(param
);
2737 param_rval
->accept(this);
2738 st_src_reg r
= this->result
;
2741 l
.file
= storage
->file
;
2742 l
.index
= storage
->index
;
2744 l
.writemask
= WRITEMASK_XYZW
;
2745 l
.cond_mask
= COND_TR
;
2747 for (i
= 0; i
< type_size(param
->type
); i
++) {
2748 emit(ir
, TGSI_OPCODE_MOV
, l
, r
);
2755 /* Emit call instruction */
2756 call_inst
= emit(ir
, TGSI_OPCODE_CAL
);
2757 call_inst
->function
= entry
;
2759 /* Process out parameters. */
2760 foreach_two_lists(formal_node
, &sig
->parameters
,
2761 actual_node
, &ir
->actual_parameters
) {
2762 ir_rvalue
*param_rval
= (ir_rvalue
*) actual_node
;
2763 ir_variable
*param
= (ir_variable
*) formal_node
;
2765 if (param
->data
.mode
== ir_var_function_out
||
2766 param
->data
.mode
== ir_var_function_inout
) {
2767 variable_storage
*storage
= find_variable_storage(param
);
2771 r
.file
= storage
->file
;
2772 r
.index
= storage
->index
;
2774 r
.swizzle
= SWIZZLE_NOOP
;
2777 param_rval
->accept(this);
2778 st_dst_reg l
= st_dst_reg(this->result
);
2780 for (i
= 0; i
< type_size(param
->type
); i
++) {
2781 emit(ir
, TGSI_OPCODE_MOV
, l
, r
);
2788 /* Process return value. */
2789 this->result
= entry
->return_reg
;
2793 glsl_to_tgsi_visitor::visit(ir_texture
*ir
)
2795 st_src_reg result_src
, coord
, cube_sc
, lod_info
, projector
, dx
, dy
;
2796 st_src_reg offset
[MAX_GLSL_TEXTURE_OFFSET
], sample_index
, component
;
2797 st_src_reg levels_src
;
2798 st_dst_reg result_dst
, coord_dst
, cube_sc_dst
;
2799 glsl_to_tgsi_instruction
*inst
= NULL
;
2800 unsigned opcode
= TGSI_OPCODE_NOP
;
2801 const glsl_type
*sampler_type
= ir
->sampler
->type
;
2802 bool is_cube_array
= false;
2805 /* if we are a cube array sampler */
2806 if ((sampler_type
->sampler_dimensionality
== GLSL_SAMPLER_DIM_CUBE
&&
2807 sampler_type
->sampler_array
)) {
2808 is_cube_array
= true;
2811 if (ir
->coordinate
) {
2812 ir
->coordinate
->accept(this);
2814 /* Put our coords in a temp. We'll need to modify them for shadow,
2815 * projection, or LOD, so the only case we'd use it as is is if
2816 * we're doing plain old texturing. The optimization passes on
2817 * glsl_to_tgsi_visitor should handle cleaning up our mess in that case.
2819 coord
= get_temp(glsl_type::vec4_type
);
2820 coord_dst
= st_dst_reg(coord
);
2821 coord_dst
.writemask
= (1 << ir
->coordinate
->type
->vector_elements
) - 1;
2822 emit(ir
, TGSI_OPCODE_MOV
, coord_dst
, this->result
);
2825 if (ir
->projector
) {
2826 ir
->projector
->accept(this);
2827 projector
= this->result
;
2830 /* Storage for our result. Ideally for an assignment we'd be using
2831 * the actual storage for the result here, instead.
2833 result_src
= get_temp(ir
->type
);
2834 result_dst
= st_dst_reg(result_src
);
2838 opcode
= (is_cube_array
&& ir
->shadow_comparitor
) ? TGSI_OPCODE_TEX2
: TGSI_OPCODE_TEX
;
2840 ir
->offset
->accept(this);
2841 offset
[0] = this->result
;
2845 if (is_cube_array
||
2846 sampler_type
== glsl_type::samplerCubeShadow_type
) {
2847 opcode
= TGSI_OPCODE_TXB2
;
2850 opcode
= TGSI_OPCODE_TXB
;
2852 ir
->lod_info
.bias
->accept(this);
2853 lod_info
= this->result
;
2855 ir
->offset
->accept(this);
2856 offset
[0] = this->result
;
2860 opcode
= is_cube_array
? TGSI_OPCODE_TXL2
: TGSI_OPCODE_TXL
;
2861 ir
->lod_info
.lod
->accept(this);
2862 lod_info
= this->result
;
2864 ir
->offset
->accept(this);
2865 offset
[0] = this->result
;
2869 opcode
= TGSI_OPCODE_TXD
;
2870 ir
->lod_info
.grad
.dPdx
->accept(this);
2872 ir
->lod_info
.grad
.dPdy
->accept(this);
2875 ir
->offset
->accept(this);
2876 offset
[0] = this->result
;
2880 opcode
= TGSI_OPCODE_TXQ
;
2881 ir
->lod_info
.lod
->accept(this);
2882 lod_info
= this->result
;
2884 case ir_query_levels
:
2885 opcode
= TGSI_OPCODE_TXQ
;
2886 lod_info
= st_src_reg(PROGRAM_IMMEDIATE
, 0, GLSL_TYPE_INT
);
2887 levels_src
= get_temp(ir
->type
);
2890 opcode
= TGSI_OPCODE_TXF
;
2891 ir
->lod_info
.lod
->accept(this);
2892 lod_info
= this->result
;
2894 ir
->offset
->accept(this);
2895 offset
[0] = this->result
;
2899 opcode
= TGSI_OPCODE_TXF
;
2900 ir
->lod_info
.sample_index
->accept(this);
2901 sample_index
= this->result
;
2904 opcode
= TGSI_OPCODE_TG4
;
2905 ir
->lod_info
.component
->accept(this);
2906 component
= this->result
;
2908 ir
->offset
->accept(this);
2909 if (ir
->offset
->type
->base_type
== GLSL_TYPE_ARRAY
) {
2910 const glsl_type
*elt_type
= ir
->offset
->type
->fields
.array
;
2911 for (i
= 0; i
< ir
->offset
->type
->length
; i
++) {
2912 offset
[i
] = this->result
;
2913 offset
[i
].index
+= i
* type_size(elt_type
);
2914 offset
[i
].type
= elt_type
->base_type
;
2915 offset
[i
].swizzle
= swizzle_for_size(elt_type
->vector_elements
);
2918 offset
[0] = this->result
;
2923 opcode
= TGSI_OPCODE_LODQ
;
2927 if (ir
->projector
) {
2928 if (opcode
== TGSI_OPCODE_TEX
) {
2929 /* Slot the projector in as the last component of the coord. */
2930 coord_dst
.writemask
= WRITEMASK_W
;
2931 emit(ir
, TGSI_OPCODE_MOV
, coord_dst
, projector
);
2932 coord_dst
.writemask
= WRITEMASK_XYZW
;
2933 opcode
= TGSI_OPCODE_TXP
;
2935 st_src_reg coord_w
= coord
;
2936 coord_w
.swizzle
= SWIZZLE_WWWW
;
2938 /* For the other TEX opcodes there's no projective version
2939 * since the last slot is taken up by LOD info. Do the
2940 * projective divide now.
2942 coord_dst
.writemask
= WRITEMASK_W
;
2943 emit(ir
, TGSI_OPCODE_RCP
, coord_dst
, projector
);
2945 /* In the case where we have to project the coordinates "by hand,"
2946 * the shadow comparator value must also be projected.
2948 st_src_reg tmp_src
= coord
;
2949 if (ir
->shadow_comparitor
) {
2950 /* Slot the shadow value in as the second to last component of the
2953 ir
->shadow_comparitor
->accept(this);
2955 tmp_src
= get_temp(glsl_type::vec4_type
);
2956 st_dst_reg tmp_dst
= st_dst_reg(tmp_src
);
2958 /* Projective division not allowed for array samplers. */
2959 assert(!sampler_type
->sampler_array
);
2961 tmp_dst
.writemask
= WRITEMASK_Z
;
2962 emit(ir
, TGSI_OPCODE_MOV
, tmp_dst
, this->result
);
2964 tmp_dst
.writemask
= WRITEMASK_XY
;
2965 emit(ir
, TGSI_OPCODE_MOV
, tmp_dst
, coord
);
2968 coord_dst
.writemask
= WRITEMASK_XYZ
;
2969 emit(ir
, TGSI_OPCODE_MUL
, coord_dst
, tmp_src
, coord_w
);
2971 coord_dst
.writemask
= WRITEMASK_XYZW
;
2972 coord
.swizzle
= SWIZZLE_XYZW
;
2976 /* If projection is done and the opcode is not TGSI_OPCODE_TXP, then the shadow
2977 * comparator was put in the correct place (and projected) by the code,
2978 * above, that handles by-hand projection.
2980 if (ir
->shadow_comparitor
&& (!ir
->projector
|| opcode
== TGSI_OPCODE_TXP
)) {
2981 /* Slot the shadow value in as the second to last component of the
2984 ir
->shadow_comparitor
->accept(this);
2986 if (is_cube_array
) {
2987 cube_sc
= get_temp(glsl_type::float_type
);
2988 cube_sc_dst
= st_dst_reg(cube_sc
);
2989 cube_sc_dst
.writemask
= WRITEMASK_X
;
2990 emit(ir
, TGSI_OPCODE_MOV
, cube_sc_dst
, this->result
);
2991 cube_sc_dst
.writemask
= WRITEMASK_X
;
2994 if ((sampler_type
->sampler_dimensionality
== GLSL_SAMPLER_DIM_2D
&&
2995 sampler_type
->sampler_array
) ||
2996 sampler_type
->sampler_dimensionality
== GLSL_SAMPLER_DIM_CUBE
) {
2997 coord_dst
.writemask
= WRITEMASK_W
;
2999 coord_dst
.writemask
= WRITEMASK_Z
;
3002 emit(ir
, TGSI_OPCODE_MOV
, coord_dst
, this->result
);
3003 coord_dst
.writemask
= WRITEMASK_XYZW
;
3007 if (ir
->op
== ir_txf_ms
) {
3008 coord_dst
.writemask
= WRITEMASK_W
;
3009 emit(ir
, TGSI_OPCODE_MOV
, coord_dst
, sample_index
);
3010 coord_dst
.writemask
= WRITEMASK_XYZW
;
3011 } else if (opcode
== TGSI_OPCODE_TXL
|| opcode
== TGSI_OPCODE_TXB
||
3012 opcode
== TGSI_OPCODE_TXF
) {
3013 /* TGSI stores LOD or LOD bias in the last channel of the coords. */
3014 coord_dst
.writemask
= WRITEMASK_W
;
3015 emit(ir
, TGSI_OPCODE_MOV
, coord_dst
, lod_info
);
3016 coord_dst
.writemask
= WRITEMASK_XYZW
;
3019 if (opcode
== TGSI_OPCODE_TXD
)
3020 inst
= emit(ir
, opcode
, result_dst
, coord
, dx
, dy
);
3021 else if (opcode
== TGSI_OPCODE_TXQ
) {
3022 if (ir
->op
== ir_query_levels
) {
3023 /* the level is stored in W */
3024 inst
= emit(ir
, opcode
, st_dst_reg(levels_src
), lod_info
);
3025 result_dst
.writemask
= WRITEMASK_X
;
3026 levels_src
.swizzle
= SWIZZLE_WWWW
;
3027 emit(ir
, TGSI_OPCODE_MOV
, result_dst
, levels_src
);
3029 inst
= emit(ir
, opcode
, result_dst
, lod_info
);
3030 } else if (opcode
== TGSI_OPCODE_TXF
) {
3031 inst
= emit(ir
, opcode
, result_dst
, coord
);
3032 } else if (opcode
== TGSI_OPCODE_TXL2
|| opcode
== TGSI_OPCODE_TXB2
) {
3033 inst
= emit(ir
, opcode
, result_dst
, coord
, lod_info
);
3034 } else if (opcode
== TGSI_OPCODE_TEX2
) {
3035 inst
= emit(ir
, opcode
, result_dst
, coord
, cube_sc
);
3036 } else if (opcode
== TGSI_OPCODE_TG4
) {
3037 if (is_cube_array
&& ir
->shadow_comparitor
) {
3038 inst
= emit(ir
, opcode
, result_dst
, coord
, cube_sc
);
3040 inst
= emit(ir
, opcode
, result_dst
, coord
, component
);
3043 inst
= emit(ir
, opcode
, result_dst
, coord
);
3045 if (ir
->shadow_comparitor
)
3046 inst
->tex_shadow
= GL_TRUE
;
3048 inst
->sampler
= _mesa_get_sampler_uniform_value(ir
->sampler
,
3049 this->shader_program
,
3053 for (i
= 0; i
< MAX_GLSL_TEXTURE_OFFSET
&& offset
[i
].file
!= PROGRAM_UNDEFINED
; i
++)
3054 inst
->tex_offsets
[i
] = offset
[i
];
3055 inst
->tex_offset_num_offset
= i
;
3058 switch (sampler_type
->sampler_dimensionality
) {
3059 case GLSL_SAMPLER_DIM_1D
:
3060 inst
->tex_target
= (sampler_type
->sampler_array
)
3061 ? TEXTURE_1D_ARRAY_INDEX
: TEXTURE_1D_INDEX
;
3063 case GLSL_SAMPLER_DIM_2D
:
3064 inst
->tex_target
= (sampler_type
->sampler_array
)
3065 ? TEXTURE_2D_ARRAY_INDEX
: TEXTURE_2D_INDEX
;
3067 case GLSL_SAMPLER_DIM_3D
:
3068 inst
->tex_target
= TEXTURE_3D_INDEX
;
3070 case GLSL_SAMPLER_DIM_CUBE
:
3071 inst
->tex_target
= (sampler_type
->sampler_array
)
3072 ? TEXTURE_CUBE_ARRAY_INDEX
: TEXTURE_CUBE_INDEX
;
3074 case GLSL_SAMPLER_DIM_RECT
:
3075 inst
->tex_target
= TEXTURE_RECT_INDEX
;
3077 case GLSL_SAMPLER_DIM_BUF
:
3078 inst
->tex_target
= TEXTURE_BUFFER_INDEX
;
3080 case GLSL_SAMPLER_DIM_EXTERNAL
:
3081 inst
->tex_target
= TEXTURE_EXTERNAL_INDEX
;
3083 case GLSL_SAMPLER_DIM_MS
:
3084 inst
->tex_target
= (sampler_type
->sampler_array
)
3085 ? TEXTURE_2D_MULTISAMPLE_ARRAY_INDEX
: TEXTURE_2D_MULTISAMPLE_INDEX
;
3088 assert(!"Should not get here.");
3091 this->result
= result_src
;
3095 glsl_to_tgsi_visitor::visit(ir_return
*ir
)
3097 if (ir
->get_value()) {
3101 assert(current_function
);
3103 ir
->get_value()->accept(this);
3104 st_src_reg r
= this->result
;
3106 l
= st_dst_reg(current_function
->return_reg
);
3108 for (i
= 0; i
< type_size(current_function
->sig
->return_type
); i
++) {
3109 emit(ir
, TGSI_OPCODE_MOV
, l
, r
);
3115 emit(ir
, TGSI_OPCODE_RET
);
3119 glsl_to_tgsi_visitor::visit(ir_discard
*ir
)
3121 if (ir
->condition
) {
3122 ir
->condition
->accept(this);
3123 this->result
