2 * Copyright (C) 2005-2007 Brian Paul All Rights Reserved.
3 * Copyright (C) 2008 VMware, Inc. All Rights Reserved.
4 * Copyright © 2010 Intel Corporation
5 * Copyright © 2011 Bryan Cain
7 * Permission is hereby granted, free of charge, to any person obtaining a
8 * copy of this software and associated documentation files (the "Software"),
9 * to deal in the Software without restriction, including without limitation
10 * the rights to use, copy, modify, merge, publish, distribute, sublicense,
11 * and/or sell copies of the Software, and to permit persons to whom the
12 * Software is furnished to do so, subject to the following conditions:
14 * The above copyright notice and this permission notice (including the next
15 * paragraph) shall be included in all copies or substantial portions of the
18 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
19 * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
20 * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL
21 * THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
22 * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING
23 * FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER
24 * DEALINGS IN THE SOFTWARE.
28 * \file glsl_to_tgsi.cpp
30 * Translate GLSL IR to TGSI.
34 #include "main/compiler.h"
36 #include "ir_visitor.h"
37 #include "ir_print_visitor.h"
38 #include "ir_expression_flattening.h"
39 #include "glsl_types.h"
40 #include "glsl_parser_extras.h"
41 #include "../glsl/program.h"
42 #include "ir_optimization.h"
45 #include "main/mtypes.h"
46 #include "main/shaderobj.h"
47 #include "program/hash_table.h"
50 #include "main/shaderapi.h"
51 #include "main/uniforms.h"
52 #include "program/prog_instruction.h"
53 #include "program/prog_optimize.h"
54 #include "program/prog_print.h"
55 #include "program/program.h"
56 #include "program/prog_parameter.h"
57 #include "program/sampler.h"
59 #include "pipe/p_compiler.h"
60 #include "pipe/p_context.h"
61 #include "pipe/p_screen.h"
62 #include "pipe/p_shader_tokens.h"
63 #include "pipe/p_state.h"
64 #include "util/u_math.h"
65 #include "tgsi/tgsi_ureg.h"
66 #include "tgsi/tgsi_info.h"
67 #include "st_context.h"
68 #include "st_program.h"
69 #include "st_glsl_to_tgsi.h"
70 #include "st_mesa_to_tgsi.h"
73 #define PROGRAM_IMMEDIATE PROGRAM_FILE_MAX
74 #define PROGRAM_ANY_CONST ((1 << PROGRAM_LOCAL_PARAM) | \
75 (1 << PROGRAM_ENV_PARAM) | \
76 (1 << PROGRAM_STATE_VAR) | \
77 (1 << PROGRAM_CONSTANT) | \
78 (1 << PROGRAM_UNIFORM))
81 * Maximum number of temporary registers.
83 * It is too big for stack allocated arrays -- it will cause stack overflow on
84 * Windows and likely Mac OS X.
86 #define MAX_TEMPS 4096
88 /* will be 4 for GLSL 4.00 */
89 #define MAX_GLSL_TEXTURE_OFFSET 1
94 static int swizzle_for_size(int size
);
97 * This struct is a corresponding struct to TGSI ureg_src.
101 st_src_reg(gl_register_file file
, int index
, const glsl_type
*type
)
105 if (type
&& (type
->is_scalar() || type
->is_vector() || type
->is_matrix()))
106 this->swizzle
= swizzle_for_size(type
->vector_elements
);
108 this->swizzle
= SWIZZLE_XYZW
;
111 this->type
= type
? type
->base_type
: GLSL_TYPE_ERROR
;
112 this->reladdr
= NULL
;
115 st_src_reg(gl_register_file file
, int index
, int type
)
121 this->swizzle
= SWIZZLE_XYZW
;
123 this->reladdr
= NULL
;
126 st_src_reg(gl_register_file file
, int index
, int type
, int index2D
)
131 this->index2D
= index2D
;
132 this->swizzle
= SWIZZLE_XYZW
;
134 this->reladdr
= NULL
;
139 this->type
= GLSL_TYPE_ERROR
;
140 this->file
= PROGRAM_UNDEFINED
;
145 this->reladdr
= NULL
;
148 explicit st_src_reg(st_dst_reg reg
);
150 gl_register_file file
; /**< PROGRAM_* from Mesa */
151 int index
; /**< temporary index, VERT_ATTRIB_*, FRAG_ATTRIB_*, etc. */
153 GLuint swizzle
; /**< SWIZZLE_XYZWONEZERO swizzles from Mesa. */
154 int negate
; /**< NEGATE_XYZW mask from mesa */
155 int type
; /** GLSL_TYPE_* from GLSL IR (enum glsl_base_type) */
156 /** Register index should be offset by the integer in this reg. */
162 st_dst_reg(gl_register_file file
, int writemask
, int type
)
166 this->writemask
= writemask
;
167 this->cond_mask
= COND_TR
;
168 this->reladdr
= NULL
;
174 this->type
= GLSL_TYPE_ERROR
;
175 this->file
= PROGRAM_UNDEFINED
;
178 this->cond_mask
= COND_TR
;
179 this->reladdr
= NULL
;
182 explicit st_dst_reg(st_src_reg reg
);
184 gl_register_file file
; /**< PROGRAM_* from Mesa */
185 int index
; /**< temporary index, VERT_ATTRIB_*, FRAG_ATTRIB_*, etc. */
186 int writemask
; /**< Bitfield of WRITEMASK_[XYZW] */
188 int type
; /** GLSL_TYPE_* from GLSL IR (enum glsl_base_type) */
189 /** Register index should be offset by the integer in this reg. */
193 st_src_reg::st_src_reg(st_dst_reg reg
)
195 this->type
= reg
.type
;
196 this->file
= reg
.file
;
197 this->index
= reg
.index
;
198 this->swizzle
= SWIZZLE_XYZW
;
200 this->reladdr
= reg
.reladdr
;
204 st_dst_reg::st_dst_reg(st_src_reg reg
)
206 this->type
= reg
.type
;
207 this->file
= reg
.file
;
208 this->index
= reg
.index
;
209 this->writemask
= WRITEMASK_XYZW
;
210 this->cond_mask
= COND_TR
;
211 this->reladdr
= reg
.reladdr
;
214 class glsl_to_tgsi_instruction
: public exec_node
{
216 /* Callers of this ralloc-based new need not call delete. It's
217 * easier to just ralloc_free 'ctx' (or any of its ancestors). */
218 static void* operator new(size_t size
, void *ctx
)
222 node
= rzalloc_size(ctx
, size
);
223 assert(node
!= NULL
);
231 /** Pointer to the ir source this tree came from for debugging */
233 GLboolean cond_update
;
235 int sampler
; /**< sampler index */
236 int tex_target
; /**< One of TEXTURE_*_INDEX */
237 GLboolean tex_shadow
;
238 struct tgsi_texture_offset tex_offsets
[MAX_GLSL_TEXTURE_OFFSET
];
239 unsigned tex_offset_num_offset
;
240 int dead_mask
; /**< Used in dead code elimination */
242 class function_entry
*function
; /* Set on TGSI_OPCODE_CAL or TGSI_OPCODE_BGNSUB */
245 class variable_storage
: public exec_node
{
247 variable_storage(ir_variable
*var
, gl_register_file file
, int index
)
248 : file(file
), index(index
), var(var
)
253 gl_register_file file
;
255 ir_variable
*var
; /* variable that maps to this, if any */
258 class immediate_storage
: public exec_node
{
260 immediate_storage(gl_constant_value
*values
, int size
, int type
)
262 memcpy(this->values
, values
, size
* sizeof(gl_constant_value
));
267 gl_constant_value values
[4];
268 int size
; /**< Number of components (1-4) */
269 int type
; /**< GL_FLOAT, GL_INT, GL_BOOL, or GL_UNSIGNED_INT */
272 class function_entry
: public exec_node
{
274 ir_function_signature
*sig
;
277 * identifier of this function signature used by the program.
279 * At the point that TGSI instructions for function calls are
280 * generated, we don't know the address of the first instruction of
281 * the function body. So we make the BranchTarget that is called a
282 * small integer and rewrite them during set_branchtargets().
287 * Pointer to first instruction of the function body.
289 * Set during function body emits after main() is processed.
291 glsl_to_tgsi_instruction
*bgn_inst
;
294 * Index of the first instruction of the function body in actual TGSI.
296 * Set after conversion from glsl_to_tgsi_instruction to TGSI.