.negate
= ~this->result
.negate
;
3124 emit(ir
, TGSI_OPCODE_KILL_IF
, undef_dst
, this->result
);
3126 /* unconditional kil */
3127 emit(ir
, TGSI_OPCODE_KILL
);
3132 glsl_to_tgsi_visitor::visit(ir_if
*ir
)
3135 glsl_to_tgsi_instruction
*if_inst
;
3137 ir
->condition
->accept(this);
3138 assert(this->result
.file
!= PROGRAM_UNDEFINED
);
3140 if_opcode
= native_integers
? TGSI_OPCODE_UIF
: TGSI_OPCODE_IF
;
3142 if_inst
= emit(ir
->condition
, if_opcode
, undef_dst
, this->result
);
3144 this->instructions
.push_tail(if_inst
);
3146 visit_exec_list(&ir
->then_instructions
, this);
3148 if (!ir
->else_instructions
.is_empty()) {
3149 emit(ir
->condition
, TGSI_OPCODE_ELSE
);
3150 visit_exec_list(&ir
->else_instructions
, this);
3153 if_inst
= emit(ir
->condition
, TGSI_OPCODE_ENDIF
);
3158 glsl_to_tgsi_visitor::visit(ir_emit_vertex
*ir
)
3160 assert(this->prog
->Target
== GL_GEOMETRY_PROGRAM_NV
);
3162 ir
->stream
->accept(this);
3163 emit(ir
, TGSI_OPCODE_EMIT
, undef_dst
, this->result
);
3167 glsl_to_tgsi_visitor::visit(ir_end_primitive
*ir
)
3169 assert(this->prog
->Target
== GL_GEOMETRY_PROGRAM_NV
);
3171 ir
->stream
->accept(this);
3172 emit(ir
, TGSI_OPCODE_ENDPRIM
, undef_dst
, this->result
);
3175 glsl_to_tgsi_visitor::glsl_to_tgsi_visitor()
3177 result
.file
= PROGRAM_UNDEFINED
;
3180 next_signature_id
= 1;
3182 current_function
= NULL
;
3183 num_address_regs
= 0;
3185 indirect_addr_consts
= false;
3187 native_integers
= false;
3188 mem_ctx
= ralloc_context(NULL
);
3191 shader_program
= NULL
;
3196 glsl_to_tgsi_visitor::~glsl_to_tgsi_visitor()
3198 ralloc_free(mem_ctx
);
3201 extern "C" void free_glsl_to_tgsi_visitor(glsl_to_tgsi_visitor
*v
)
3208 * Count resources used by the given gpu program (number of texture
3212 count_resources(glsl_to_tgsi_visitor
*v
, gl_program
*prog
)
3214 v
->samplers_used
= 0;
3216 foreach_in_list(glsl_to_tgsi_instruction
, inst
, &v
->instructions
) {
3217 if (is_tex_instruction(inst
->op
)) {
3218 v
->samplers_used
|= 1 << inst
->sampler
;
3220 if (inst
->tex_shadow
) {
3221 prog
->ShadowSamplers
|= 1 << inst
->sampler
;
3226 prog
->SamplersUsed
= v
->samplers_used
;
3228 if (v
->shader_program
!= NULL
)
3229 _mesa_update_shader_textures_used(v
->shader_program
, prog
);
3233 set_uniform_initializer(struct gl_context
*ctx
, void *mem_ctx
,
3234 struct gl_shader_program
*shader_program
,
3235 const char *name
, const glsl_type
*type
,
3238 if (type
->is_record()) {
3239 ir_constant
*field_constant
;
3241 field_constant
= (ir_constant
*)val
->components
.get_head();
3243 for (unsigned int i
= 0; i
< type
->length
; i
++) {
3244 const glsl_type
*field_type
= type
->fields
.structure
[i
].type
;
3245 const char *field_name
= ralloc_asprintf(mem_ctx
, "%s.%s", name
,
3246 type
->fields
.structure
[i
].name
);
3247 set_uniform_initializer(ctx
, mem_ctx
, shader_program
, field_name
,
3248 field_type
, field_constant
);
3249 field_constant
= (ir_constant
*)field_constant
->next
;
3255 unsigned index
= _mesa_get_uniform_location(ctx
, shader_program
, name
,
3257 if (offset
== GL_INVALID_INDEX
) {
3258 fail_link(shader_program
,
3259 "Couldn't find uniform for initializer %s\n", name
);
3262 int loc
= _mesa_uniform_merge_location_offset(shader_program
, index
, offset
);
3264 for (unsigned int i
= 0; i
< (type
->is_array() ? type
->length
: 1); i
++) {
3265 ir_constant
*element
;
3266 const glsl_type
*element_type
;
3267 if (type
->is_array()) {
3268 element
= val
->array_elements
[i
];
3269 element_type
= type
->fields
.array
;
3272 element_type
= type
;
3277 if (element_type
->base_type
== GLSL_TYPE_BOOL
) {
3278 int *conv
= ralloc_array(mem_ctx
, int, element_type
->components());
3279 for (unsigned int j
= 0; j
< element_type
->components(); j
++) {
3280 conv
[j
] = element
->value
.b
[j
];
3282 values
= (void *)conv
;
3283 element_type
= glsl_type::get_instance(GLSL_TYPE_INT
,
3284 element_type
->vector_elements
,
3287 values
= &element
->value
;
3290 if (element_type
->is_matrix()) {
3291 _mesa_uniform_matrix(ctx
, shader_program
,
3292 element_type
->matrix_columns
,
3293 element_type
->vector_elements
,
3294 loc
, 1, GL_FALSE
, (GLfloat
*)values
);
3296 _mesa_uniform(ctx
, shader_program
, loc
, element_type
->matrix_columns
,
3297 values
, element_type
->gl_type
);
3305 * Returns the mask of channels (bitmask of WRITEMASK_X,Y,Z,W) which
3306 * are read from the given src in this instruction
3309 get_src_arg_mask(st_dst_reg dst
, st_src_reg src
)
3311 int read_mask
= 0, comp
;
3313 /* Now, given the src swizzle and the written channels, find which
3314 * components are actually read
3316 for (comp
= 0; comp
< 4; ++comp
) {
3317 const unsigned coord
= GET_SWZ(src
.swizzle
, comp
);
3319 if (dst
.writemask
& (1 << comp
) && coord
<= SWIZZLE_W
)
3320 read_mask
|= 1 << coord
;
3327 * This pass replaces CMP T0, T1 T2 T0 with MOV T0, T2 when the CMP
3328 * instruction is the first instruction to write to register T0. There are
3329 * several lowering passes done in GLSL IR (e.g. branches and
3330 * relative addressing) that create a large number of conditional assignments
3331 * that ir_to_mesa converts to CMP instructions like the one mentioned above.
3333 * Here is why this conversion is safe:
3334 * CMP T0, T1 T2 T0 can be expanded to:
3340 * If (T1 < 0.0) evaluates to true then our replacement MOV T0, T2 is the same
3341 * as the original program. If (T1 < 0.0) evaluates to false, executing
3342 * MOV T0, T0 will store a garbage value in T0 since T0 is uninitialized.
3343 * Therefore, it doesn't matter that we are replacing MOV T0, T0 with MOV T0, T2
3344 * because any instruction that was going to read from T0 after this was going
3345 * to read a garbage value anyway.
3348 glsl_to_tgsi_visitor::simplify_cmp(void)
3350 unsigned *tempWrites
;
3351 unsigned outputWrites
[MAX_PROGRAM_OUTPUTS
];
3353 tempWrites
= new unsigned[MAX_TEMPS
];
3357 memset(tempWrites
, 0, sizeof(unsigned) * MAX_TEMPS
);
3358 memset(outputWrites
, 0, sizeof(outputWrites
));
3360 foreach_in_list(glsl_to_tgsi_instruction
, inst
, &this->instructions
) {
3361 unsigned prevWriteMask
= 0;
3363 /* Give up if we encounter relative addressing or flow control. */
3364 if (inst
->dst
.reladdr
||
3365 tgsi_get_opcode_info(inst
->op
)->is_branch
||
3366 inst
->op
== TGSI_OPCODE_BGNSUB
||
3367 inst
->op
== TGSI_OPCODE_CONT
||
3368 inst
->op
== TGSI_OPCODE_END
||
3369 inst
->op
== TGSI_OPCODE_ENDSUB
||
3370 inst
->op
== TGSI_OPCODE_RET
) {
3374 if (inst
->dst
.file
== PROGRAM_OUTPUT
) {
3375 assert(inst
->dst
.index
< MAX_PROGRAM_OUTPUTS
);
3376 prevWriteMask
= outputWrites
[inst
->dst
.index
];
3377 outputWrites
[inst
->dst
.index
] |= inst
->dst
.writemask
;
3378 } else if (inst
->dst
.file
== PROGRAM_TEMPORARY
) {
3379 assert(inst
->dst
.index
< MAX_TEMPS
);
3380 prevWriteMask
= tempWrites
[inst
->dst
.index
];
3381 tempWrites
[inst
->dst
.index
] |= inst
->dst
.writemask
;
3385 /* For a CMP to be considered a conditional write, the destination
3386 * register and source register two must be the same. */
3387 if (inst
->op
== TGSI_OPCODE_CMP
3388 && !(inst
->dst
.writemask
& prevWriteMask
)
3389 && inst
->src
[2].file
== inst
->dst
.file
3390 && inst
->src
[2].index
== inst
->dst
.index
3391 && inst
->dst
.writemask
== get_src_arg_mask(inst
->dst
, inst
->src
[2])) {
3393 inst
->op
= TGSI_OPCODE_MOV
;
3394 inst
->src
[0] = inst
->src
[1];
3398 delete [] tempWrites
;
3401 /* Replaces all references to a temporary register index with another index. */
3403 glsl_to_tgsi_visitor::rename_temp_register(int index
, int new_index
)
3405 foreach_in_list(glsl_to_tgsi_instruction
, inst
, &this->instructions
) {
3408 for (j
=0; j
< num_inst_src_regs(inst
->op
); j
++) {
3409 if (inst
->src
[j
].file
== PROGRAM_TEMPORARY
&&
3410 inst
->src
[j
].index
== index
) {
3411 inst
->src
[j
].index
= new_index
;
3415 for (j
=0; j
< inst
->tex_offset_num_offset
; j
++) {
3416 if (inst
->tex_offsets
[j
].file
== PROGRAM_TEMPORARY
&&
3417 inst
->tex_offsets
[j
].index
== index
) {
3418 inst
->tex_offsets
[j
].index
= new_index
;
3422 if (inst
->dst
.file
== PROGRAM_TEMPORARY
&& inst
->dst
.index
== index
) {
3423 inst
->dst
.index
= new_index
;
3429 glsl_to_tgsi_visitor::get_first_temp_read(int index
)
3431 int depth
= 0; /* loop depth */
3432 int loop_start
= -1; /* index of the first active BGNLOOP (if any) */
3435 foreach_in_list(glsl_to_tgsi_instruction
, inst
, &this->instructions
) {
3436 for (j
=0; j
< num_inst_src_regs(inst
->op
); j
++) {
3437 if (inst
->src
[j
].file
== PROGRAM_TEMPORARY
&&
3438 inst
->src
[j
].index
== index
) {
3439 return (depth
== 0) ? i
: loop_start
;
3442 for (j
=0; j
< inst
->tex_offset_num_offset
; j
++) {
3443 if (inst
->tex_offsets
[j
].file
== PROGRAM_TEMPORARY
&&
3444 inst
->tex_offsets
[j
].index
== index
) {
3445 return (depth
== 0) ? i
: loop_start
;
3449 if (inst
->op
== TGSI_OPCODE_BGNLOOP
) {
3452 } else if (inst
->op
== TGSI_OPCODE_ENDLOOP
) {
3465 glsl_to_tgsi_visitor::get_first_temp_write(int index
)
3467 int depth
= 0; /* loop depth */
3468 int loop_start
= -1; /* index of the first active BGNLOOP (if any) */
3471 foreach_in_list(glsl_to_tgsi_instruction
, inst
, &this->instructions
) {
3472 if (inst
->dst
.file
== PROGRAM_TEMPORARY
&& inst
->dst
.index
== index
) {
3473 return (depth
== 0) ? i
: loop_start
;
3476 if (inst
->op
== TGSI_OPCODE_BGNLOOP
) {
3479 } else if (inst
->op
== TGSI_OPCODE_ENDLOOP
) {
3492 glsl_to_tgsi_visitor::get_last_temp_read(int index
)
3494 int depth
= 0; /* loop depth */
3495 int last
= -1; /* index of last instruction that reads the temporary */
3498 foreach_in_list(glsl_to_tgsi_instruction
, inst
, &this->instructions
) {
3499 for (j
=0; j
< num_inst_src_regs(inst
->op
); j
++) {
3500 if (inst
->src
[j
].file
== PROGRAM_TEMPORARY
&&
3501 inst
->src
[j
].index
== index
) {
3502 last
= (depth
== 0) ? i
: -2;
3505 for (j
=0; j
< inst
->tex_offset_num_offset
; j
++) {
3506 if (inst
->tex_offsets
[j
].file
== PROGRAM_TEMPORARY
&&
3507 inst
->tex_offsets
[j
].index
== index
)
3508 last
= (depth
== 0) ? i
: -2;
3511 if (inst
->op
== TGSI_OPCODE_BGNLOOP
)
3513 else if (inst
->op
== TGSI_OPCODE_ENDLOOP
)
3514 if (--depth
== 0 && last
== -2)
3526 glsl_to_tgsi_visitor::get_last_temp_write(int index
)
3528 int depth
= 0; /* loop depth */
3529 int last
= -1; /* index of last instruction that writes to the temporary */
3532 foreach_in_list(glsl_to_tgsi_instruction
, inst
, &this->instructions
) {
3533 if (inst
->dst
.file
== PROGRAM_TEMPORARY
&& inst
->dst
.index
== index
)
3534 last
= (depth
== 0) ? i
: -2;
3536 if (inst
->op
== TGSI_OPCODE_BGNLOOP
)
3538 else if (inst
->op
== TGSI_OPCODE_ENDLOOP
)
3539 if (--depth
== 0 && last
== -2)
3551 * On a basic block basis, tracks available PROGRAM_TEMPORARY register
3552 * channels for copy propagation and updates following instructions to
3553 * use the original versions.
3555 * The glsl_to_tgsi_visitor lazily produces code assuming that this pass
3556 * will occur. As an example, a TXP production before this pass:
3558 * 0: MOV TEMP[1], INPUT[4].xyyy;
3559 * 1: MOV TEMP[1].w, INPUT[4].wwww;
3560 * 2: TXP TEMP[2], TEMP[1], texture[0], 2D;
3564 * 0: MOV TEMP[1], INPUT[4].xyyy;
3565 * 1: MOV TEMP[1].w, INPUT[4].wwww;
3566 * 2: TXP TEMP[2], INPUT[4].xyyw, texture[0], 2D;
3568 * which allows for dead code elimination on TEMP[1]'s writes.
3571 glsl_to_tgsi_visitor::copy_propagate(void)
3573 glsl_to_tgsi_instruction
**acp
= rzalloc_array(mem_ctx
,
3574 glsl_to_tgsi_instruction
*,
3575 this->next_temp
* 4);
3576 int *acp_level
= rzalloc_array(mem_ctx
, int, this->next_temp
* 4);
3579 foreach_in_list(glsl_to_tgsi_instruction
, inst
, &this->instructions
) {
3580 assert(inst
->dst
.file
!= PROGRAM_TEMPORARY
3581 || inst
->dst
.index
< this->next_temp
);
3583 /* First, do any copy propagation possible into the src regs. */
3584 for (int r
= 0; r
< 3; r
++) {
3585 glsl_to_tgsi_instruction
*first
= NULL
;
3587 int acp_base
= inst
->src
[r
].index
* 4;
3589 if (inst
->src
[r
].file
!= PROGRAM_TEMPORARY
||
3590 inst
->src
[r
].reladdr
||
3591 inst
->src
[r
].reladdr2
)
3594 /* See if we can find entries in the ACP consisting of MOVs
3595 * from the same src register for all the swizzled channels
3596 * of this src register reference.
3598 for (int i
= 0; i
< 4; i
++) {
3599 int src_chan
= GET_SWZ(inst
->src
[r
].swizzle
, i
);
3600 glsl_to_tgsi_instruction
*copy_chan
= acp
[acp_base
+ src_chan
];
3607 assert(acp_level
[acp_base
+ src_chan
] <= level
);
3612 if (first
->src
[0].file
!= copy_chan
->src
[0].file
||
3613 first
->src
[0].index
!= copy_chan
->src
[0].index
) {
3621 /* We've now validated that we can copy-propagate to
3622 * replace this src register reference. Do it.
3624 inst
->src
[r
].file
= first
->src
[0].file
;
3625 inst
->src
[r
].index
= first
->src
[0].index
;
3626 inst
->src
[r
].index2D
= first
->src
[0].index2D
;
3627 inst
->src
[r
].has_index2
= first
->src
[0].has_index2
;
3630 for (int i
= 0; i
< 4; i
++) {
3631 int src_chan
= GET_SWZ(inst
->src
[r
].swizzle
, i
);
3632 glsl_to_tgsi_instruction
*copy_inst
= acp
[acp_base
+ src_chan
];
3633 swizzle
|= (GET_SWZ(copy_inst
->src
[0].swizzle
, src_chan
) <<
3636 inst
->src
[r
].swizzle
= swizzle
;
3641 case TGSI_OPCODE_BGNLOOP
:
3642 case TGSI_OPCODE_ENDLOOP
:
3643 /* End of a basic block, clear the ACP entirely. */
3644 memset(acp
, 0, sizeof(*acp
) * this->next_temp
* 4);
3647 case TGSI_OPCODE_IF
:
3648 case TGSI_OPCODE_UIF
:
3652 case TGSI_OPCODE_ENDIF
:
3653 case TGSI_OPCODE_ELSE
:
3654 /* Clear all channels written inside the block from the ACP, but
3655 * leaving those that were not touched.
3657 for (int r
= 0; r
< this->next_temp
; r
++) {
3658 for (int c
= 0; c
< 4; c
++) {
3659 if (!acp
[4 * r
+ c
])
3662 if (acp_level
[4 * r
+ c
] >= level
)
3663 acp
[4 * r
+ c
] = NULL
;
3666 if (inst
->op
== TGSI_OPCODE_ENDIF
)
3671 /* Continuing the block, clear any written channels from
3674 if (inst
->dst
.file
== PROGRAM_TEMPORARY
&& inst
->dst
.reladdr
) {
3675 /* Any temporary might be written, so no copy propagation
3676 * across this instruction.
3678 memset(acp
, 0, sizeof(*acp
) * this->next_temp
* 4);
3679 } else if (inst
->dst
.file
== PROGRAM_OUTPUT
&&
3680 inst
->dst
.reladdr
) {
3681 /* Any output might be written, so no copy propagation
3682 * from outputs across this instruction.
3684 for (int r
= 0; r
< this->next_temp
; r
++) {
3685 for (int c
= 0; c
< 4; c
++) {
3686 if (!acp
[4 * r
+ c
])
3689 if (acp
[4 * r
+ c
]->src
[0].file
== PROGRAM_OUTPUT
)
3690 acp
[4 * r
+ c
] = NULL
;
3693 } else if (inst
->dst
.file
== PROGRAM_TEMPORARY
||
3694 inst
->dst
.file
== PROGRAM_OUTPUT
) {
3695 /* Clear where it's used as dst. */
3696 if (inst
->dst
.file
== PROGRAM_TEMPORARY
) {
3697 for (int c
= 0; c
< 4; c
++) {
3698 if (inst
->dst
.writemask
& (1 << c
)) {
3699 acp
[4 * inst
->dst
.index
+ c
] = NULL
;
3704 /* Clear where it's used as src. */
3705 for (int r
= 0; r
< this->next_temp
; r
++) {
3706 for (int c
= 0; c
< 4; c
++) {
3707 if (!acp
[4 * r
+ c
])
3710 int src_chan
= GET_SWZ(acp
[4 * r
+ c
]->src
[0].swizzle
, c
);
3712 if (acp
[4 * r
+ c
]->src
[0].file
== inst
->dst
.file
&&
3713 acp
[4 * r
+ c
]->src
[0].index
== inst
->dst
.index
&&
3714 inst
->dst
.writemask
& (1 << src_chan
))
3716 acp
[4 * r
+ c
] = NULL
;
3724 /* If this is a copy, add it to the ACP. */
3725 if (inst
->op
== TGSI_OPCODE_MOV
&&
3726 inst
->dst
.file
== PROGRAM_TEMPORARY
&&
3727 !(inst
->dst
.file
== inst
->src
[0].file
&&
3728 inst
->dst
.index
== inst
->src
[0].index
) &&
3729 !inst
->dst
.reladdr
&&
3731 !inst
->src
[0].reladdr
&&
3732 !inst
->src
[0].reladdr2
&&
3733 !inst
->src
[0].negate
) {
3734 for (int i
= 0; i
< 4; i
++) {
3735 if (inst
->dst
.writemask
& (1 << i
)) {
3736 acp
[4 * inst
->dst
.index
+ i
] = inst
;
3737 acp_level
[4 * inst
->dst
.index
+ i
] = level
;
3743 ralloc_free(acp_level
);
3748 * On a basic block basis, tracks available PROGRAM_TEMPORARY registers for dead
3751 * The glsl_to_tgsi_visitor lazily produces code assuming that this pass
3752 * will occur. As an example, a TXP production after copy propagation but
3755 * 0: MOV TEMP[1], INPUT[4].xyyy;
3756 * 1: MOV TEMP[1].w, INPUT[4].wwww;
3757 * 2: TXP TEMP[2], INPUT[4].xyyw, texture[0], 2D;
3759 * and after this pass:
3761 * 0: TXP TEMP[2], INPUT[4].xyyw, texture[0], 2D;
3764 glsl_to_tgsi_visitor::eliminate_dead_code(void)
3766 glsl_to_tgsi_instruction
**writes
= rzalloc_array(mem_ctx
,
3767 glsl_to_tgsi_instruction
*,
3768 this->next_temp
* 4);
3769 int *write_level
= rzalloc_array(mem_ctx
, int, this->next_temp
* 4);
3773 foreach_in_list(glsl_to_tgsi_instruction
, inst
, &this->instructions
) {
3774 assert(inst
->dst
.file
!= PROGRAM_TEMPORARY
3775 || inst
->dst
.index
< this->next_temp
);
3778 case TGSI_OPCODE_BGNLOOP
:
3779 case TGSI_OPCODE_ENDLOOP
:
3780 case TGSI_OPCODE_CONT
:
3781 case TGSI_OPCODE_BRK
:
3782 /* End of a basic block, clear the write array entirely.
3784 * This keeps us from killing dead code when the writes are
3785 * on either side of a loop, even when the register isn't touched
3786 * inside the loop. However, glsl_to_tgsi_visitor doesn't seem to emit
3787 * dead code of this type, so it shouldn't make a difference as long as
3788 * the dead code elimination pass in the GLSL compiler does its job.
3790 memset(writes
, 0, sizeof(*writes
) * this->next_temp
* 4);
3793 case TGSI_OPCODE_ENDIF
:
3794 case TGSI_OPCODE_ELSE
:
3795 /* Promote the recorded level of all channels written inside the
3796 * preceding if or else block to the level above the if/else block.
3798 for (int r
= 0; r
< this->next_temp
; r
++) {
3799 for (int c
= 0; c
< 4; c
++) {
3800 if (!writes
[4 * r
+ c
])
3803 if (write_level
[4 * r
+ c
] == level
)
3804 write_level
[4 * r
+ c
] = level
-1;
3808 if(inst
->op
== TGSI_OPCODE_ENDIF
)
3813 case TGSI_OPCODE_IF
:
3814 case TGSI_OPCODE_UIF
:
3816 /* fallthrough to default case to mark the condition as read */
3819 /* Continuing the block, clear any channels from the write array that
3820 * are read by this instruction.
3822 for (unsigned i
= 0; i
< Elements(inst
->src
); i
++) {
3823 if (inst
->src
[i
].file
== PROGRAM_TEMPORARY
&& inst
->src
[i
].reladdr
){
3824 /* Any temporary might be read, so no dead code elimination
3825 * across this instruction.
3827 memset(writes
, 0, sizeof(*writes
) * this->next_temp
* 4);
3828 } else if (inst
->src
[i
].file
== PROGRAM_TEMPORARY
) {
3829 /* Clear where it's used as src. */
3830 int src_chans
= 1 << GET_SWZ(inst
->src
[i
].swizzle
, 0);
3831 src_chans
|= 1 << GET_SWZ(inst
->src
[i
].swizzle
, 1);
3832 src_chans
|= 1 << GET_SWZ(inst
->src
[i
].swizzle
, 2);
3833 src_chans
|= 1 << GET_SWZ(inst
->src
[i
].swizzle
, 3);
3835 for (int c
= 0; c
< 4; c
++) {
3836 if (src_chans
& (1 << c
)) {
3837 writes
[4 * inst
->src
[i
].index
+ c
] = NULL
;
3842 for (unsigned i
= 0; i
< inst
->tex_offset_num_offset
; i
++) {
3843 if (inst
->tex_offsets
[i
].file
== PROGRAM_TEMPORARY
&& inst
->tex_offsets
[i
].reladdr
){
3844 /* Any temporary might be read, so no dead code elimination
3845 * across this instruction.