300 /** Storage for the return value. */
301 st_src_reg return_reg
;
304 struct glsl_to_tgsi_visitor
: public ir_visitor
{
306 glsl_to_tgsi_visitor();
307 ~glsl_to_tgsi_visitor();
309 function_entry
*current_function
;
311 struct gl_context
*ctx
;
312 struct gl_program
*prog
;
313 struct gl_shader_program
*shader_program
;
314 struct gl_shader_compiler_options
*options
;
318 int num_address_regs
;
320 bool indirect_addr_temps
;
321 bool indirect_addr_consts
;
324 bool native_integers
;
326 variable_storage
*find_variable_storage(ir_variable
*var
);
328 int add_constant(gl_register_file file
, gl_constant_value values
[4],
329 int size
, int datatype
, GLuint
*swizzle_out
);
331 function_entry
*get_function_signature(ir_function_signature
*sig
);
333 st_src_reg
get_temp(const glsl_type
*type
);
334 void reladdr_to_temp(ir_instruction
*ir
, st_src_reg
*reg
, int *num_reladdr
);
336 st_src_reg
st_src_reg_for_float(float val
);
337 st_src_reg
st_src_reg_for_int(int val
);
338 st_src_reg
st_src_reg_for_type(int type
, int val
);
341 * \name Visit methods
343 * As typical for the visitor pattern, there must be one \c visit method for
344 * each concrete subclass of \c ir_instruction. Virtual base classes within
345 * the hierarchy should not have \c visit methods.
348 virtual void visit(ir_variable
*);
349 virtual void visit(ir_loop
*);
350 virtual void visit(ir_loop_jump
*);
351 virtual void visit(ir_function_signature
*);
352 virtual void visit(ir_function
*);
353 virtual void visit(ir_expression
*);
354 virtual void visit(ir_swizzle
*);
355 virtual void visit(ir_dereference_variable
*);
356 virtual void visit(ir_dereference_array
*);
357 virtual void visit(ir_dereference_record
*);
358 virtual void visit(ir_assignment
*);
359 virtual void visit(ir_constant
*);
360 virtual void visit(ir_call
*);
361 virtual void visit(ir_return
*);
362 virtual void visit(ir_discard
*);
363 virtual void visit(ir_texture
*);
364 virtual void visit(ir_if
*);
369 /** List of variable_storage */
372 /** List of immediate_storage */
373 exec_list immediates
;
374 unsigned num_immediates
;
376 /** List of function_entry */
377 exec_list function_signatures
;
378 int next_signature_id
;
380 /** List of glsl_to_tgsi_instruction */
381 exec_list instructions
;
383 glsl_to_tgsi_instruction
*emit(ir_instruction
*ir
, unsigned op
);
385 glsl_to_tgsi_instruction
*emit(ir_instruction
*ir
, unsigned op
,
386 st_dst_reg dst
, st_src_reg src0
);
388 glsl_to_tgsi_instruction
*emit(ir_instruction
*ir
, unsigned op
,
389 st_dst_reg dst
, st_src_reg src0
, st_src_reg src1
);
391 glsl_to_tgsi_instruction
*emit(ir_instruction
*ir
, unsigned op
,
393 st_src_reg src0
, st_src_reg src1
, st_src_reg src2
);
395 unsigned get_opcode(ir_instruction
*ir
, unsigned op
,
397 st_src_reg src0
, st_src_reg src1
);
400 * Emit the correct dot-product instruction for the type of arguments
402 glsl_to_tgsi_instruction
*emit_dp(ir_instruction
*ir
,
408 void emit_scalar(ir_instruction
*ir
, unsigned op
,
409 st_dst_reg dst
, st_src_reg src0
);
411 void emit_scalar(ir_instruction
*ir
, unsigned op
,
412 st_dst_reg dst
, st_src_reg src0
, st_src_reg src1
);
414 void try_emit_float_set(ir_instruction
*ir
, unsigned op
, st_dst_reg dst
);
416 void emit_arl(ir_instruction
*ir
, st_dst_reg dst
, st_src_reg src0
);
418 void emit_scs(ir_instruction
*ir
, unsigned op
,
419 st_dst_reg dst
, const st_src_reg
&src
);
421 bool try_emit_mad(ir_expression
*ir
,
423 bool try_emit_mad_for_and_not(ir_expression
*ir
,
425 bool try_emit_sat(ir_expression
*ir
);
427 void emit_swz(ir_expression
*ir
);
429 bool process_move_condition(ir_rvalue
*ir
);
431 void simplify_cmp(void);
433 void rename_temp_register(int index
, int new_index
);
434 int get_first_temp_read(int index
);
435 int get_first_temp_write(int index
);
436 int get_last_temp_read(int index
);
437 int get_last_temp_write(int index
);
439 void copy_propagate(void);
440 void eliminate_dead_code(void);
441 int eliminate_dead_code_advanced(void);
442 void merge_registers(void);
443 void renumber_registers(void);
448 static st_src_reg undef_src
= st_src_reg(PROGRAM_UNDEFINED
, 0, GLSL_TYPE_ERROR
);
450 static st_dst_reg undef_dst
= st_dst_reg(PROGRAM_UNDEFINED
, SWIZZLE_NOOP
, GLSL_TYPE_ERROR
);
452 static st_dst_reg address_reg
= st_dst_reg(PROGRAM_ADDRESS
, WRITEMASK_X
, GLSL_TYPE_FLOAT
);
455 fail_link(struct gl_shader_program
*prog
, const char *fmt
, ...) PRINTFLIKE(2, 3);
458 fail_link(struct gl_shader_program
*prog
, const char *fmt
, ...)
462 ralloc_vasprintf_append(&prog
->InfoLog
, fmt
, args
);
465 prog
->LinkStatus
= GL_FALSE
;
469 swizzle_for_size(int size
)
471 int size_swizzles
[4] = {
472 MAKE_SWIZZLE4(SWIZZLE_X
, SWIZZLE_X
, SWIZZLE_X
, SWIZZLE_X
),
473 MAKE_SWIZZLE4(SWIZZLE_X
, SWIZZLE_Y
, SWIZZLE_Y
, SWIZZLE_Y
),
474 MAKE_SWIZZLE4(SWIZZLE_X
, SWIZZLE_Y
, SWIZZLE_Z
, SWIZZLE_Z
),
475 MAKE_SWIZZLE4(SWIZZLE_X
, SWIZZLE_Y
, SWIZZLE_Z
, SWIZZLE_W
),
478 assert((size
>= 1) && (size
<= 4));
479 return size_swizzles
[size
- 1];
483 is_tex_instruction(unsigned opcode
)
485 const tgsi_opcode_info
* info
= tgsi_get_opcode_info(opcode
);
490 num_inst_dst_regs(unsigned opcode
)
492 const tgsi_opcode_info
* info
= tgsi_get_opcode_info(opcode
);
493 return info
->num_dst
;
497 num_inst_src_regs(unsigned opcode
)
499 const tgsi_opcode_info
* info
= tgsi_get_opcode_info(opcode
);
500 return info
->is_tex
? info
->num_src
- 1 : info
->num_src
;
503 glsl_to_tgsi_instruction
*
504 glsl_to_tgsi_visitor::emit(ir_instruction
*ir
, unsigned op
,
506 st_src_reg src0
, st_src_reg src1
, st_src_reg src2
)
508 glsl_to_tgsi_instruction
*inst
= new(mem_ctx
) glsl_to_tgsi_instruction();
509 int num_reladdr
= 0, i
;
511 op
= get_opcode(ir
, op
, dst
, src0
, src1
);
513 /* If we have to do relative addressing, we want to load the ARL
514 * reg directly for one of the regs, and preload the other reladdr
515 * sources into temps.
517 num_reladdr
+= dst
.reladdr
!= NULL
;
518 num_reladdr
+= src0
.reladdr
!= NULL
;
519 num_reladdr
+= src1
.reladdr
!= NULL
;
520 num_reladdr
+= src2
.reladdr
!= NULL
;
522 reladdr_to_temp(ir
, &src2
, &num_reladdr
);
523 reladdr_to_temp(ir
, &src1
, &num_reladdr
);
524 reladdr_to_temp(ir
, &src0
, &num_reladdr
);
527 emit_arl(ir
, address_reg
, *dst
.reladdr
);
530 assert(num_reladdr
== 0);
540 inst
->function
= NULL
;
542 if (op
== TGSI_OPCODE_ARL
|| op
== TGSI_OPCODE_UARL
)
543 this->num_address_regs
= 1;
545 /* Update indirect addressing status used by TGSI */
548 case PROGRAM_TEMPORARY
:
549 this->indirect_addr_temps
= true;
551 case PROGRAM_LOCAL_PARAM
:
552 case PROGRAM_ENV_PARAM
:
553 case PROGRAM_STATE_VAR
:
554 case PROGRAM_CONSTANT
:
555 case PROGRAM_UNIFORM
:
556 this->indirect_addr_consts
= true;
558 case PROGRAM_IMMEDIATE
:
559 assert(!"immediates should not have indirect addressing");
566 for (i
=0; i
<3; i
++) {
567 if(inst
->src
[i
].reladdr
) {
568 switch(inst
->src
[i
].file
) {
569 case PROGRAM_TEMPORARY
:
570 this->indirect_addr_temps
= true;
572 case PROGRAM_LOCAL_PARAM
:
573 case PROGRAM_ENV_PARAM
:
574 case PROGRAM_STATE_VAR
:
575 case PROGRAM_CONSTANT
:
576 case PROGRAM_UNIFORM
:
577 this->indirect_addr_consts
= true;
579 case PROGRAM_IMMEDIATE
:
580 assert(!"immediates should not have indirect addressing");
589 this->instructions
.push_tail(inst
);
592 try_emit_float_set(ir
, op
, dst
);
598 glsl_to_tgsi_instruction
*
599 glsl_to_tgsi_visitor::emit(ir_instruction
*ir
, unsigned op
,
600 st_dst_reg dst
, st_src_reg src0
, st_src_reg src1
)
602 return emit(ir
, op
, dst
, src0
, src1
, undef_src
);
605 glsl_to_tgsi_instruction
*
606 glsl_to_tgsi_visitor::emit(ir_instruction
*ir
, unsigned op
,
607 st_dst_reg dst
, st_src_reg src0
)
609 assert(dst
.writemask
!= 0);
610 return emit(ir
, op
, dst
, src0
, undef_src
, undef_src
);
613 glsl_to_tgsi_instruction
*
614 glsl_to_tgsi_visitor::emit(ir_instruction
*ir
, unsigned op
)
616 return emit(ir
, op
, undef_dst
, undef_src
, undef_src
, undef_src
);
620 * Emits the code to convert the result of float SET instructions to integers.
623 glsl_to_tgsi_visitor::try_emit_float_set(ir_instruction
*ir
, unsigned op
,
626 if ((op
== TGSI_OPCODE_SEQ
||
627 op
== TGSI_OPCODE_SNE
||
628 op
== TGSI_OPCODE_SGE
||
629 op
== TGSI_OPCODE_SLT
))
631 st_src_reg src
= st_src_reg(dst
);
632 src
.negate
= ~src
.negate
;
633 dst
.type
= GLSL_TYPE_FLOAT
;
634 emit(ir
, TGSI_OPCODE_F2I
, dst
, src
);
639 * Determines whether to use an integer, unsigned integer, or float opcode
640 * based on the operands and input opcode, then emits the result.
643 glsl_to_tgsi_visitor::get_opcode(ir_instruction
*ir
, unsigned op
,
645 st_src_reg src0
, st_src_reg src1
)
647 int type
= GLSL_TYPE_FLOAT
;
649 assert(src0
.type
!= GLSL_TYPE_ARRAY
);
650 assert(src0
.type
!= GLSL_TYPE_STRUCT
);
651 assert(src1
.type
!= GLSL_TYPE_ARRAY
);
652 assert(src1
.type
!= GLSL_TYPE_STRUCT
);
654 if (src0
.type
== GLSL_TYPE_FLOAT
|| src1
.type
== GLSL_TYPE_FLOAT
)
655 type
= GLSL_TYPE_FLOAT
;
656 else if (native_integers
)
657 type
= src0
.type
== GLSL_TYPE_BOOL
? GLSL_TYPE_INT
: src0
.type
;
659 #define case4(c, f, i, u) \
660 case TGSI_OPCODE_##c: \
661 if (type == GLSL_TYPE_INT) op = TGSI_OPCODE_##i; \
662 else if (type == GLSL_TYPE_UINT) op = TGSI_OPCODE_##u; \
663 else op = TGSI_OPCODE_##f; \
665 #define case3(f, i, u) case4(f, f, i, u)
666 #define case2fi(f, i) case4(f, f, i, i)
667 #define case2iu(i, u) case4(i, LAST, i, u)
673 case3(DIV
, IDIV
, UDIV
);
674 case3(MAX
, IMAX
, UMAX
);
675 case3(MIN
, IMIN
, UMIN
);
680 case3(SGE
, ISGE
, USGE
);
681 case3(SLT
, ISLT
, USLT
);
686 case3(ABS
, IABS
, IABS
);
691 assert(op
!= TGSI_OPCODE_LAST
);
695 glsl_to_tgsi_instruction
*
696 glsl_to_tgsi_visitor::emit_dp(ir_instruction
*ir
,
697 st_dst_reg dst
, st_src_reg src0
, st_src_reg src1
,
700 static const unsigned dot_opcodes
[] = {
701 TGSI_OPCODE_DP2
, TGSI_OPCODE_DP3
, TGSI_OPCODE_DP4
704 return emit(ir
, dot_opcodes
[elements
- 2], dst
, src0
, src1
);
708 * Emits TGSI scalar opcodes to produce unique answers across channels.
710 * Some TGSI opcodes are scalar-only, like ARB_fp/vp. The src X
711 * channel determines the result across all channels. So to do a vec4
712 * of this operation, we want to emit a scalar per source channel used
713 * to produce dest channels.
716 glsl_to_tgsi_visitor::emit_scalar(ir_instruction
*ir
, unsigned op
,
718 st_src_reg orig_src0
, st_src_reg orig_src1
)
721 int done_mask
= ~dst
.writemask
;
723 /* TGSI RCP is a scalar operation splatting results to all channels,
724 * like ARB_fp/vp. So emit as many RCPs as necessary to cover our
727 for (i
= 0; i
< 4; i
++) {
728 GLuint this_mask
= (1 << i
);
729 glsl_to_tgsi_instruction
*inst
;
730 st_src_reg src0
= orig_src0
;
731 st_src_reg src1
= orig_src1
;
733 if (done_mask
& this_mask
)
736 GLuint src0_swiz
= GET_SWZ(src0
.swizzle
, i
);
737 GLuint src1_swiz
= GET_SWZ(src1
.swizzle
, i
);
738 for (j
= i
+ 1; j
< 4; j
++) {
739 /* If there is another enabled component in the destination that is
740 * derived from the same inputs, generate its value on this pass as
743 if (!(done_mask
& (1 << j
)) &&
744 GET_SWZ(src0
.swizzle
, j
) == src0_swiz
&&
745 GET_SWZ(src1
.swizzle
, j
) == src1_swiz
) {
746 this_mask
|= (1 << j
);
749 src0
.swizzle
= MAKE_SWIZZLE4(src0_swiz
, src0_swiz
,
750 src0_swiz
, src0_swiz
);
751 src1
.swizzle
= MAKE_SWIZZLE4(src1_swiz
, src1_swiz
,
752 src1_swiz
, src1_swiz
);
754 inst
= emit(ir
, op
, dst
, src0
, src1
);
755 inst
->dst
.writemask
= this_mask
;
756 done_mask
|= this_mask
;
761 glsl_to_tgsi_visitor::emit_scalar(ir_instruction
*ir
, unsigned op
,
762 st_dst_reg dst
, st_src_reg src0
)
764 st_src_reg undef
= undef_src
;
766 undef
.swizzle
= SWIZZLE_XXXX
;
768 emit_scalar(ir
, op
, dst
, src0
, undef
);
772 glsl_to_tgsi_visitor::emit_arl(ir_instruction
*ir
,
773 st_dst_reg dst
, st_src_reg src0
)
775 int op
= TGSI_OPCODE_ARL
;
777 if (src0
.type
== GLSL_TYPE_INT
|| src0
.type
== GLSL_TYPE_UINT
)
778 op
= TGSI_OPCODE_UARL
;
780 emit(NULL
, op
, dst
, src0
);
784 * Emit an TGSI_OPCODE_SCS instruction
786 * The \c SCS opcode functions a bit differently than the other TGSI opcodes.
787 * Instead of splatting its result across all four components of the
788 * destination, it writes one value to the \c x component and another value to
789 * the \c y component.
791 * \param ir IR instruction being processed
792 * \param op Either \c TGSI_OPCODE_SIN or \c TGSI_OPCODE_COS depending
793 * on which value is desired.
794 * \param dst Destination register
795 * \param src Source register
798 glsl_to_tgsi_visitor::emit_scs(ir_instruction
*ir
, unsigned op
,
800 const st_src_reg
&src
)
802 /* Vertex programs cannot use the SCS opcode.
804 if (this->prog
->Target
== GL_VERTEX_PROGRAM_ARB
) {
805 emit_scalar(ir
, op
, dst
, src
);
809 const unsigned component
= (op
== TGSI_OPCODE_SIN
) ? 0 : 1;
810 const unsigned scs_mask
= (1U << component
);
811 int done_mask
= ~dst
.writemask
;
814 assert(op
== TGSI_OPCODE_SIN
|| op
== TGSI_OPCODE_COS
);
816 /* If there are compnents in the destination that differ from the component
817 * that will be written by the SCS instrution, we'll need a temporary.
819 if (scs_mask
!= unsigned(dst
.writemask
)) {
820 tmp
= get_temp(glsl_type::vec4_type
);
823 for (unsigned i
= 0; i
< 4; i
++) {
824 unsigned this_mask
= (1U << i
);
825 st_src_reg src0
= src
;
827 if ((done_mask
& this_mask
) != 0)
830 /* The source swizzle specified which component of the source generates
831 * sine / cosine for the current component in the destination. The SCS
832 * instruction requires that this value be swizzle to the X component.
833 * Replace the current swizzle with a swizzle that puts the source in
836 unsigned src0_swiz
= GET_SWZ(src
.swizzle
, i
);
838 src0
.swizzle
= MAKE_SWIZZLE4(src0_swiz
, src0_swiz
,
839 src0_swiz
, src0_swiz
);
840 for (unsigned j
= i
+ 1; j
< 4; j
++) {
841 /* If there is another enabled component in the destination that is
842 * derived from the same inputs, generate its value on this pass as
845 if (!(done_mask
& (1 << j
)) &&
846 GET_SWZ(src0
.swizzle
, j
) == src0_swiz
) {
847 this_mask
|= (1 << j
);
851 if (this_mask
!= scs_mask
) {
852 glsl_to_tgsi_instruction
*inst
;
853 st_dst_reg tmp_dst
= st_dst_reg(tmp
);
855 /* Emit the SCS instruction.