3847 memset(writes
, 0, sizeof(*writes
) * this->next_temp
* 4);
3848 } else if (inst
->tex_offsets
[i
].file
== PROGRAM_TEMPORARY
) {
3849 /* Clear where it's used as src. */
3850 int src_chans
= 1 << GET_SWZ(inst
->tex_offsets
[i
].swizzle
, 0);
3851 src_chans
|= 1 << GET_SWZ(inst
->tex_offsets
[i
].swizzle
, 1);
3852 src_chans
|= 1 << GET_SWZ(inst
->tex_offsets
[i
].swizzle
, 2);
3853 src_chans
|= 1 << GET_SWZ(inst
->tex_offsets
[i
].swizzle
, 3);
3855 for (int c
= 0; c
< 4; c
++) {
3856 if (src_chans
& (1 << c
)) {
3857 writes
[4 * inst
->tex_offsets
[i
].index
+ c
] = NULL
;
3865 /* If this instruction writes to a temporary, add it to the write array.
3866 * If there is already an instruction in the write array for one or more
3867 * of the channels, flag that channel write as dead.
3869 if (inst
->dst
.file
== PROGRAM_TEMPORARY
&&
3870 !inst
->dst
.reladdr
&&
3872 for (int c
= 0; c
< 4; c
++) {
3873 if (inst
->dst
.writemask
& (1 << c
)) {
3874 if (writes
[4 * inst
->dst
.index
+ c
]) {
3875 if (write_level
[4 * inst
->dst
.index
+ c
] < level
)
3878 writes
[4 * inst
->dst
.index
+ c
]->dead_mask
|= (1 << c
);
3880 writes
[4 * inst
->dst
.index
+ c
] = inst
;
3881 write_level
[4 * inst
->dst
.index
+ c
] = level
;
3887 /* Anything still in the write array at this point is dead code. */
3888 for (int r
= 0; r
< this->next_temp
; r
++) {
3889 for (int c
= 0; c
< 4; c
++) {
3890 glsl_to_tgsi_instruction
*inst
= writes
[4 * r
+ c
];
3892 inst
->dead_mask
|= (1 << c
);
3896 /* Now actually remove the instructions that are completely dead and update
3897 * the writemask of other instructions with dead channels.
3899 foreach_in_list_safe(glsl_to_tgsi_instruction
, inst
, &this->instructions
) {
3900 if (!inst
->dead_mask
|| !inst
->dst
.writemask
)
3902 else if ((inst
->dst
.writemask
& ~inst
->dead_mask
) == 0) {
3907 inst
->dst
.writemask
&= ~(inst
->dead_mask
);
3910 ralloc_free(write_level
);
3911 ralloc_free(writes
);
3916 /* Merges temporary registers together where possible to reduce the number of
3917 * registers needed to run a program.
3919 * Produces optimal code only after copy propagation and dead code elimination
3922 glsl_to_tgsi_visitor::merge_registers(void)
3924 int *last_reads
= rzalloc_array(mem_ctx
, int, this->next_temp
);
3925 int *first_writes
= rzalloc_array(mem_ctx
, int, this->next_temp
);
3928 /* Read the indices of the last read and first write to each temp register
3929 * into an array so that we don't have to traverse the instruction list as
3931 for (i
=0; i
< this->next_temp
; i
++) {
3932 last_reads
[i
] = get_last_temp_read(i
);
3933 first_writes
[i
] = get_first_temp_write(i
);
3936 /* Start looking for registers with non-overlapping usages that can be
3937 * merged together. */
3938 for (i
=0; i
< this->next_temp
; i
++) {
3939 /* Don't touch unused registers. */
3940 if (last_reads
[i
] < 0 || first_writes
[i
] < 0) continue;
3942 for (j
=0; j
< this->next_temp
; j
++) {
3943 /* Don't touch unused registers. */
3944 if (last_reads
[j
] < 0 || first_writes
[j
] < 0) continue;
3946 /* We can merge the two registers if the first write to j is after or
3947 * in the same instruction as the last read from i. Note that the
3948 * register at index i will always be used earlier or at the same time
3949 * as the register at index j. */
3950 if (first_writes
[i
] <= first_writes
[j
] &&
3951 last_reads
[i
] <= first_writes
[j
])
3953 rename_temp_register(j
, i
); /* Replace all references to j with i.*/
3955 /* Update the first_writes and last_reads arrays with the new
3956 * values for the merged register index, and mark the newly unused
3957 * register index as such. */
3958 last_reads
[i
] = last_reads
[j
];
3959 first_writes
[j
] = -1;
3965 ralloc_free(last_reads
);
3966 ralloc_free(first_writes
);
3969 /* Reassign indices to temporary registers by reusing unused indices created
3970 * by optimization passes. */
3972 glsl_to_tgsi_visitor::renumber_registers(void)
3977 for (i
=0; i
< this->next_temp
; i
++) {
3978 if (get_first_temp_read(i
) < 0) continue;
3980 rename_temp_register(i
, new_index
);
3984 this->next_temp
= new_index
;
3988 * Returns a fragment program which implements the current pixel transfer ops.
3989 * Based on get_pixel_transfer_program in st_atom_pixeltransfer.c.
3992 get_pixel_transfer_visitor(struct st_fragment_program
*fp
,
3993 glsl_to_tgsi_visitor
*original
,
3994 int scale_and_bias
, int pixel_maps
)
3996 glsl_to_tgsi_visitor
*v
= new glsl_to_tgsi_visitor();
3997 struct st_context
*st
= st_context(original
->ctx
);
3998 struct gl_program
*prog
= &fp
->Base
.Base
;
3999 struct gl_program_parameter_list
*params
= _mesa_new_parameter_list();
4000 st_src_reg coord
, src0
;
4002 glsl_to_tgsi_instruction
*inst
;
4004 /* Copy attributes of the glsl_to_tgsi_visitor in the original shader. */
4005 v
->ctx
= original
->ctx
;
4007 v
->shader_program
= NULL
;
4009 v
->glsl_version
= original
->glsl_version
;
4010 v
->native_integers
= original
->native_integers
;
4011 v
->options
= original
->options
;
4012 v
->next_temp
= original
->next_temp
;
4013 v
->num_address_regs
= original
->num_address_regs
;
4014 v
->samplers_used
= prog
->SamplersUsed
= original
->samplers_used
;
4015 v
->indirect_addr_consts
= original
->indirect_addr_consts
;
4016 memcpy(&v
->immediates
, &original
->immediates
, sizeof(v
->immediates
));
4017 v
->num_immediates
= original
->num_immediates
;
4020 * Get initial pixel color from the texture.
4021 * TEX colorTemp, fragment.texcoord[0], texture[0], 2D;
4023 coord
= st_src_reg(PROGRAM_INPUT
, VARYING_SLOT_TEX0
, glsl_type::vec2_type
);
4024 src0
= v
->get_temp(glsl_type::vec4_type
);
4025 dst0
= st_dst_reg(src0
);
4026 inst
= v
->emit(NULL
, TGSI_OPCODE_TEX
, dst0
, coord
);
4028 inst
->tex_target
= TEXTURE_2D_INDEX
;
4030 prog
->InputsRead
|= VARYING_BIT_TEX0
;
4031 prog
->SamplersUsed
|= (1 << 0); /* mark sampler 0 as used */
4032 v
->samplers_used
|= (1 << 0);
4034 if (scale_and_bias
) {
4035 static const gl_state_index scale_state
[STATE_LENGTH
] =
4036 { STATE_INTERNAL
, STATE_PT_SCALE
,
4037 (gl_state_index
) 0, (gl_state_index
) 0, (gl_state_index
) 0 };
4038 static const gl_state_index bias_state
[STATE_LENGTH
] =
4039 { STATE_INTERNAL
, STATE_PT_BIAS
,
4040 (gl_state_index
) 0, (gl_state_index
) 0, (gl_state_index
) 0 };
4041 GLint scale_p
, bias_p
;
4042 st_src_reg scale
, bias
;
4044 scale_p
= _mesa_add_state_reference(params
, scale_state
);
4045 bias_p
= _mesa_add_state_reference(params
, bias_state
);
4047 /* MAD colorTemp, colorTemp, scale, bias; */
4048 scale
= st_src_reg(PROGRAM_STATE_VAR
, scale_p
, GLSL_TYPE_FLOAT
);
4049 bias
= st_src_reg(PROGRAM_STATE_VAR
, bias_p
, GLSL_TYPE_FLOAT
);
4050 inst
= v
->emit(NULL
, TGSI_OPCODE_MAD
, dst0
, src0
, scale
, bias
);
4054 st_src_reg temp
= v
->get_temp(glsl_type::vec4_type
);
4055 st_dst_reg temp_dst
= st_dst_reg(temp
);
4057 assert(st
->pixel_xfer
.pixelmap_texture
);
4059 /* With a little effort, we can do four pixel map look-ups with
4060 * two TEX instructions:
4063 /* TEX temp.rg, colorTemp.rgba, texture[1], 2D; */
4064 temp_dst
.writemask
= WRITEMASK_XY
; /* write R,G */
4065 inst
= v
->emit(NULL
, TGSI_OPCODE_TEX
, temp_dst
, src0
);
4067 inst
->tex_target
= TEXTURE_2D_INDEX
;
4069 /* TEX temp.ba, colorTemp.baba, texture[1], 2D; */
4070 src0
.swizzle
= MAKE_SWIZZLE4(SWIZZLE_Z
, SWIZZLE_W
, SWIZZLE_Z
, SWIZZLE_W
);
4071 temp_dst
.writemask
= WRITEMASK_ZW
; /* write B,A */
4072 inst
= v
->emit(NULL
, TGSI_OPCODE_TEX
, temp_dst
, src0
);
4074 inst
->tex_target
= TEXTURE_2D_INDEX
;
4076 prog
->SamplersUsed
|= (1 << 1); /* mark sampler 1 as used */
4077 v
->samplers_used
|= (1 << 1);
4079 /* MOV colorTemp, temp; */
4080 inst
= v
->emit(NULL
, TGSI_OPCODE_MOV
, dst0
, temp
);
4083 /* Now copy the instructions from the original glsl_to_tgsi_visitor into the
4085 foreach_in_list(glsl_to_tgsi_instruction
, inst
, &original
->instructions
) {
4086 glsl_to_tgsi_instruction
*newinst
;
4087 st_src_reg src_regs
[3];
4089 if (inst
->dst
.file
== PROGRAM_OUTPUT
)
4090 prog
->OutputsWritten
|= BITFIELD64_BIT(inst
->dst
.index
);
4092 for (int i
=0; i
<3; i
++) {
4093 src_regs
[i
] = inst
->src
[i
];
4094 if (src_regs
[i
].file
== PROGRAM_INPUT
&&
4095 src_regs
[i
].index
== VARYING_SLOT_COL0
)
4097 src_regs
[i
].file
= PROGRAM_TEMPORARY
;
4098 src_regs
[i
].index
= src0
.index
;
4100 else if (src_regs
[i
].file
== PROGRAM_INPUT
)
4101 prog
->InputsRead
|= BITFIELD64_BIT(src_regs
[i
].index
);
4104 newinst
= v
->emit(NULL
, inst
->op
, inst
->dst
, src_regs
[0], src_regs
[1], src_regs
[2]);
4105 newinst
->tex_target
= inst
->tex_target
;
4108 /* Make modifications to fragment program info. */
4109 prog
->Parameters
= _mesa_combine_parameter_lists(params
,
4110 original
->prog
->Parameters
);
4111 _mesa_free_parameter_list(params
);
4112 count_resources(v
, prog
);
4113 fp
->glsl_to_tgsi
= v
;
4117 * Make fragment program for glBitmap:
4118 * Sample the texture and kill the fragment if the bit is 0.
4119 * This program will be combined with the user's fragment program.
4121 * Based on make_bitmap_fragment_program in st_cb_bitmap.c.