857 inst
= emit(ir
, TGSI_OPCODE_SCS
, tmp_dst
, src0
);
858 inst
->dst
.writemask
= scs_mask
;
860 /* Move the result of the SCS instruction to the desired location in
863 tmp
.swizzle
= MAKE_SWIZZLE4(component
, component
,
864 component
, component
);
865 inst
= emit(ir
, TGSI_OPCODE_SCS
, dst
, tmp
);
866 inst
->dst
.writemask
= this_mask
;
868 /* Emit the SCS instruction to write directly to the destination.
870 glsl_to_tgsi_instruction
*inst
= emit(ir
, TGSI_OPCODE_SCS
, dst
, src0
);
871 inst
->dst
.writemask
= scs_mask
;
874 done_mask
|= this_mask
;
879 glsl_to_tgsi_visitor::add_constant(gl_register_file file
,
880 gl_constant_value values
[4], int size
, int datatype
,
883 if (file
== PROGRAM_CONSTANT
) {
884 return _mesa_add_typed_unnamed_constant(this->prog
->Parameters
, values
,
885 size
, datatype
, swizzle_out
);
888 immediate_storage
*entry
;
889 assert(file
== PROGRAM_IMMEDIATE
);
891 /* Search immediate storage to see if we already have an identical
892 * immediate that we can use instead of adding a duplicate entry.
894 foreach_iter(exec_list_iterator
, iter
, this->immediates
) {
895 entry
= (immediate_storage
*)iter
.get();
897 if (entry
->size
== size
&&
898 entry
->type
== datatype
&&
899 !memcmp(entry
->values
, values
, size
* sizeof(gl_constant_value
))) {
905 /* Add this immediate to the list. */
906 entry
= new(mem_ctx
) immediate_storage(values
, size
, datatype
);
907 this->immediates
.push_tail(entry
);
908 this->num_immediates
++;
914 glsl_to_tgsi_visitor::st_src_reg_for_float(float val
)
916 st_src_reg
src(PROGRAM_IMMEDIATE
, -1, GLSL_TYPE_FLOAT
);
917 union gl_constant_value uval
;
920 src
.index
= add_constant(src
.file
, &uval
, 1, GL_FLOAT
, &src
.swizzle
);
926 glsl_to_tgsi_visitor::st_src_reg_for_int(int val
)
928 st_src_reg
src(PROGRAM_IMMEDIATE
, -1, GLSL_TYPE_INT
);
929 union gl_constant_value uval
;
931 assert(native_integers
);
934 src
.index
= add_constant(src
.file
, &uval
, 1, GL_INT
, &src
.swizzle
);
940 glsl_to_tgsi_visitor::st_src_reg_for_type(int type
, int val
)
943 return type
== GLSL_TYPE_FLOAT
? st_src_reg_for_float(val
) :
944 st_src_reg_for_int(val
);
946 return st_src_reg_for_float(val
);
950 type_size(const struct glsl_type
*type
)
955 switch (type
->base_type
) {
958 case GLSL_TYPE_FLOAT
:
960 if (type
->is_matrix()) {
961 return type
->matrix_columns
;
963 /* Regardless of size of vector, it gets a vec4. This is bad
964 * packing for things like floats, but otherwise arrays become a
965 * mess. Hopefully a later pass over the code can pack scalars
966 * down if appropriate.
970 case GLSL_TYPE_ARRAY
:
971 assert(type
->length
> 0);
972 return type_size(type
->fields
.array
) * type
->length
;
973 case GLSL_TYPE_STRUCT
:
975 for (i
= 0; i
< type
->length
; i
++) {
976 size
+= type_size(type
->fields
.structure
[i
].type
);
979 case GLSL_TYPE_SAMPLER
:
980 /* Samplers take up one slot in UNIFORMS[], but they're baked in
991 * In the initial pass of codegen, we assign temporary numbers to
992 * intermediate results. (not SSA -- variable assignments will reuse
996 glsl_to_tgsi_visitor::get_temp(const glsl_type
*type
)
1000 src
.type
= native_integers
? type
->base_type
: GLSL_TYPE_FLOAT
;
1001 src
.file
= PROGRAM_TEMPORARY
;
1002 src
.index
= next_temp
;
1004 next_temp
+= type_size(type
);
1006 if (type
->is_array() || type
->is_record()) {
1007 src
.swizzle
= SWIZZLE_NOOP
;
1009 src
.swizzle
= swizzle_for_size(type
->vector_elements
);
1017 glsl_to_tgsi_visitor::find_variable_storage(ir_variable
*var
)
1020 variable_storage
*entry
;
1022 foreach_iter(exec_list_iterator
, iter
, this->variables
) {
1023 entry
= (variable_storage
*)iter
.get();
1025 if (entry
->var
== var
)
1033 glsl_to_tgsi_visitor::visit(ir_variable
*ir
)
1035 if (strcmp(ir
->name
, "gl_FragCoord") == 0) {
1036 struct gl_fragment_program
*fp
= (struct gl_fragment_program
*)this->prog
;
1038 fp
->OriginUpperLeft
= ir
->origin_upper_left
;
1039 fp
->PixelCenterInteger
= ir
->pixel_center_integer
;
1042 if (ir
->mode
== ir_var_uniform
&& strncmp(ir
->name
, "gl_", 3) == 0) {
1044 const ir_state_slot
*const slots
= ir
->state_slots
;
1045 assert(ir
->state_slots
!= NULL
);
1047 /* Check if this statevar's setup in the STATE file exactly
1048 * matches how we'll want to reference it as a
1049 * struct/array/whatever. If not, then we need to move it into
1050 * temporary storage and hope that it'll get copy-propagated
1053 for (i
= 0; i
< ir
->num_state_slots
; i
++) {
1054 if (slots
[i
].swizzle
!= SWIZZLE_XYZW
) {
1059 variable_storage
*storage
;
1061 if (i
== ir
->num_state_slots
) {
1062 /* We'll set the index later. */
1063 storage
= new(mem_ctx
) variable_storage(ir
, PROGRAM_STATE_VAR
, -1);
1064 this->variables
.push_tail(storage
);
1068 /* The variable_storage constructor allocates slots based on the size
1069 * of the type. However, this had better match the number of state
1070 * elements that we're going to copy into the new temporary.
1072 assert((int) ir
->num_state_slots
== type_size(ir
->type
));
1074 storage
= new(mem_ctx
) variable_storage(ir
, PROGRAM_TEMPORARY
,
1076 this->variables
.push_tail(storage
);
1077 this->next_temp
+= type_size(ir
->type
);
1079 dst
= st_dst_reg(st_src_reg(PROGRAM_TEMPORARY
, storage
->index
,
1080 native_integers
? ir
->type
->base_type
: GLSL_TYPE_FLOAT
));
1084 for (unsigned int i
= 0; i
< ir
->num_state_slots
; i
++) {
1085 int index
= _mesa_add_state_reference(this->prog
->Parameters
,
1086 (gl_state_index
*)slots
[i
].tokens
);
1088 if (storage
->file
== PROGRAM_STATE_VAR
) {
1089 if (storage
->index
== -1) {
1090 storage
->index
= index
;
1092 assert(index
== storage
->index
+ (int)i
);
1095 /* We use GLSL_TYPE_FLOAT here regardless of the actual type of
1096 * the data being moved since MOV does not care about the type of
1097 * data it is moving, and we don't want to declare registers with
1098 * array or struct types.
1100 st_src_reg
src(PROGRAM_STATE_VAR
, index
, GLSL_TYPE_FLOAT
);
1101 src
.swizzle
= slots
[i
].swizzle
;
1102 emit(ir
, TGSI_OPCODE_MOV
, dst
, src
);
1103 /* even a float takes up a whole vec4 reg in a struct/array. */
1108 if (storage
->file
== PROGRAM_TEMPORARY
&&
1109 dst
.index
!= storage
->index
+ (int) ir
->num_state_slots
) {
1110 fail_link(this->shader_program
,
1111 "failed to load builtin uniform `%s' (%d/%d regs loaded)\n",
1112 ir
->name
, dst
.index
- storage
->index
,
1113 type_size(ir
->type
));
1119 glsl_to_tgsi_visitor::visit(ir_loop
*ir
)
1121 ir_dereference_variable
*counter
= NULL
;
1123 if (ir
->counter
!= NULL
)
1124 counter
= new(ir
) ir_dereference_variable(ir
->counter
);
1126 if (ir
->from
!= NULL
) {
1127 assert(ir
->counter
!= NULL
);
1129 ir_assignment
*a
= new(ir
) ir_assignment(counter
, ir
->from
, NULL
);
1135 emit(NULL
, TGSI_OPCODE_BGNLOOP
);
1139 new(ir
) ir_expression(ir
->cmp
, glsl_type::bool_type
,
1141 ir_if
*if_stmt
= new(ir
) ir_if(e
);
1143 ir_loop_jump
*brk
= new(ir
) ir_loop_jump(ir_loop_jump::jump_break
);
1145 if_stmt
->then_instructions
.push_tail(brk
);
1147 if_stmt
->accept(this);
1154 visit_exec_list(&ir
->body_instructions
, this);
1156 if (ir
->increment
) {
1158 new(ir
) ir_expression(ir_binop_add
, counter
->type
,
1159 counter
, ir
->increment
);
1161 ir_assignment
*a
= new(ir
) ir_assignment(counter
, e
, NULL
);
1168 emit(NULL
, TGSI_OPCODE_ENDLOOP
);
1172 glsl_to_tgsi_visitor::visit(ir_loop_jump
*ir
)
1175 case ir_loop_jump::jump_break
:
1176 emit(NULL
, TGSI_OPCODE_BRK
);
1178 case ir_loop_jump::jump_continue
:
1179 emit(NULL
, TGSI_OPCODE_CONT
);
1186 glsl_to_tgsi_visitor::visit(ir_function_signature
*ir
)
1193 glsl_to_tgsi_visitor::visit(ir_function
*ir
)
1195 /* Ignore function bodies other than main() -- we shouldn't see calls to
1196 * them since they should all be inlined before we get to glsl_to_tgsi.
1198 if (strcmp(ir
->name
, "main") == 0) {
1199 const ir_function_signature
*sig
;
1202 sig
= ir
->matching_signature(&empty
);
1206 foreach_iter(exec_list_iterator
, iter
, sig
->body
) {
1207 ir_instruction
*ir
= (ir_instruction
*)iter
.get();
1215 glsl_to_tgsi_visitor::try_emit_mad(ir_expression
*ir
, int mul_operand
)
1217 int nonmul_operand
= 1 - mul_operand
;
1219 st_dst_reg result_dst
;
1221 ir_expression
*expr
= ir
->operands
[mul_operand
]->as_expression();
1222 if (!expr
|| expr
->operation
!= ir_binop_mul
)
1225 expr
->operands
[0]->accept(this);
1227 expr
->operands
[1]->accept(this);
1229 ir
->operands
[nonmul_operand
]->accept(this);
1232 this->result
= get_temp(ir
->type
);
1233 result_dst
= st_dst_reg(this->result
);
1234 result_dst
.writemask
= (1 << ir
->type
->vector_elements
) - 1;
1235 emit(ir
, TGSI_OPCODE_MAD
, result_dst
, a
, b
, c
);
1241 * Emit MAD(a, -b, a) instead of AND(a, NOT(b))
1243 * The logic values are 1.0 for true and 0.0 for false. Logical-and is
1244 * implemented using multiplication, and logical-or is implemented using
1245 * addition. Logical-not can be implemented as (true - x), or (1.0 - x).
1246 * As result, the logical expression (a & !b) can be rewritten as:
1250 * - (a * 1) - (a * b)
1254 * This final expression can be implemented as a single MAD(a, -b, a)
1258 glsl_to_tgsi_visitor::try_emit_mad_for_and_not(ir_expression
*ir
, int try_operand
)
1260 const int other_operand
= 1 - try_operand
;
1263 ir_expression
*expr
= ir
->operands
[try_operand
]->as_expression();
1264 if (!expr
|| expr
->operation
!= ir_unop_logic_not
)
1267 ir
->operands
[other_operand
]->accept(this);
1269 expr
->operands
[0]->accept(this);
1272 b
.negate
= ~b
.negate
;
1274 this->result
= get_temp(ir
->type
);
1275 emit(ir
, TGSI_OPCODE_MAD
, st_dst_reg(this->result
), a
, b
, a
);
1281 glsl_to_tgsi_visitor::try_emit_sat(ir_expression
*ir
)
1283 /* Saturates were only introduced to vertex programs in
1284 * NV_vertex_program3, so don't give them to drivers in the VP.
1286 if (this->prog
->Target
== GL_VERTEX_PROGRAM_ARB
)
1289 ir_rvalue
*sat_src
= ir
->as_rvalue_to_saturate();
1293 sat_src
->accept(this);
1294 st_src_reg src
= this->result
;
1296 /* If we generated an expression instruction into a temporary in
1297 * processing the saturate's operand, apply the saturate to that
1298 * instruction. Otherwise, generate a MOV to do the saturate.
1300 * Note that we have to be careful to only do this optimization if
1301 * the instruction in question was what generated src->result. For
1302 * example, ir_dereference_array might generate a MUL instruction
1303 * to create the reladdr, and return us a src reg using that
1304 * reladdr. That MUL result is not the value we're trying to
1307 ir_expression
*sat_src_expr
= sat_src
->as_expression();
1308 if (sat_src_expr
&& (sat_src_expr
->operation
== ir_binop_mul
||
1309 sat_src_expr
->operation
== ir_binop_add
||
1310 sat_src_expr
->operation
== ir_binop_dot
)) {
1311 glsl_to_tgsi_instruction
*new_inst
;
1312 new_inst
= (glsl_to_tgsi_instruction
*)this->instructions
.get_tail();
1313 new_inst
->saturate
= true;
1315 this->result
= get_temp(ir
->type
);
1316 st_dst_reg result_dst
= st_dst_reg(this->result
);
1317 result_dst
.writemask
= (1 << ir
->type
->vector_elements
) - 1;
1318 glsl_to_tgsi_instruction
*inst
;
1319 inst
= emit(ir
, TGSI_OPCODE_MOV
, result_dst
, src
);
1320 inst
->saturate
= true;
1327 glsl_to_tgsi_visitor::reladdr_to_temp(ir_instruction
*ir
,
1328 st_src_reg
*reg
, int *num_reladdr
)
1333 emit_arl(ir
, address_reg
, *reg
->reladdr
);
1335 if (*num_reladdr
!= 1) {
1336 st_src_reg temp
= get_temp(glsl_type::vec4_type
);
1338 emit(ir
, TGSI_OPCODE_MOV
, st_dst_reg(temp
), *reg
);
1346 glsl_to_tgsi_visitor::visit(ir_expression
*ir
)
1348 unsigned int operand
;
1349 st_src_reg op
[Elements(ir
->operands
)];
1350 st_src_reg result_src
;
1351 st_dst_reg result_dst
;
1353 /* Quick peephole: Emit MAD(a, b, c) instead of ADD(MUL(a, b), c)
1355 if (ir
->operation
== ir_binop_add
) {
1356 if (try_emit_mad(ir
, 1))
1358 if (try_emit_mad(ir
, 0))
1362 /* Quick peephole: Emit OPCODE_MAD(-a, -b, a) instead of AND(a, NOT(b))
1364 if (ir
->operation
== ir_binop_logic_and
) {
1365 if (try_emit_mad_for_and_not(ir
, 1))
1367 if (try_emit_mad_for_and_not(ir
, 0))
1371 if (try_emit_sat(ir
))
1374 if (ir
->operation
== ir_quadop_vector
)
1375 assert(!"ir_quadop_vector should have been lowered");
1377 for (operand
= 0; operand
< ir
->get_num_operands(); operand
++) {
1378 this->result
.file
= PROGRAM_UNDEFINED
;
1379 ir
->operands
[operand
]->accept(this);
1380 if (this->result
.file
== PROGRAM_UNDEFINED
) {
1382 printf("Failed to get tree for expression operand:\n");
1383 ir
->operands
[operand
]->accept(&v
);
1386 op
[operand
] = this->result
;
1388 /* Matrix expression operands should have been broken down to vector
1389 * operations already.
1391 assert(!ir
->operands
[operand
]->type
->is_matrix());
1394 int vector_elements
= ir
->operands
[0]->type
->vector_elements
;
1395 if (ir
->operands
[1]) {
1396 vector_elements
= MAX2(vector_elements
,
1397 ir
->operands
[1]->type
->vector_elements
);
1400 this->result
.file
= PROGRAM_UNDEFINED
;
1402 /* Storage for our result. Ideally for an assignment we'd be using
1403 * the actual storage for the result here, instead.
1405 result_src
= get_temp(ir
->type
);
1406 /* convenience for the emit functions below. */
1407 result_dst
= st_dst_reg(result_src
);
1408 /* Limit writes to the channels that will be used by result_src later.
1409 * This does limit this temp's use as a temporary for multi-instruction
1412 result_dst
.writemask
= (1 << ir
->type
->vector_elements
) - 1;
1414 switch (ir
->operation
) {
1415 case ir_unop_logic_not
:
1416 if (result_dst
.type
!= GLSL_TYPE_FLOAT
)
1417 emit(ir
, TGSI_OPCODE_NOT
, result_dst
, op
[0]);
1419 /* Previously 'SEQ dst, src, 0.0' was used for this. However, many
1420 * older GPUs implement SEQ using multiple instructions (i915 uses two
1421 * SGE instructions and a MUL instruction). Since our logic values are
1422 * 0.0 and 1.0, 1-x also implements !x.