4124 get_bitmap_visitor(struct st_fragment_program
*fp
,
4125 glsl_to_tgsi_visitor
*original
, int samplerIndex
)
4127 glsl_to_tgsi_visitor
*v
= new glsl_to_tgsi_visitor();
4128 struct st_context
*st
= st_context(original
->ctx
);
4129 struct gl_program
*prog
= &fp
->Base
.Base
;
4130 st_src_reg coord
, src0
;
4132 glsl_to_tgsi_instruction
*inst
;
4134 /* Copy attributes of the glsl_to_tgsi_visitor in the original shader. */
4135 v
->ctx
= original
->ctx
;
4137 v
->shader_program
= NULL
;
4139 v
->glsl_version
= original
->glsl_version
;
4140 v
->native_integers
= original
->native_integers
;
4141 v
->options
= original
->options
;
4142 v
->next_temp
= original
->next_temp
;
4143 v
->num_address_regs
= original
->num_address_regs
;
4144 v
->samplers_used
= prog
->SamplersUsed
= original
->samplers_used
;
4145 v
->indirect_addr_consts
= original
->indirect_addr_consts
;
4146 memcpy(&v
->immediates
, &original
->immediates
, sizeof(v
->immediates
));
4147 v
->num_immediates
= original
->num_immediates
;
4149 /* TEX tmp0, fragment.texcoord[0], texture[0], 2D; */
4150 coord
= st_src_reg(PROGRAM_INPUT
, VARYING_SLOT_TEX0
, glsl_type::vec2_type
);
4151 src0
= v
->get_temp(glsl_type::vec4_type
);
4152 dst0
= st_dst_reg(src0
);
4153 inst
= v
->emit(NULL
, TGSI_OPCODE_TEX
, dst0
, coord
);
4154 inst
->sampler
= samplerIndex
;
4155 inst
->tex_target
= TEXTURE_2D_INDEX
;
4157 prog
->InputsRead
|= VARYING_BIT_TEX0
;
4158 prog
->SamplersUsed
|= (1 << samplerIndex
); /* mark sampler as used */
4159 v
->samplers_used
|= (1 << samplerIndex
);
4161 /* KIL if -tmp0 < 0 # texel=0 -> keep / texel=0 -> discard */
4162 src0
.negate
= NEGATE_XYZW
;
4163 if (st
->bitmap
.tex_format
== PIPE_FORMAT_L8_UNORM
)
4164 src0
.swizzle
= SWIZZLE_XXXX
;
4165 inst
= v
->emit(NULL
, TGSI_OPCODE_KILL_IF
, undef_dst
, src0
);
4167 /* Now copy the instructions from the original glsl_to_tgsi_visitor into the
4169 foreach_in_list(glsl_to_tgsi_instruction
, inst
, &original
->instructions
) {
4170 glsl_to_tgsi_instruction
*newinst
;
4171 st_src_reg src_regs
[3];
4173 if (inst
->dst
.file
== PROGRAM_OUTPUT
)
4174 prog
->OutputsWritten
|= BITFIELD64_BIT(inst
->dst
.index
);
4176 for (int i
=0; i
<3; i
++) {
4177 src_regs
[i
] = inst
->src
[i
];
4178 if (src_regs
[i
].file
== PROGRAM_INPUT
)
4179 prog
->InputsRead
|= BITFIELD64_BIT(src_regs
[i
].index
);
4182 newinst
= v
->emit(NULL
, inst
->op
, inst
->dst
, src_regs
[0], src_regs
[1], src_regs
[2]);
4183 newinst
->tex_target
= inst
->tex_target
;
4186 /* Make modifications to fragment program info. */
4187 prog
->Parameters
= _mesa_clone_parameter_list(original
->prog
->Parameters
);
4188 count_resources(v
, prog
);
4189 fp
->glsl_to_tgsi
= v
;
4192 /* ------------------------- TGSI conversion stuff -------------------------- */
4194 unsigned branch_target
;
4199 * Intermediate state used during shader translation.
4201 struct st_translate
{
4202 struct ureg_program
*ureg
;
4204 struct ureg_dst temps
[MAX_TEMPS
];
4205 struct ureg_dst arrays
[MAX_ARRAYS
];
4206 struct ureg_src
*constants
;
4207 struct ureg_src
*immediates
;
4208 struct ureg_dst outputs
[PIPE_MAX_SHADER_OUTPUTS
];
4209 struct ureg_src inputs
[PIPE_MAX_SHADER_INPUTS
];
4210 struct ureg_dst address
[2];
4211 struct ureg_src samplers
[PIPE_MAX_SAMPLERS
];
4212 struct ureg_src systemValues
[SYSTEM_VALUE_MAX
];
4213 struct tgsi_texture_offset tex_offsets
[MAX_GLSL_TEXTURE_OFFSET
];
4214 unsigned array_sizes
[MAX_ARRAYS
];
4216 const GLuint
*inputMapping
;
4217 const GLuint
*outputMapping
;
4219 /* For every instruction that contains a label (eg CALL), keep
4220 * details so that we can go back afterwards and emit the correct
4221 * tgsi instruction number for each label.
4223 struct label
*labels
;
4224 unsigned labels_size
;
4225 unsigned labels_count
;
4227 /* Keep a record of the tgsi instruction number that each mesa
4228 * instruction starts at, will be used to fix up labels after
4233 unsigned insn_count
;
4235 unsigned procType
; /**< TGSI_PROCESSOR_VERTEX/FRAGMENT */
4240 /** Map Mesa's SYSTEM_VALUE_x to TGSI_SEMANTIC_x */
4241 static unsigned mesa_sysval_to_semantic
[SYSTEM_VALUE_MAX
] = {
4243 TGSI_SEMANTIC_VERTEXID
,
4244 TGSI_SEMANTIC_INSTANCEID
,
4245 TGSI_SEMANTIC_SAMPLEID
,
4246 TGSI_SEMANTIC_SAMPLEPOS
,
4247 TGSI_SEMANTIC_SAMPLEMASK
,
4248 TGSI_SEMANTIC_INVOCATIONID
,
4252 * Make note of a branch to a label in the TGSI code.
4253 * After we've emitted all instructions, we'll go over the list
4254 * of labels built here and patch the TGSI code with the actual
4255 * location of each label.
4257 static unsigned *get_label(struct st_translate
*t
, unsigned branch_target
)
4261 if (t
->labels_count
+ 1 >= t
->labels_size
) {
4262 t
->labels_size
= 1 << (util_logbase2(t
->labels_size
) + 1);
4263 t
->labels
= (struct label
*)realloc(t
->labels
,
4264 t
->labels_size
* sizeof(struct label
));
4265 if (t
->labels
== NULL
) {
4266 static unsigned dummy
;
4272 i
= t
->labels_count
++;
4273 t
->labels
[i
].branch_target
= branch_target
;
4274 return &t
->labels
[i
].token
;
4278 * Called prior to emitting the TGSI code for each instruction.
4279 * Allocate additional space for instructions if needed.
4280 * Update the insn[] array so the next glsl_to_tgsi_instruction points to
4281 * the next TGSI instruction.
4283 static void set_insn_start(struct st_translate
*t
, unsigned start
)
4285 if (t
->insn_count
+ 1 >= t
->insn_size
) {
4286 t
->insn_size
= 1 << (util_logbase2(t
->insn_size
) + 1);
4287 t
->insn
= (unsigned *)realloc(t
->insn
, t
->insn_size
* sizeof(t
->insn
[0]));
4288 if (t
->insn
== NULL
) {
4294 t
->insn
[t
->insn_count
++] = start
;
4298 * Map a glsl_to_tgsi constant/immediate to a TGSI immediate.
4300 static struct ureg_src
4301 emit_immediate(struct st_translate
*t
,
4302 gl_constant_value values
[4],
4305 struct ureg_program
*ureg
= t
->ureg
;
4310 return ureg_DECL_immediate(ureg
, &values
[0].f
, size
);
4312 return ureg_DECL_immediate_int(ureg
, &values
[0].i
, size
);
4313 case GL_UNSIGNED_INT
:
4315 return ureg_DECL_immediate_uint(ureg
, &values
[0].u
, size
);
4317 assert(!"should not get here - type must be float, int, uint, or bool");
4318 return ureg_src_undef();
4323 * Map a glsl_to_tgsi dst register to a TGSI ureg_dst register.
4325 static struct ureg_dst
4326 dst_register(struct st_translate
*t
,
4327 gl_register_file file
,
4333 case PROGRAM_UNDEFINED
:
4334 return ureg_dst_undef();
4336 case PROGRAM_TEMPORARY
:
4338 assert(index
< (int) Elements(t
->temps
));
4340 if (ureg_dst_is_undef(t
->temps
[index
]))
4341 t
->temps
[index
] = ureg_DECL_local_temporary(t
->ureg
);
4343 return t
->temps
[index
];
4346 array
= index
>> 16;
4349 assert(array
< (int) Elements(t
->arrays
));
4351 if (ureg_dst_is_undef(t
->arrays
[array
]))
4352 t
->arrays
[array
] = ureg_DECL_array_temporary(
4353 t
->ureg
, t
->array_sizes
[array
], TRUE
);
4355 return ureg_dst_array_offset(t
->arrays
[array
],
4356 (int)(index
& 0xFFFF) - 0x8000);
4358 case PROGRAM_OUTPUT
:
4359 if (t
->procType
== TGSI_PROCESSOR_VERTEX
)
4360 assert(index
< VARYING_SLOT_MAX
);
4361 else if (t
->procType
== TGSI_PROCESSOR_FRAGMENT
)
4362 assert(index
< FRAG_RESULT_MAX
);
4364 assert(index
< VARYING_SLOT_MAX
);
4366 assert(t
->outputMapping
[index
] < Elements(t
->outputs
));
4368 return t
->outputs
[t
->outputMapping
[index
]];
4370 case PROGRAM_ADDRESS
:
4371 return t
->address
[index
];
4374 assert(!"unknown dst register file");
4375 return ureg_dst_undef();
4380 * Map a glsl_to_tgsi src register to a TGSI ureg_src register.
4382 static struct ureg_src
4383 src_register(struct st_translate
*t
, const struct st_src_reg
*reg
)
4386 case PROGRAM_UNDEFINED
:
4387 return ureg_src_undef();
4389 case PROGRAM_TEMPORARY
:
4391 return ureg_src(dst_register(t
, reg
->file
, reg
->index
));
4393 case PROGRAM_UNIFORM
:
4394 assert(reg
->index
>= 0);
4395 return t
->constants
[reg
->index
];
4396 case PROGRAM_STATE_VAR
:
4397 case PROGRAM_CONSTANT
: /* ie, immediate */
4398 if (reg
->has_index2
)
4399 return ureg_src_register(TGSI_FILE_CONSTANT
, reg
->index
);
4400 else if (reg
->index
< 0)
4401 return ureg_DECL_constant(t
->ureg
, 0);
4403 return t
->constants
[reg
->index
];
4405 case PROGRAM_IMMEDIATE
:
4406 return t
->immediates
[reg
->index
];
4409 assert(t
->inputMapping
[reg
->index
] < Elements(t
->inputs
));
4410 return t
->inputs
[t
->inputMapping
[reg
->index
]];
4412 case PROGRAM_OUTPUT
:
4413 assert(t
->outputMapping
[reg
->index
] < Elements(t
->outputs
));
4414 return ureg_src(t
->outputs
[t
->outputMapping
[reg
->index
]]); /* not needed? */
4416 case PROGRAM_ADDRESS
:
4417 return ureg_src(t
->address
[reg
->index
]);
4419 case PROGRAM_SYSTEM_VALUE
:
4420 assert(reg
->index
< (int) Elements(t
->systemValues
));
4421 return t
->systemValues
[reg
->index
];
4424 assert(!"unknown src register file");
4425 return ureg_src_undef();
4430 * Create a TGSI ureg_dst register from an st_dst_reg.
4432 static struct ureg_dst
4433 translate_dst(struct st_translate
*t
,
4434 const st_dst_reg
*dst_reg
,
4435 bool saturate
, bool clamp_color
)
4437 struct ureg_dst dst
= dst_register(t
,
4441 dst
= ureg_writemask(dst
, dst_reg
->writemask
);
4444 dst
= ureg_saturate(dst
);
4445 else if (clamp_color
&& dst_reg
->file
== PROGRAM_OUTPUT
) {
4446 /* Clamp colors for ARB_color_buffer_float. */
4447 switch (t
->procType
) {
4448 case TGSI_PROCESSOR_VERTEX
:
4449 /* XXX if the geometry shader is present, this must be done there
4450 * instead of here. */
4451 if (dst_reg
->index
== VARYING_SLOT_COL0
||
4452 dst_reg
->index
== VARYING_SLOT_COL1
||
4453 dst_reg
->index
== VARYING_SLOT_BFC0
||
4454 dst_reg
->index
== VARYING_SLOT_BFC1
) {
4455 dst
= ureg_saturate(dst
);
4459 case TGSI_PROCESSOR_FRAGMENT
:
4460 if (dst_reg
->index
== FRAG_RESULT_COLOR
||
4461 dst_reg
->index
>= FRAG_RESULT_DATA0
) {
4462 dst
= ureg_saturate(dst
);
4468 if (dst_reg
->reladdr
!= NULL
) {
4469 assert(dst_reg
->file
!= PROGRAM_TEMPORARY
);
4470 dst
= ureg_dst_indirect(dst
, ureg_src(t
->address
[0]));
4477 * Create a TGSI ureg_src register from an st_src_reg.
4479 static struct ureg_src
4480 translate_src(struct st_translate
*t
, const st_src_reg
*src_reg
)
4482 struct ureg_src src
= src_register(t
, src_reg
);
4484 if (src_reg
->has_index2
) {
4485 /* 2D indexes occur with geometry shader inputs (attrib, vertex)
4486 * and UBO constant buffers (buffer, position).