1424 op
[0].negate
= ~op
[0].negate
;
1425 emit(ir
, TGSI_OPCODE_ADD
, result_dst
, op
[0], st_src_reg_for_float(1.0));
1429 if (result_dst
.type
== GLSL_TYPE_INT
|| result_dst
.type
== GLSL_TYPE_UINT
)
1430 emit(ir
, TGSI_OPCODE_INEG
, result_dst
, op
[0]);
1432 op
[0].negate
= ~op
[0].negate
;
1437 emit(ir
, TGSI_OPCODE_ABS
, result_dst
, op
[0]);
1440 emit(ir
, TGSI_OPCODE_SSG
, result_dst
, op
[0]);
1443 emit_scalar(ir
, TGSI_OPCODE_RCP
, result_dst
, op
[0]);
1447 emit_scalar(ir
, TGSI_OPCODE_EX2
, result_dst
, op
[0]);
1451 assert(!"not reached: should be handled by ir_explog_to_explog2");
1454 emit_scalar(ir
, TGSI_OPCODE_LG2
, result_dst
, op
[0]);
1457 emit_scalar(ir
, TGSI_OPCODE_SIN
, result_dst
, op
[0]);
1460 emit_scalar(ir
, TGSI_OPCODE_COS
, result_dst
, op
[0]);
1462 case ir_unop_sin_reduced
:
1463 emit_scs(ir
, TGSI_OPCODE_SIN
, result_dst
, op
[0]);
1465 case ir_unop_cos_reduced
:
1466 emit_scs(ir
, TGSI_OPCODE_COS
, result_dst
, op
[0]);
1470 emit(ir
, TGSI_OPCODE_DDX
, result_dst
, op
[0]);
1474 /* The X component contains 1 or -1 depending on whether the framebuffer
1475 * is a FBO or the window system buffer, respectively.
1476 * It is then multiplied with the source operand of DDY.
1478 static const gl_state_index transform_y_state
[STATE_LENGTH
]
1479 = { STATE_INTERNAL
, STATE_FB_WPOS_Y_TRANSFORM
};
1481 unsigned transform_y_index
=
1482 _mesa_add_state_reference(this->prog
->Parameters
,
1485 st_src_reg transform_y
= st_src_reg(PROGRAM_STATE_VAR
,
1487 glsl_type::vec4_type
);
1488 transform_y
.swizzle
= SWIZZLE_XXXX
;
1490 st_src_reg temp
= get_temp(glsl_type::vec4_type
);
1492 emit(ir
, TGSI_OPCODE_MUL
, st_dst_reg(temp
), transform_y
, op
[0]);
1493 emit(ir
, TGSI_OPCODE_DDY
, result_dst
, temp
);
1497 case ir_unop_noise
: {
1498 /* At some point, a motivated person could add a better
1499 * implementation of noise. Currently not even the nvidia
1500 * binary drivers do anything more than this. In any case, the
1501 * place to do this is in the GL state tracker, not the poor
1504 emit(ir
, TGSI_OPCODE_MOV
, result_dst
, st_src_reg_for_float(0.5));
1509 emit(ir
, TGSI_OPCODE_ADD
, result_dst
, op
[0], op
[1]);
1512 emit(ir
, TGSI_OPCODE_SUB
, result_dst
, op
[0], op
[1]);
1516 emit(ir
, TGSI_OPCODE_MUL
, result_dst
, op
[0], op
[1]);
1519 if (result_dst
.type
== GLSL_TYPE_FLOAT
)
1520 assert(!"not reached: should be handled by ir_div_to_mul_rcp");
1522 emit(ir
, TGSI_OPCODE_DIV
, result_dst
, op
[0], op
[1]);
1525 if (result_dst
.type
== GLSL_TYPE_FLOAT
)
1526 assert(!"ir_binop_mod should have been converted to b * fract(a/b)");
1528 emit(ir
, TGSI_OPCODE_MOD
, result_dst
, op
[0], op
[1]);
1532 emit(ir
, TGSI_OPCODE_SLT
, result_dst
, op
[0], op
[1]);
1534 case ir_binop_greater
:
1535 emit(ir
, TGSI_OPCODE_SLT
, result_dst
, op
[1], op
[0]);
1537 case ir_binop_lequal
:
1538 emit(ir
, TGSI_OPCODE_SGE
, result_dst
, op
[1], op
[0]);
1540 case ir_binop_gequal
:
1541 emit(ir
, TGSI_OPCODE_SGE
, result_dst
, op
[0], op
[1]);
1543 case ir_binop_equal
:
1544 emit(ir
, TGSI_OPCODE_SEQ
, result_dst
, op
[0], op
[1]);
1546 case ir_binop_nequal
:
1547 emit(ir
, TGSI_OPCODE_SNE
, result_dst
, op
[0], op
[1]);
1549 case ir_binop_all_equal
:
1550 /* "==" operator producing a scalar boolean. */
1551 if (ir
->operands
[0]->type
->is_vector() ||
1552 ir
->operands
[1]->type
->is_vector()) {
1553 st_src_reg temp
= get_temp(native_integers
?
1554 glsl_type::get_instance(ir
->operands
[0]->type
->base_type
, 4, 1) :
1555 glsl_type::vec4_type
);
1557 if (native_integers
) {
1558 st_dst_reg temp_dst
= st_dst_reg(temp
);
1559 st_src_reg temp1
= st_src_reg(temp
), temp2
= st_src_reg(temp
);
1561 emit(ir
, TGSI_OPCODE_SEQ
, st_dst_reg(temp
), op
[0], op
[1]);
1563 /* Emit 1-3 AND operations to combine the SEQ results. */
1564 switch (ir
->operands
[0]->type
->vector_elements
) {
1568 temp_dst
.writemask
= WRITEMASK_Y
;
1569 temp1
.swizzle
= SWIZZLE_YYYY
;
1570 temp2
.swizzle
= SWIZZLE_ZZZZ
;
1571 emit(ir
, TGSI_OPCODE_AND
, temp_dst
, temp1
, temp2
);
1574 temp_dst
.writemask
= WRITEMASK_X
;
1575 temp1
.swizzle
= SWIZZLE_XXXX
;
1576 temp2
.swizzle
= SWIZZLE_YYYY
;
1577 emit(ir
, TGSI_OPCODE_AND
, temp_dst
, temp1
, temp2
);
1578 temp_dst
.writemask
= WRITEMASK_Y
;
1579 temp1
.swizzle
= SWIZZLE_ZZZZ
;
1580 temp2
.swizzle
= SWIZZLE_WWWW
;
1581 emit(ir
, TGSI_OPCODE_AND
, temp_dst
, temp1
, temp2
);
1584 temp1
.swizzle
= SWIZZLE_XXXX
;
1585 temp2
.swizzle
= SWIZZLE_YYYY
;
1586 emit(ir
, TGSI_OPCODE_AND
, result_dst
, temp1
, temp2
);
1588 emit(ir
, TGSI_OPCODE_SNE
, st_dst_reg(temp
), op
[0], op
[1]);
1590 /* After the dot-product, the value will be an integer on the
1591 * range [0,4]. Zero becomes 1.0, and positive values become zero.
1593 emit_dp(ir
, result_dst
, temp
, temp
, vector_elements
);
1595 /* Negating the result of the dot-product gives values on the range
1596 * [-4, 0]. Zero becomes 1.0, and negative values become zero.
1597 * This is achieved using SGE.
1599 st_src_reg sge_src
= result_src
;
1600 sge_src
.negate
= ~sge_src
.negate
;
1601 emit(ir
, TGSI_OPCODE_SGE
, result_dst
, sge_src
, st_src_reg_for_float(0.0));
1604 emit(ir
, TGSI_OPCODE_SEQ
, result_dst
, op
[0], op
[1]);
1607 case ir_binop_any_nequal
:
1608 /* "!=" operator producing a scalar boolean. */
1609 if (ir
->operands
[0]->type
->is_vector() ||
1610 ir
->operands
[1]->type
->is_vector()) {
1611 st_src_reg temp
= get_temp(native_integers
?
1612 glsl_type::get_instance(ir
->operands
[0]->type
->base_type
, 4, 1) :
1613 glsl_type::vec4_type
);
1614 emit(ir
, TGSI_OPCODE_SNE
, st_dst_reg(temp
), op
[0], op
[1]);
1616 if (native_integers
) {
1617 st_dst_reg temp_dst
= st_dst_reg(temp
);
1618 st_src_reg temp1
= st_src_reg(temp
), temp2
= st_src_reg(temp
);
1620 /* Emit 1-3 OR operations to combine the SNE results. */
1621 switch (ir
->operands
[0]->type
->vector_elements
) {
1625 temp_dst
.writemask
= WRITEMASK_Y
;
1626 temp1
.swizzle
= SWIZZLE_YYYY
;
1627 temp2
.swizzle
= SWIZZLE_ZZZZ
;
1628 emit(ir
, TGSI_OPCODE_OR
, temp_dst
, temp1
, temp2
);
1631 temp_dst
.writemask
= WRITEMASK_X
;
1632 temp1
.swizzle
= SWIZZLE_XXXX
;
1633 temp2
.swizzle
= SWIZZLE_YYYY
;
1634 emit(ir
, TGSI_OPCODE_OR
, temp_dst
, temp1
, temp2
);
1635 temp_dst
.writemask
= WRITEMASK_Y
;
1636 temp1
.swizzle
= SWIZZLE_ZZZZ
;
1637 temp2
.swizzle
= SWIZZLE_WWWW
;
1638 emit(ir
, TGSI_OPCODE_OR
, temp_dst
, temp1
, temp2
);
1641 temp1
.swizzle
= SWIZZLE_XXXX
;
1642 temp2
.swizzle
= SWIZZLE_YYYY
;
1643 emit(ir
, TGSI_OPCODE_OR
, result_dst
, temp1
, temp2
);
1645 /* After the dot-product, the value will be an integer on the
1646 * range [0,4]. Zero stays zero, and positive values become 1.0.
1648 glsl_to_tgsi_instruction
*const dp
=
1649 emit_dp(ir
, result_dst
, temp
, temp
, vector_elements
);
1650 if (this->prog
->Target
== GL_FRAGMENT_PROGRAM_ARB
) {
1651 /* The clamping to [0,1] can be done for free in the fragment
1652 * shader with a saturate.
1654 dp
->saturate
= true;
1656 /* Negating the result of the dot-product gives values on the range
1657 * [-4, 0]. Zero stays zero, and negative values become 1.0. This
1658 * achieved using SLT.
1660 st_src_reg slt_src
= result_src
;
1661 slt_src
.negate
= ~slt_src
.negate
;
1662 emit(ir
, TGSI_OPCODE_SLT
, result_dst
, slt_src
, st_src_reg_for_float(0.0));
1666 emit(ir
, TGSI_OPCODE_SNE
, result_dst
, op
[0], op
[1]);
1671 assert(ir
->operands
[0]->type
->is_vector());
1673 /* After the dot-product, the value will be an integer on the
1674 * range [0,4]. Zero stays zero, and positive values become 1.0.
1676 glsl_to_tgsi_instruction
*const dp
=
1677 emit_dp(ir
, result_dst
, op
[0], op
[0],
1678 ir
->operands
[0]->type
->vector_elements
);
1679 if (this->prog
->Target
== GL_FRAGMENT_PROGRAM_ARB
&&
1680 result_dst
.type
== GLSL_TYPE_FLOAT
) {
1681 /* The clamping to [0,1] can be done for free in the fragment
1682 * shader with a saturate.
1684 dp
->saturate
= true;
1685 } else if (result_dst
.type
== GLSL_TYPE_FLOAT
) {
1686 /* Negating the result of the dot-product gives values on the range
1687 * [-4, 0]. Zero stays zero, and negative values become 1.0. This
1688 * is achieved using SLT.
1690 st_src_reg slt_src
= result_src
;
1691 slt_src
.negate
= ~slt_src
.negate
;
1692 emit(ir
, TGSI_OPCODE_SLT
, result_dst
, slt_src
, st_src_reg_for_float(0.0));
1695 /* Use SNE 0 if integers are being used as boolean values. */
1696 emit(ir
, TGSI_OPCODE_SNE
, result_dst
, result_src
, st_src_reg_for_int(0));
1701 case ir_binop_logic_xor
:
1702 if (native_integers
)
1703 emit(ir
, TGSI_OPCODE_XOR
, result_dst
, op
[0], op
[1]);
1705 emit(ir
, TGSI_OPCODE_SNE
, result_dst
, op
[0], op
[1]);
1708 case ir_binop_logic_or
: {
1709 if (native_integers
) {
1710 /* If integers are used as booleans, we can use an actual "or"
1713 assert(native_integers
);
1714 emit(ir
, TGSI_OPCODE_OR
, result_dst
, op
[0], op
[1]);
1716 /* After the addition, the value will be an integer on the
1717 * range [0,2]. Zero stays zero, and positive values become 1.0.
1719 glsl_to_tgsi_instruction
*add
=
1720 emit(ir
, TGSI_OPCODE_ADD
, result_dst
, op
[0], op
[1]);
1721 if (this->prog
->Target
== GL_FRAGMENT_PROGRAM_ARB
) {
1722 /* The clamping to [0,1] can be done for free in the fragment
1723 * shader with a saturate if floats are being used as boolean values.
1725 add
->saturate
= true;
1727 /* Negating the result of the addition gives values on the range
1728 * [-2, 0]. Zero stays zero, and negative values become 1.0. This
1729 * is achieved using SLT.
1731 st_src_reg slt_src
= result_src
;
1732 slt_src
.negate
= ~slt_src
.negate
;
1733 emit(ir
, TGSI_OPCODE_SLT
, result_dst
, slt_src
, st_src_reg_for_float(0.0));
1739 case ir_binop_logic_and
:
1740 /* If native integers are disabled, the bool args are stored as float 0.0
1741 * or 1.0, so "mul" gives us "and". If they're enabled, just use the
1742 * actual AND opcode.
1744 if (native_integers
)
1745 emit(ir
, TGSI_OPCODE_AND
, result_dst
, op
[0], op
[1]);
1747 emit(ir
, TGSI_OPCODE_MUL
, result_dst
, op
[0], op
[1]);
1751 assert(ir
->operands
[0]->type
->is_vector());
1752 assert(ir
->operands
[0]->type
== ir
->operands
[1]->type
);
1753 emit_dp(ir
, result_dst
, op
[0], op
[1],
1754 ir
->operands
[0]->type
->vector_elements
);
1758 /* sqrt(x) = x * rsq(x). */
1759 emit_scalar(ir
, TGSI_OPCODE_RSQ
, result_dst
, op
[0]);
1760 emit(ir
, TGSI_OPCODE_MUL
, result_dst
, result_src
, op
[0]);
1761 /* For incoming channels <= 0, set the result to 0. */
1762 op
[0].negate
= ~op
[0].negate
;
1763 emit(ir
, TGSI_OPCODE_CMP
, result_dst
,
1764 op
[0], result_src
, st_src_reg_for_float(0.0));
1767 emit_scalar(ir
, TGSI_OPCODE_RSQ
, result_dst
, op
[0]);
1770 if (native_integers
) {
1771 emit(ir
, TGSI_OPCODE_I2F
, result_dst
, op
[0]);
1774 /* fallthrough to next case otherwise */
1776 if (native_integers
) {
1777 emit(ir
, TGSI_OPCODE_AND
, result_dst
, op
[0], st_src_reg_for_float(1.0));
1780 /* fallthrough to next case otherwise */
1783 /* Converting between signed and unsigned integers is a no-op. */
1787 if (native_integers
) {
1788 /* Booleans are stored as integers using ~0 for true and 0 for false.
1789 * GLSL requires that int(bool) return 1 for true and 0 for false.
1790 * This conversion is done with AND, but it could be done with NEG.
1792 emit(ir
, TGSI_OPCODE_AND
, result_dst
, op
[0], st_src_reg_for_int(1));
1794 /* Booleans and integers are both stored as floats when native
1795 * integers are disabled.
1801 if (native_integers
)
1802 emit(ir
, TGSI_OPCODE_F2I
, result_dst
, op
[0]);
1804 emit(ir
, TGSI_OPCODE_TRUNC
, result_dst
, op
[0]);
1807 if (native_integers
)
1808 emit(ir
, TGSI_OPCODE_F2U
, result_dst
, op
[0]);
1810 emit(ir
, TGSI_OPCODE_TRUNC
, result_dst
, op
[0]);
1812 case ir_unop_bitcast_f2i
:
1813 case ir_unop_bitcast_f2u
:
1814 case ir_unop_bitcast_i2f
:
1815 case ir_unop_bitcast_u2f
:
1819 emit(ir
, TGSI_OPCODE_SNE
, result_dst
, op
[0], st_src_reg_for_float(0.0));
1822 if (native_integers
)
1823 emit(ir
, TGSI_OPCODE_INEG
, result_dst
, op
[0]);
1825 emit(ir
, TGSI_OPCODE_SNE
, result_dst
, op
[0], st_src_reg_for_float(0.0));
1828 emit(ir
, TGSI_OPCODE_TRUNC
, result_dst
, op
[0]);
1831 emit(ir
, TGSI_OPCODE_CEIL
, result_dst
, op
[0]);
1834 emit(ir
, TGSI_OPCODE_FLR
, result_dst
, op
[0]);
1836 case ir_unop_round_even
:
1837 emit(ir
, TGSI_OPCODE_ROUND
, result_dst
, op
[0]);
1840 emit(ir
, TGSI_OPCODE_FRC
, result_dst
, op
[0]);
1844 emit(ir
, TGSI_OPCODE_MIN
, result_dst
, op
[0], op
[1]);
1847 emit(ir
, TGSI_OPCODE_MAX
, result_dst
, op
[0], op
[1]);
1850 emit_scalar(ir
, TGSI_OPCODE_POW
, result_dst
, op
[0], op
[1]);
1853 case ir_unop_bit_not
:
1854 if (native_integers
) {
1855 emit(ir
, TGSI_OPCODE_NOT
, result_dst
, op
[0]);
1859 if (native_integers
) {
1860 emit(ir
, TGSI_OPCODE_U2F
, result_dst
, op
[0]);
1863 case ir_binop_lshift
:
1864 if (native_integers
) {
1865 emit(ir
, TGSI_OPCODE_SHL
, result_dst
, op
[0], op
[1]);
1868 case ir_binop_rshift
:
1869 if (native_integers
) {
1870 emit(ir
, TGSI_OPCODE_ISHR
, result_dst
, op
[0], op
[1]);
1873 case ir_binop_bit_and
:
1874 if (native_integers
) {
1875 emit(ir
, TGSI_OPCODE_AND
, result_dst
, op
[0], op
[1]);
1878 case ir_binop_bit_xor
:
1879 if (native_integers
) {
1880 emit(ir
, TGSI_OPCODE_XOR
, result_dst
, op
[0], op
[1]);
1883 case ir_binop_bit_or
:
1884 if (native_integers
) {
1885 emit(ir
, TGSI_OPCODE_OR
, result_dst
, op
[0], op
[1]);
1889 assert(!"GLSL 1.30 features unsupported");
1892 case ir_binop_ubo_load
: {
1893 ir_constant
*uniform_block
= ir
->operands
[0]->as_constant();
1894 st_src_reg index_reg
= get_temp(glsl_type::uint_type
);
1897 cbuf
.type
= glsl_type::vec4_type
->base_type
;
1898 cbuf
.file
= PROGRAM_CONSTANT
;
1900 cbuf
.index2D
= uniform_block
->value
.u
[0] + 1;
1901 cbuf
.reladdr
= NULL
;
1904 assert(ir
->type
->is_vector() || ir
->type
->is_scalar());
1906 emit(ir
, TGSI_OPCODE_USHR
, st_dst_reg(index_reg
), op
[1], st_src_reg_for_int(4));
1908 cbuf
.swizzle
= swizzle_for_size(ir
->type
->vector_elements
);
1909 cbuf
.reladdr
= ralloc(mem_ctx
, st_src_reg
);
1910 memcpy(cbuf
.reladdr
, &index_reg
, sizeof(index_reg
));
1912 emit(ir
, TGSI_OPCODE_MOV
, result_dst
, cbuf
);
1915 case ir_quadop_vector
:
1916 /* This operation should have already been handled.