4488 if (src_reg
->reladdr2
)
4489 src
= ureg_src_dimension_indirect(src
, ureg_src(t
->address
[1]),
4492 src
= ureg_src_dimension(src
, src_reg
->index2D
);
4495 src
= ureg_swizzle(src
,
4496 GET_SWZ(src_reg
->swizzle
, 0) & 0x3,
4497 GET_SWZ(src_reg
->swizzle
, 1) & 0x3,
4498 GET_SWZ(src_reg
->swizzle
, 2) & 0x3,
4499 GET_SWZ(src_reg
->swizzle
, 3) & 0x3);
4501 if ((src_reg
->negate
& 0xf) == NEGATE_XYZW
)
4502 src
= ureg_negate(src
);
4504 if (src_reg
->reladdr
!= NULL
) {
4505 assert(src_reg
->file
!= PROGRAM_TEMPORARY
);
4506 src
= ureg_src_indirect(src
, ureg_src(t
->address
[0]));
4512 static struct tgsi_texture_offset
4513 translate_tex_offset(struct st_translate
*t
,
4514 const st_src_reg
*in_offset
, int idx
)
4516 struct tgsi_texture_offset offset
;
4517 struct ureg_src imm_src
;
4518 struct ureg_dst dst
;
4521 switch (in_offset
->file
) {
4522 case PROGRAM_IMMEDIATE
:
4523 imm_src
= t
->immediates
[in_offset
->index
];
4525 offset
.File
= imm_src
.File
;
4526 offset
.Index
= imm_src
.Index
;
4527 offset
.SwizzleX
= imm_src
.SwizzleX
;
4528 offset
.SwizzleY
= imm_src
.SwizzleY
;
4529 offset
.SwizzleZ
= imm_src
.SwizzleZ
;
4532 case PROGRAM_TEMPORARY
:
4533 imm_src
= ureg_src(t
->temps
[in_offset
->index
]);
4534 offset
.File
= imm_src
.File
;
4535 offset
.Index
= imm_src
.Index
;
4536 offset
.SwizzleX
= GET_SWZ(in_offset
->swizzle
, 0);
4537 offset
.SwizzleY
= GET_SWZ(in_offset
->swizzle
, 1);
4538 offset
.SwizzleZ
= GET_SWZ(in_offset
->swizzle
, 2);
4542 array
= in_offset
->index
>> 16;
4545 assert(array
< (int) Elements(t
->arrays
));
4547 dst
= t
->arrays
[array
];
4548 offset
.File
= dst
.File
;
4549 offset
.Index
= dst
.Index
+ (in_offset
->index
& 0xFFFF) - 0x8000;
4550 offset
.SwizzleX
= GET_SWZ(in_offset
->swizzle
, 0);
4551 offset
.SwizzleY
= GET_SWZ(in_offset
->swizzle
, 1);
4552 offset
.SwizzleZ
= GET_SWZ(in_offset
->swizzle
, 2);
4562 compile_tgsi_instruction(struct st_translate
*t
,
4563 const glsl_to_tgsi_instruction
*inst
,
4564 bool clamp_dst_color_output
)
4566 struct ureg_program
*ureg
= t
->ureg
;
4568 struct ureg_dst dst
[1];
4569 struct ureg_src src
[4];
4570 struct tgsi_texture_offset texoffsets
[MAX_GLSL_TEXTURE_OFFSET
];
4574 unsigned tex_target
;
4576 num_dst
= num_inst_dst_regs(inst
->op
);
4577 num_src
= num_inst_src_regs(inst
->op
);
4580 dst
[0] = translate_dst(t
,
4583 clamp_dst_color_output
);
4585 for (i
= 0; i
< num_src
; i
++)
4586 src
[i
] = translate_src(t
, &inst
->src
[i
]);
4589 case TGSI_OPCODE_BGNLOOP
:
4590 case TGSI_OPCODE_CAL
:
4591 case TGSI_OPCODE_ELSE
:
4592 case TGSI_OPCODE_ENDLOOP
:
4593 case TGSI_OPCODE_IF
:
4594 case TGSI_OPCODE_UIF
:
4595 assert(num_dst
== 0);
4596 ureg_label_insn(ureg
,
4600 inst
->op
== TGSI_OPCODE_CAL
? inst
->function
->sig_id
: 0));
4603 case TGSI_OPCODE_TEX
:
4604 case TGSI_OPCODE_TXB
:
4605 case TGSI_OPCODE_TXD
:
4606 case TGSI_OPCODE_TXL
:
4607 case TGSI_OPCODE_TXP
:
4608 case TGSI_OPCODE_TXQ
:
4609 case TGSI_OPCODE_TXF
:
4610 case TGSI_OPCODE_TEX2
:
4611 case TGSI_OPCODE_TXB2
:
4612 case TGSI_OPCODE_TXL2
:
4613 case TGSI_OPCODE_TG4
:
4614 case TGSI_OPCODE_LODQ
:
4615 src
[num_src
++] = t
->samplers
[inst
->sampler
];
4616 for (i
= 0; i
< inst
->tex_offset_num_offset
; i
++) {
4617 texoffsets
[i
] = translate_tex_offset(t
, &inst
->tex_offsets
[i
], i
);
4619 tex_target
= st_translate_texture_target(inst
->tex_target
, inst
->tex_shadow
);
4625 texoffsets
, inst
->tex_offset_num_offset
,
4629 case TGSI_OPCODE_SCS
:
4630 dst
[0] = ureg_writemask(dst
[0], TGSI_WRITEMASK_XY
);
4631 ureg_insn(ureg
, inst
->op
, dst
, num_dst
, src
, num_src
);
4644 * Emit the TGSI instructions for inverting and adjusting WPOS.
4645 * This code is unavoidable because it also depends on whether
4646 * a FBO is bound (STATE_FB_WPOS_Y_TRANSFORM).
4649 emit_wpos_adjustment( struct st_translate
*t
,
4650 const struct gl_program
*program
,
4652 GLfloat adjX
, GLfloat adjY
[2])
4654 struct ureg_program
*ureg
= t
->ureg
;
4656 /* Fragment program uses fragment position input.
4657 * Need to replace instances of INPUT[WPOS] with temp T
4658 * where T = INPUT[WPOS] by y is inverted.
4660 static const gl_state_index wposTransformState
[STATE_LENGTH
]
4661 = { STATE_INTERNAL
, STATE_FB_WPOS_Y_TRANSFORM
,
4662 (gl_state_index
)0, (gl_state_index
)0, (gl_state_index
)0 };
4664 /* XXX: note we are modifying the incoming shader here! Need to
4665 * do this before emitting the constant decls below, or this
4668 unsigned wposTransConst
= _mesa_add_state_reference(program
->Parameters
,
4669 wposTransformState
);
4671 struct ureg_src wpostrans
= ureg_DECL_constant( ureg
, wposTransConst
);
4672 struct ureg_dst wpos_temp
= ureg_DECL_temporary( ureg
);
4673 struct ureg_src wpos_input
= t
->inputs
[t
->inputMapping
[VARYING_SLOT_POS
]];
4675 /* First, apply the coordinate shift: */
4676 if (adjX
|| adjY
[0] || adjY
[1]) {
4677 if (adjY
[0] != adjY
[1]) {
4678 /* Adjust the y coordinate by adjY[1] or adjY[0] respectively
4679 * depending on whether inversion is actually going to be applied
4680 * or not, which is determined by testing against the inversion
4681 * state variable used below, which will be either +1 or -1.
4683 struct ureg_dst adj_temp
= ureg_DECL_local_temporary(ureg
);
4685 ureg_CMP(ureg
, adj_temp
,
4686 ureg_scalar(wpostrans
, invert
? 2 : 0),
4687 ureg_imm4f(ureg
, adjX
, adjY
[0], 0.0f
, 0.0f
),
4688 ureg_imm4f(ureg
, adjX
, adjY
[1], 0.0f
, 0.0f
));
4689 ureg_ADD(ureg
, wpos_temp
, wpos_input
, ureg_src(adj_temp
));
4691 ureg_ADD(ureg
, wpos_temp
, wpos_input
,
4692 ureg_imm4f(ureg
, adjX
, adjY
[0], 0.0f
, 0.0f
));
4694 wpos_input
= ureg_src(wpos_temp
);
4696 /* MOV wpos_temp, input[wpos]
4698 ureg_MOV( ureg
, wpos_temp
, wpos_input
);
4701 /* Now the conditional y flip: STATE_FB_WPOS_Y_TRANSFORM.xy/zw will be
4702 * inversion/identity, or the other way around if we're drawing to an FBO.
4705 /* MAD wpos_temp.y, wpos_input, wpostrans.xxxx, wpostrans.yyyy
4708 ureg_writemask(wpos_temp
, TGSI_WRITEMASK_Y
),
4710 ureg_scalar(wpostrans
, 0),
4711 ureg_scalar(wpostrans
, 1));
4713 /* MAD wpos_temp.y, wpos_input, wpostrans.zzzz, wpostrans.wwww
4716 ureg_writemask(wpos_temp
, TGSI_WRITEMASK_Y
),
4718 ureg_scalar(wpostrans
, 2),
4719 ureg_scalar(wpostrans
, 3));
4722 /* Use wpos_temp as position input from here on:
4724 t
->inputs
[t
->inputMapping
[VARYING_SLOT_POS
]] = ureg_src(wpos_temp
);
4729 * Emit fragment position/ooordinate code.
4732 emit_wpos(struct st_context
*st
,
4733 struct st_translate
*t
,
4734 const struct gl_program
*program
,
4735 struct ureg_program
*ureg
)
4737 const struct gl_fragment_program
*fp
=
4738 (const struct gl_fragment_program
*) program
;
4739 struct pipe_screen
*pscreen
= st
->pipe
->screen
;
4740 GLfloat adjX
= 0.0f
;
4741 GLfloat adjY
[2] = { 0.0f
, 0.0f
};
4742 boolean invert
= FALSE
;
4744 /* Query the pixel center conventions supported by the pipe driver and set
4745 * adjX, adjY to help out if it cannot handle the requested one internally.
4747 * The bias of the y-coordinate depends on whether y-inversion takes place
4748 * (adjY[1]) or not (adjY[0]), which is in turn dependent on whether we are
4749 * drawing to an FBO (causes additional inversion), and whether the the pipe
4750 * driver origin and the requested origin differ (the latter condition is
4751 * stored in the 'invert' variable).
4753 * For height = 100 (i = integer, h = half-integer, l = lower, u = upper):
4755 * center shift only:
4760 * l,i -> u,i: ( 0.0 + 1.0) * -1 + 100 = 99
4761 * l,h -> u,h: ( 0.5 + 0.0) * -1 + 100 = 99.5
4762 * u,i -> l,i: (99.0 + 1.0) * -1 + 100 = 0
4763 * u,h -> l,h: (99.5 + 0.0) * -1 + 100 = 0.5
4765 * inversion and center shift:
4766 * l,i -> u,h: ( 0.0 + 0.5) * -1 + 100 = 99.5
4767 * l,h -> u,i: ( 0.5 + 0.5) * -1 + 100 = 99
4768 * u,i -> l,h: (99.0 + 0.5) * -1 + 100 = 0.5
4769 * u,h -> l,i: (99.5 + 0.5) * -1 + 100 = 0
4771 if (fp
->OriginUpperLeft
) {
4772 /* Fragment shader wants origin in upper-left */
4773 if (pscreen
->get_param(pscreen
, PIPE_CAP_TGSI_FS_COORD_ORIGIN_UPPER_LEFT
)) {
4774 /* the driver supports upper-left origin */
4776 else if (pscreen
->get_param(pscreen
, PIPE_CAP_TGSI_FS_COORD_ORIGIN_LOWER_LEFT
)) {
4777 /* the driver supports lower-left origin, need to invert Y */
4778 ureg_property_fs_coord_origin(ureg
, TGSI_FS_COORD_ORIGIN_LOWER_LEFT
);
4785 /* Fragment shader wants origin in lower-left */
4786 if (pscreen
->get_param(pscreen
, PIPE_CAP_TGSI_FS_COORD_ORIGIN_LOWER_LEFT
))
4787 /* the driver supports lower-left origin */
4788 ureg_property_fs_coord_origin(ureg
, TGSI_FS_COORD_ORIGIN_LOWER_LEFT
);
4789 else if (pscreen
->get_param(pscreen
, PIPE_CAP_TGSI_FS_COORD_ORIGIN_UPPER_LEFT
))
4790 /* the driver supports upper-left origin, need to invert Y */
4796 if (fp
->PixelCenterInteger
) {
4797 /* Fragment shader wants pixel center integer */
4798 if (pscreen
->get_param(pscreen
, PIPE_CAP_TGSI_FS_COORD_PIXEL_CENTER_INTEGER
)) {
4799 /* the driver supports pixel center integer */
4801 ureg_property_fs_coord_pixel_center(ureg
, TGSI_FS_COORD_PIXEL_CENTER_INTEGER
);
4803 else if (pscreen
->get_param(pscreen
, PIPE_CAP_TGSI_FS_COORD_PIXEL_CENTER_HALF_INTEGER
)) {
4804 /* the driver supports pixel center half integer, need to bias X,Y */
4813 /* Fragment shader wants pixel center half integer */
4814 if (pscreen
->get_param(pscreen
, PIPE_CAP_TGSI_FS_COORD_PIXEL_CENTER_HALF_INTEGER
)) {
4815 /* the driver supports pixel center half integer */
4817 else if (pscreen
->get_param(pscreen
, PIPE_CAP_TGSI_FS_COORD_PIXEL_CENTER_INTEGER
)) {
4818 /* the driver supports pixel center integer, need to bias X,Y */
4819 adjX
= adjY
[0] = adjY
[1] = 0.5f
;
4820 ureg_property_fs_coord_pixel_center(ureg
, TGSI_FS_COORD_PIXEL_CENTER_INTEGER
);
4826 /* we invert after adjustment so that we avoid the MOV to temporary,
4827 * and reuse the adjustment ADD instead */
4828 emit_wpos_adjustment(t
, program
, invert
, adjX
, adjY
);
4832 * OpenGL's fragment gl_FrontFace input is 1 for front-facing, 0 for back.
4833 * TGSI uses +1 for front, -1 for back.
4834 * This function converts the TGSI value to the GL value. Simply clamping/
4835 * saturating the value to [0,1] does the job.
4838 emit_face_var(struct st_translate
*t
)
4840 struct ureg_program
*ureg
= t
->ureg
;
4841 struct ureg_dst face_temp
= ureg_DECL_temporary(ureg
);
4842 struct ureg_src face_input
= t
->inputs
[t
->inputMapping
[VARYING_SLOT_FACE
]];
4844 /* MOV_SAT face_temp, input[face] */
4845 face_temp
= ureg_saturate(face_temp
);
4846 ureg_MOV(ureg
, face_temp
, face_input
);
4848 /* Use face_temp as face input from here on: */
4849 t
->inputs
[t
->inputMapping
[VARYING_SLOT_FACE
]] = ureg_src(face_temp
);
4853 emit_edgeflags(struct st_translate
*t
)
4855 struct ureg_program
*ureg
= t
->ureg
;
4856 struct ureg_dst edge_dst
= t
->outputs
[t
->outputMapping
[VARYING_SLOT_EDGE
]];
4857 struct ureg_src edge_src
= t
->inputs
[t
->inputMapping
[VERT_ATTRIB_EDGEFLAG
]];
4859 ureg_MOV(ureg
, edge_dst
, edge_src
);
4863 * Translate intermediate IR (glsl_to_tgsi_instruction) to TGSI format.
4864 * \param program the program to translate
4865 * \param numInputs number of input registers used
4866 * \param inputMapping maps Mesa fragment program inputs to TGSI generic
4868 * \param inputSemanticName the TGSI_SEMANTIC flag for each input
4869 * \param inputSemanticIndex the semantic index (ex: which texcoord) for
4871 * \param interpMode the TGSI_INTERPOLATE_LINEAR/PERSP mode for each input
4872 * \param interpLocation the TGSI_INTERPOLATE_LOC_* location for each input
4873 * \param numOutputs number of output registers used
4874 * \param outputMapping maps Mesa fragment program outputs to TGSI
4876 * \param outputSemanticName the TGSI_SEMANTIC flag for each output
4877 * \param outputSemanticIndex the semantic index (ex: which texcoord) for
4880 * \return PIPE_OK or PIPE_ERROR_OUT_OF_MEMORY
4882 extern "C" enum pipe_error
4883 st_translate_program(
4884 struct gl_context
*ctx
,
4886 struct ureg_program
*ureg
,
4887 glsl_to_tgsi_visitor
*program
,
4888 const struct gl_program
*proginfo
,
4890 const GLuint inputMapping
[],
4891 const ubyte inputSemanticName
[],
4892 const ubyte inputSemanticIndex
[],
4893 const GLuint interpMode
[],
4894 const GLuint interpLocation
[],
4896 const GLuint outputMapping
[],
4897 const ubyte outputSemanticName
[],
4898 const ubyte outputSemanticIndex
[],
4899 boolean passthrough_edgeflags
,
4900 boolean clamp_color
)
4902 struct st_translate
*t
;
4904 enum pipe_error ret
= PIPE_OK
;
4906 assert(numInputs
<= Elements(t
->inputs
));
4907 assert(numOutputs
<= Elements(t
->outputs
));
4909 t
= CALLOC_STRUCT(st_translate
);
4911 ret
= PIPE_ERROR_OUT_OF_MEMORY
;
4915 memset(t
, 0, sizeof *t
);
4917 t
->procType
= procType
;
4918 t
->inputMapping
= inputMapping
;
4919 t
->outputMapping
= outputMapping
;
4922 if (program
->shader_program
) {
4923 for (i
= 0; i
< program
->shader_program
->NumUserUniformStorage
; i
++) {
4924 struct gl_uniform_storage
*const storage
=
4925 &program
->shader_program
->UniformStorage
[i
];
4927 _mesa_uniform_detach_all_driver_storage(storage
);
4932 * Declare input attributes.
4934 if (procType
== TGSI_PROCESSOR_FRAGMENT
) {
4935 for (i
= 0; i
< numInputs
; i
++) {
4936 t
->inputs
[i
] = ureg_DECL_fs_input_cyl_centroid(ureg
,
4937 inputSemanticName
[i
],
4938 inputSemanticIndex
[i
],
4943 if (proginfo
->InputsRead
& VARYING_BIT_POS
) {
4944 /* Must do this after setting up t->inputs, and before
4945 * emitting constant references, below:
4947 emit_wpos(st_context(ctx
), t
, proginfo
, ureg
);
4950 if (proginfo
->InputsRead
& VARYING_BIT_FACE
)
4954 * Declare output attributes.
4956 for (i
= 0; i
< numOutputs
; i
++) {
4957 switch (outputSemanticName
[i
]) {
4958 case TGSI_SEMANTIC_POSITION
:
4959 t
->outputs
[i
] = ureg_DECL_output(ureg
,
4960 TGSI_SEMANTIC_POSITION
, /* Z/Depth */
4961 outputSemanticIndex
[i
]);
4962 t
->outputs
[i
] = ureg_writemask(t
->outputs
[i
], TGSI_WRITEMASK_Z
);
4964 case TGSI_SEMANTIC_STENCIL
:
4965 t
->outputs
[i
] = ureg_DECL_output(ureg
,
4966 TGSI_SEMANTIC_STENCIL
, /* Stencil */
4967 outputSemanticIndex
[i
]);
4968 t
->outputs
[i
] = ureg_writemask(t
->outputs
[i
], TGSI_WRITEMASK_Y
);
4970 case TGSI_SEMANTIC_COLOR
:
4971 t
->outputs
[i
] = ureg_DECL_output(ureg
,
4972 TGSI_SEMANTIC_COLOR
,
4973 outputSemanticIndex
[i
]);
4975 case TGSI_SEMANTIC_SAMPLEMASK
:
4976 t
->outputs
[i
] = ureg_DECL_output(ureg
,
4977 TGSI_SEMANTIC_SAMPLEMASK
,
4978 outputSemanticIndex
[i
]);
4979 /* TODO: If we ever support more than 32 samples, this will have
4980 * to become an array.
4982 t
->outputs
[i
] = ureg_writemask(t
->outputs
[i
], TGSI_WRITEMASK_X
);
4985 assert(!"fragment shader outputs must be POSITION/STENCIL/COLOR");
4986 ret
= PIPE_ERROR_BAD_INPUT
;
4991 else if (procType
== TGSI_PROCESSOR_GEOMETRY
) {
4992 for (i
= 0; i
< numInputs
; i
++) {
4993 t
->inputs
[i
] = ureg_DECL_gs_input(ureg
,
4995 inputSemanticName
[i
],
4996 inputSemanticIndex
[i
]);
4999 for (i
= 0; i
< numOutputs
; i
++) {
5000 t
->outputs
[i
] = ureg_DECL_output(ureg
,
5001 outputSemanticName
[i
],
5002 outputSemanticIndex
[i
]);
5006 assert(procType
== TGSI_PROCESSOR_VERTEX
);
5008 for (i
= 0; i
< numInputs
; i
++) {
5009 t
->inputs
[i
] = ureg_DECL_vs_input(ureg
, i
);
5012 for (i
= 0; i
< numOutputs
; i
++) {
5013 t
->outputs
[i
] = ureg_DECL_output(ureg
,
5014 outputSemanticName
[i
],
5015 outputSemanticIndex
[i
]);
5016 if (outputSemanticName
[i
] == TGSI_SEMANTIC_FOG
) {
5017 /* force register to contain a fog coordinate in the form (F, 0, 0, 1). */
5019 ureg_writemask(t
->outputs
[i
], TGSI_WRITEMASK_YZW
),
5020 ureg_imm4f(ureg
, 0.0f
, 0.0f
, 0.0f
, 1.0f
));
5021 t
->outputs
[i
] = ureg_writemask(t
->outputs
[i
], TGSI_WRITEMASK_X
);
5024 if (passthrough_edgeflags
)
5028 /* Declare address register.
5030 if (program
->num_address_regs
> 0) {
5031 assert(program
->num_address_regs
<= 2);
5032 t
->address
[0] = ureg_DECL_address(ureg
);
5033 if (program
->num_address_regs
== 2)
5034 t
->address
[1] = ureg_DECL_address(ureg
);
5037 /* Declare misc input registers
5040 GLbitfield sysInputs
= proginfo
->SystemValuesRead
;
5041 unsigned numSys
= 0;
5042 for (i
= 0; sysInputs
; i
++) {
5043 if (sysInputs
& (1 << i
)) {
5044 unsigned semName
= mesa_sysval_to_semantic
[i
];
5045 t
->systemValues
[i
] = ureg_DECL_system_value(ureg
, numSys
, semName
, 0);
5046 if (semName
== TGSI_SEMANTIC_INSTANCEID
||
5047 semName
== TGSI_SEMANTIC_VERTEXID
) {
5048 /* From Gallium perspective, these system values are always
5049 * integer, and require native integer support. However, if
5050 * native integer is supported on the vertex stage but not the
5051 * pixel stage (e.g, i915g + draw), Mesa will generate IR that
5052 * assumes these system values are floats. To resolve the
5053 * inconsistency, we insert a U2F.
5055 struct st_context
*st
= st_context(ctx
);
5056 struct pipe_screen
*pscreen
= st
->pipe
->screen
;
5057 assert(procType
== TGSI_PROCESSOR_VERTEX
);
5058 assert(pscreen
->get_shader_param(pscreen
, PIPE_SHADER_VERTEX
, PIPE_SHADER_CAP_INTEGERS
));
5059 if (!ctx
->Const
.NativeIntegers
) {
5060 struct ureg_dst temp
= ureg_DECL_local_temporary(t
->ureg
);
5061 ureg_U2F( t
->ureg
, ureg_writemask(temp
, TGSI_WRITEMASK_X
), t
->systemValues
[i
]);
5062 t
->systemValues
[i
] = ureg_scalar(ureg_src(temp
), 0);
5066 sysInputs
&= ~(1 << i
);
5071 /* Copy over array sizes
5073 memcpy(t
->array_sizes
, program
->array_sizes
, sizeof(unsigned) * program
->next_array
);
5075 /* Emit constants and uniforms. TGSI uses a single index space for these,
5076 * so we put all the translated regs in t->constants.
5078 if (proginfo
->Parameters
) {
5079 t
->constants
= (struct ureg_src
*)
5080 calloc(proginfo
->Parameters
->NumParameters
, sizeof(t
->constants
[0]));
5081 if (t
->constants
== NULL
) {
5082 ret
= PIPE_ERROR_OUT_OF_MEMORY
;
5086 for (i
= 0; i
< proginfo
->Parameters
->NumParameters
; i
++) {
5087 switch (proginfo
->Parameters
->Parameters
[i
].Type
) {
5088 case PROGRAM_STATE_VAR
:
5089 case PROGRAM_UNIFORM
:
5090 t
->constants
[i
] = ureg_DECL_constant(ureg
, i
);
5093 /* Emit immediates for PROGRAM_CONSTANT only when there's no indirect
5094 * addressing of the const buffer.
5095 * FIXME: Be smarter and recognize param arrays:
5096 * indirect addressing is only valid within the referenced
5099 case PROGRAM_CONSTANT
:
5100 if (program
->indirect_addr_consts
)
5101 t
->constants
[i
] = ureg_DECL_constant(ureg
, i
);
5103 t
->constants
[i
] = emit_immediate(t
,
5104 proginfo
->Parameters
->ParameterValues
[i
],
5105 proginfo
->Parameters
->Parameters
[i
].DataType
,
5114 if (program
->shader
) {
5115 unsigned num_ubos
= program
->shader
->NumUniformBlocks
;
5117 for (i
= 0; i
< num_ubos
; i
++) {
5119 program
->shader_program
->UniformBlocks
[i
].UniformBufferSize
;
5120 unsigned num_const_vecs
= (size
+ 15) / 16;
5121 unsigned first
, last
;
5122 assert(num_const_vecs
> 0);
5124 last
= num_const_vecs
> 0 ? num_const_vecs
- 1 : 0;
5125 ureg_DECL_constant2D(t
->ureg
, first
, last
, i
+ 1);
5129 /* Emit immediate values.
5131 t
->immediates
= (struct ureg_src
*)
5132 calloc(program
->num_immediates
, sizeof(struct ureg_src
));
5133 if (t
->immediates
== NULL
) {
5134 ret
= PIPE_ERROR_OUT_OF_MEMORY
;
5138 foreach_in_list(immediate_storage
, imm
, &program
->immediates
) {
5139 assert(i
< program
->num_immediates
);
5140 t
->immediates
[i
++] = emit_immediate(t
, imm
->values
, imm
->type
, imm
->size
);
5142 assert(i
== program
->num_immediates
);
5144 /* texture samplers */
5145 for (i
= 0; i
< ctx
->Const
.Program
[MESA_SHADER_FRAGMENT
].MaxTextureImageUnits
; i
++) {
5146 if (program
->samplers_used
& (1 << i
)) {
5147 t
->samplers
[i
] = ureg_DECL_sampler(ureg
, i
);
5151 /* Emit each instruction in turn:
5153 foreach_in_list(glsl_to_tgsi_instruction
, inst
, &program
->instructions
) {
5154 set_insn_start(t
, ureg_get_instruction_number(ureg
));
5155 compile_tgsi_instruction(t
, inst
, clamp_color
);
5158 /* Fix up all emitted labels:
5160 for (i
= 0; i
< t
->labels_count
; i
++) {
5161 ureg_fixup_label(ureg
, t
->labels
[i
].token
,
5162 t
->insn
[t
->labels
[i
].branch_target
]);
5165 if (program
->shader_program
) {
5166 /* This has to be done last. Any operation the can cause
5167 * prog->ParameterValues to get reallocated (e.g., anything that adds a
5168 * program constant) has to happen before creating this linkage.
5170 for (unsigned i
= 0; i
< MESA_SHADER_STAGES
; i
++) {
5171 if (program
->shader_program
->_LinkedShaders
[i
] == NULL
)
5174 _mesa_associate_uniform_storage(ctx
, program
->shader_program
,
5175 program
->shader_program
->_LinkedShaders
[i
]->Program
->Parameters
);
5184 free(t
->immediates
);
5187 debug_printf("%s: translate error flag set\n", __FUNCTION__
);
5195 /* ----------------------------- End TGSI code ------------------------------ */
5199 shader_stage_to_ptarget(gl_shader_stage stage
)
5202 case MESA_SHADER_VERTEX
:
5203 return PIPE_SHADER_VERTEX
;
5204 case MESA_SHADER_FRAGMENT
:
5205 return PIPE_SHADER_FRAGMENT
;
5206 case MESA_SHADER_GEOMETRY
:
5207 return PIPE_SHADER_GEOMETRY
;
5208 case MESA_SHADER_COMPUTE
:
5209 return PIPE_SHADER_COMPUTE
;
5212 assert(!"should not be reached");
5213 return PIPE_SHADER_VERTEX
;
5218 * Convert a shader's GLSL IR into a Mesa gl_program, although without
5219 * generating Mesa IR.
5221 static struct gl_program
*
5222 get_mesa_program(struct gl_context
*ctx
,
5223 struct gl_shader_program
*shader_program
,
5224 struct gl_shader
*shader
)
5226 glsl_to_tgsi_visitor
* v
;
5227 struct gl_program
*prog
;
5228 GLenum target
= _mesa_shader_stage_to_program(shader
->Stage
);
5230 struct gl_shader_compiler_options
*options
=
5231 &ctx
->ShaderCompilerOptions
[_mesa_shader_enum_to_shader_stage(shader
->Type
)];
5232 struct pipe_screen
*pscreen
= ctx
->st
->pipe
->screen
;
5233 unsigned ptarget
= shader_stage_to_ptarget(shader
->Stage
);
5235 validate_ir_tree(shader
->ir
);
5237 prog
= ctx
->Driver
.NewProgram(ctx
, target
, shader_program
->Name
);
5240 prog
->Parameters
= _mesa_new_parameter_list();
5241 v
= new glsl_to_tgsi_visitor();
5244 v
->shader_program
= shader_program
;
5246 v
->options
= options
;
5247 v
->glsl_version
= ctx
->Const
.GLSLVersion
;
5248 v
->native_integers
= ctx
->Const
.NativeIntegers
;
5250 v
->have_sqrt
= pscreen
->get_shader_param(pscreen
, ptarget
,
5251 PIPE_SHADER_CAP_TGSI_SQRT_SUPPORTED
);
5253 _mesa_generate_parameters_list_for_uniforms(shader_program
, shader
,
5256 /* Remove reads from output registers. */
5257 lower_output_reads(shader
->ir
);
5259 /* Emit intermediate IR for main(). */
5260 visit_exec_list(shader
->ir
, v
);
5262 /* Now emit bodies for any functions that were used. */
5264 progress
= GL_FALSE
;
5266 foreach_in_list(function_entry
, entry
, &v
->function_signatures
) {
5267 if (!entry
->bgn_inst
) {
5268 v
->current_function
= entry
;
5270 entry
->bgn_inst
= v
->emit(NULL
, TGSI_OPCODE_BGNSUB
);
5271 entry
->bgn_inst
->function
= entry
;
5273 visit_exec_list(&entry
->sig
->body
, v
);
5275 glsl_to_tgsi_instruction
*last
;
5276 last
= (glsl_to_tgsi_instruction
*)v
->instructions
.get_tail();
5277 if (last
->op
!= TGSI_OPCODE_RET
)
5278 v
->emit(NULL
, TGSI_OPCODE_RET
);
5280 glsl_to_tgsi_instruction
*end
;
5281 end
= v
->emit(NULL
, TGSI_OPCODE_ENDSUB
);
5282 end
->function
= entry
;
5290 /* Print out some information (for debugging purposes) used by the
5291 * optimization passes. */
5292 for (i
=0; i
< v
->next_temp
; i
++) {
5293 int fr
= v
->get_first_temp_read(i
);
5294 int fw
= v
->get_first_temp_write(i
);
5295 int lr
= v
->get_last_temp_read(i
);
5296 int lw
= v
->get_last_temp_write(i
);
5298 printf("Temp %d: FR=%3d FW=%3d LR=%3d LW=%3d\n", i
, fr
, fw
, lr
, lw
);
5303 /* Perform optimizations on the instructions in the glsl_to_tgsi_visitor. */
5305 v
->copy_propagate();
5306 while (v
->eliminate_dead_code());
5308 v
->merge_registers();
5309 v
->renumber_registers();
5311 /* Write the END instruction. */
5312 v
->emit(NULL
, TGSI_OPCODE_END
);
5314 if (ctx
->_Shader
->Flags
& GLSL_DUMP
) {
5316 printf("GLSL IR for linked %s program %d:\n",
5317 _mesa_shader_stage_to_string(shader
->Stage
),
5318 shader_program
->Name
);
5319 _mesa_print_ir(stdout
, shader
->ir
, NULL
);
5325 prog
->Instructions
= NULL
;
5326 prog
->NumInstructions
= 0;
5328 do_set_program_inouts(shader
->ir
, prog
, shader
->Stage
);
5329 count_resources(v
, prog
);
5331 _mesa_reference_program(ctx
, &shader
->Program
, prog
);
5333 /* This has to be done last. Any operation the can cause
5334 * prog->ParameterValues to get reallocated (e.g., anything that adds a
5335 * program constant) has to happen before creating this linkage.
5337 _mesa_associate_uniform_storage(ctx
, shader_program
, prog
->Parameters
);
5338 if (!shader_program
->LinkStatus
) {
5342 struct st_vertex_program
*stvp
;
5343 struct st_fragment_program
*stfp
;
5344 struct st_geometry_program
*stgp
;
5346 switch (shader
->Type
) {
5347 case GL_VERTEX_SHADER
:
5348 stvp
= (struct st_vertex_program
*)prog
;
5349 stvp
->glsl_to_tgsi
= v
;
5351 case GL_FRAGMENT_SHADER
:
5352 stfp
= (struct st_fragment_program
*)prog
;
5353 stfp
->glsl_to_tgsi
= v
;
5355 case GL_GEOMETRY_SHADER
:
5356 stgp
= (struct st_geometry_program
*)prog
;
5357 stgp
->glsl_to_tgsi
= v
;
5358 stgp
->Base
.InputType
= shader_program
->Geom
.InputType
;
5359 stgp
->Base
.OutputType
= shader_program
->Geom
.OutputType
;
5360 stgp
->Base
.VerticesOut
= shader_program
->Geom
.VerticesOut
;
5361 stgp
->Base
.Invocations
= shader_program
->Geom
.Invocations
;
5364 assert(!"should not be reached");
5374 st_new_shader(struct gl_context
*ctx
, GLuint name
, GLuint type
)
5376 struct gl_shader
*shader
;
5377 assert(type
== GL_FRAGMENT_SHADER
|| type
== GL_VERTEX_SHADER
||
5378 type
== GL_GEOMETRY_SHADER_ARB
);
5379 shader
= rzalloc(NULL
, struct gl_shader
);
5381 shader
->Type
= type
;
5382 shader
->Stage
= _mesa_shader_enum_to_shader_stage(type
);
5383 shader
->Name
= name
;
5384 _mesa_init_shader(ctx
, shader
);
5389 struct gl_shader_program
*
5390 st_new_shader_program(struct gl_context
*ctx
, GLuint name
)
5392 struct gl_shader_program
*shProg
;
5393 shProg
= rzalloc(NULL
, struct gl_shader_program
);
5395 shProg
->Name
= name
;
5396 _mesa_init_shader_program(ctx
, shProg
);
5403 * Called via ctx->Driver.LinkShader()
5404 * This actually involves converting GLSL IR into an intermediate TGSI-like IR
5405 * with code lowering and other optimizations.
5408 st_link_shader(struct gl_context
*ctx
, struct gl_shader_program
*prog
)
5410 struct pipe_screen
*pscreen
= ctx
->st
->pipe
->screen
;
5411 assert(prog
->LinkStatus
);
5413 for (unsigned i
= 0; i
< MESA_SHADER_STAGES
; i
++) {
5414 if (prog
->_LinkedShaders
[i
] == NULL
)
5418 exec_list
*ir
= prog
->_LinkedShaders
[i
]->ir
;
5419 const struct gl_shader_compiler_options
*options
=
5420 &ctx
->ShaderCompilerOptions
[_mesa_shader_enum_to_shader_stage(prog
->_LinkedShaders
[i
]->Type
)];
5422 /* If there are forms of indirect addressing that the driver
5423 * cannot handle, perform the lowering pass.
5425 if (options
->EmitNoIndirectInput
|| options
->EmitNoIndirectOutput
||
5426 options
->EmitNoIndirectTemp
|| options
->EmitNoIndirectUniform
) {
5427 lower_variable_index_to_cond_assign(ir
,
5428 options
->EmitNoIndirectInput
,
5429 options
->EmitNoIndirectOutput
,
5430 options
->EmitNoIndirectTemp
,
5431 options
->EmitNoIndirectUniform
);
5434 if (ctx
->Extensions
.ARB_shading_language_packing
) {
5435 unsigned lower_inst
= LOWER_PACK_SNORM_2x16
|
5436 LOWER_UNPACK_SNORM_2x16
|
5437 LOWER_PACK_UNORM_2x16
|
5438 LOWER_UNPACK_UNORM_2x16
|
5439 LOWER_PACK_SNORM_4x8
|
5440 LOWER_UNPACK_SNORM_4x8
|
5441 LOWER_UNPACK_UNORM_4x8
|
5442 LOWER_PACK_UNORM_4x8
|
5443 LOWER_PACK_HALF_2x16
|
5444 LOWER_UNPACK_HALF_2x16
;
5446 lower_packing_builtins(ir
, lower_inst
);
5449 if (!pscreen
->get_param(pscreen
, PIPE_CAP_TEXTURE_GATHER_OFFSETS
))
5450 lower_offset_arrays(ir
);
5451 do_mat_op_to_vec(ir
);
5452 lower_instructions(ir
,
5460 (options
->EmitNoPow
? POW_TO_EXP2
: 0) |
5461 (!ctx
->Const
.NativeIntegers
? INT_DIV_TO_MUL_RCP
: 0));
5463 lower_ubo_reference(prog
->_LinkedShaders
[i
], ir
);
5464 do_vec_index_to_cond_assign(ir
);
5465 lower_vector_insert(ir
, true);
5466 lower_quadop_vector(ir
, false);
5468 if (options
->MaxIfDepth
== 0) {
5475 progress
= do_lower_jumps(ir
, true, true, options
->EmitNoMainReturn
, options
->EmitNoCont
, options
->EmitNoLoops
) || progress
;
5477 progress
= do_common_optimization(ir
, true, true, options
,
5478 ctx
->Const
.NativeIntegers
)
5481 progress
= lower_if_to_cond_assign(ir
, options
->MaxIfDepth
) || progress
;
5485 validate_ir_tree(ir
);
5488 for (unsigned i
= 0; i
< MESA_SHADER_STAGES
; i
++) {
5489 struct gl_program
*linked_prog
;
5491 if (prog
->_LinkedShaders
[i
] == NULL
)
5494 linked_prog
= get_mesa_program(ctx
, prog
, prog
->_LinkedShaders
[i
]);
5497 _mesa_reference_program(ctx
, &prog
->_LinkedShaders
[i
]->Program
,
5499 if (!ctx
->Driver
.ProgramStringNotify(ctx
,
5500 _mesa_shader_stage_to_program(i
),
5502 _mesa_reference_program(ctx
, &prog
->_LinkedShaders
[i
]->Program
,
5504 _mesa_reference_program(ctx
, &linked_prog
, NULL
);
5509 _mesa_reference_program(ctx
, &linked_prog
, NULL
);
5516 st_translate_stream_output_info(glsl_to_tgsi_visitor
*glsl_to_tgsi
,
5517 const GLuint outputMapping
[],
5518 struct pipe_stream_output_info
*so
)
5521 struct gl_transform_feedback_info
*info
=
5522 &glsl_to_tgsi
->shader_program
->LinkedTransformFeedback
;
5524 for (i
= 0; i
< info
->NumOutputs
; i
++) {
5525 so
->output
[i
].register_index
=
5526 outputMapping
[info
->Outputs
[i
].OutputRegister
];
5527 so
->output
[i
].start_component
= info
->Outputs
[i
].ComponentOffset
;
5528 so
->output
[i
].num_components
= info
->Outputs
[i
].NumComponents
;
5529 so
->output
[i
].output_buffer
= info
->Outputs
[i
].OutputBuffer
;
5530 so
->output
[i
].dst_offset
= info
->Outputs
[i
].DstOffset
;
5531 so
->output
[i
].stream
= info
->Outputs
[i
].StreamId
;
5534 for (i
= 0; i
< PIPE_MAX_SO_BUFFERS
; i
++) {
5535 so
->stride
[i
] = info
->BufferStride
[i
];
5537 so
->num_outputs
= info
->NumOutputs
;