1918 assert(!"Should not get here.");
1922 this->result
= result_src
;
1927 glsl_to_tgsi_visitor::visit(ir_swizzle
*ir
)
1933 /* Note that this is only swizzles in expressions, not those on the left
1934 * hand side of an assignment, which do write masking. See ir_assignment
1938 ir
->val
->accept(this);
1940 assert(src
.file
!= PROGRAM_UNDEFINED
);
1942 for (i
= 0; i
< 4; i
++) {
1943 if (i
< ir
->type
->vector_elements
) {
1946 swizzle
[i
] = GET_SWZ(src
.swizzle
, ir
->mask
.x
);
1949 swizzle
[i
] = GET_SWZ(src
.swizzle
, ir
->mask
.y
);
1952 swizzle
[i
] = GET_SWZ(src
.swizzle
, ir
->mask
.z
);
1955 swizzle
[i
] = GET_SWZ(src
.swizzle
, ir
->mask
.w
);
1959 /* If the type is smaller than a vec4, replicate the last
1962 swizzle
[i
] = swizzle
[ir
->type
->vector_elements
- 1];
1966 src
.swizzle
= MAKE_SWIZZLE4(swizzle
[0], swizzle
[1], swizzle
[2], swizzle
[3]);
1972 glsl_to_tgsi_visitor::visit(ir_dereference_variable
*ir
)
1974 variable_storage
*entry
= find_variable_storage(ir
->var
);
1975 ir_variable
*var
= ir
->var
;
1978 switch (var
->mode
) {
1979 case ir_var_uniform
:
1980 entry
= new(mem_ctx
) variable_storage(var
, PROGRAM_UNIFORM
,
1982 this->variables
.push_tail(entry
);
1986 /* The linker assigns locations for varyings and attributes,
1987 * including deprecated builtins (like gl_Color), user-assign
1988 * generic attributes (glBindVertexLocation), and
1989 * user-defined varyings.
1991 * FINISHME: We would hit this path for function arguments. Fix!
1993 assert(var
->location
!= -1);
1994 entry
= new(mem_ctx
) variable_storage(var
,
1999 assert(var
->location
!= -1);
2000 entry
= new(mem_ctx
) variable_storage(var
,
2002 var
->location
+ var
->index
);
2004 case ir_var_system_value
:
2005 entry
= new(mem_ctx
) variable_storage(var
,
2006 PROGRAM_SYSTEM_VALUE
,
2010 case ir_var_temporary
:
2011 entry
= new(mem_ctx
) variable_storage(var
, PROGRAM_TEMPORARY
,
2013 this->variables
.push_tail(entry
);
2015 next_temp
+= type_size(var
->type
);
2020 printf("Failed to make storage for %s\n", var
->name
);
2025 this->result
= st_src_reg(entry
->file
, entry
->index
, var
->type
);
2026 if (!native_integers
)
2027 this->result
.type
= GLSL_TYPE_FLOAT
;
2031 glsl_to_tgsi_visitor::visit(ir_dereference_array
*ir
)
2035 int element_size
= type_size(ir
->type
);
2037 index
= ir
->array_index
->constant_expression_value();
2039 ir
->array
->accept(this);
2043 src
.index
+= index
->value
.i
[0] * element_size
;
2045 /* Variable index array dereference. It eats the "vec4" of the
2046 * base of the array and an index that offsets the TGSI register
2049 ir
->array_index
->accept(this);
2051 st_src_reg index_reg
;
2053 if (element_size
== 1) {
2054 index_reg
= this->result
;
2056 index_reg
= get_temp(native_integers
?
2057 glsl_type::int_type
: glsl_type::float_type
);
2059 emit(ir
, TGSI_OPCODE_MUL
, st_dst_reg(index_reg
),
2060 this->result
, st_src_reg_for_type(index_reg
.type
, element_size
));
2063 /* If there was already a relative address register involved, add the
2064 * new and the old together to get the new offset.
2066 if (src
.reladdr
!= NULL
) {
2067 st_src_reg accum_reg
= get_temp(native_integers
?
2068 glsl_type::int_type
: glsl_type::float_type
);
2070 emit(ir
, TGSI_OPCODE_ADD
, st_dst_reg(accum_reg
),
2071 index_reg
, *src
.reladdr
);
2073 index_reg
= accum_reg
;
2076 src
.reladdr
= ralloc(mem_ctx
, st_src_reg
);
2077 memcpy(src
.reladdr
, &index_reg
, sizeof(index_reg
));
2080 /* If the type is smaller than a vec4, replicate the last channel out. */
2081 if (ir
->type
->is_scalar() || ir
->type
->is_vector())
2082 src
.swizzle
= swizzle_for_size(ir
->type
->vector_elements
);
2084 src
.swizzle
= SWIZZLE_NOOP
;
2086 /* Change the register type to the element type of the array. */
2087 src
.type
= ir
->type
->base_type
;
2093 glsl_to_tgsi_visitor::visit(ir_dereference_record
*ir
)
2096 const glsl_type
*struct_type
= ir
->record
->type
;
2099 ir
->record
->accept(this);
2101 for (i
= 0; i
< struct_type
->length
; i
++) {
2102 if (strcmp(struct_type
->fields
.structure
[i
].name
, ir
->field
) == 0)
2104 offset
+= type_size(struct_type
->fields
.structure
[i
].type
);
2107 /* If the type is smaller than a vec4, replicate the last channel out. */
2108 if (ir
->type
->is_scalar() || ir
->type
->is_vector())
2109 this->result
.swizzle
= swizzle_for_size(ir
->type
->vector_elements
);
2111 this->result
.swizzle
= SWIZZLE_NOOP
;
2113 this->result
.index
+= offset
;
2114 this->result
.type
= ir
->type
->base_type
;
2118 * We want to be careful in assignment setup to hit the actual storage
2119 * instead of potentially using a temporary like we might with the
2120 * ir_dereference handler.
2123 get_assignment_lhs(ir_dereference
*ir
, glsl_to_tgsi_visitor
*v
)
2125 /* The LHS must be a dereference. If the LHS is a variable indexed array
2126 * access of a vector, it must be separated into a series conditional moves
2127 * before reaching this point (see ir_vec_index_to_cond_assign).
2129 assert(ir
->as_dereference());
2130 ir_dereference_array
*deref_array
= ir
->as_dereference_array();
2132 assert(!deref_array
->array
->type
->is_vector());
2135 /* Use the rvalue deref handler for the most part. We'll ignore
2136 * swizzles in it and write swizzles using writemask, though.
2139 return st_dst_reg(v
->result
);
2143 * Process the condition of a conditional assignment
2145 * Examines the condition of a conditional assignment to generate the optimal
2146 * first operand of a \c CMP instruction. If the condition is a relational
2147 * operator with 0 (e.g., \c ir_binop_less), the value being compared will be
2148 * used as the source for the \c CMP instruction. Otherwise the comparison
2149 * is processed to a boolean result, and the boolean result is used as the
2150 * operand to the CMP instruction.
2153 glsl_to_tgsi_visitor::process_move_condition(ir_rvalue
*ir
)
2155 ir_rvalue
*src_ir
= ir
;
2157 bool switch_order
= false;
2159 ir_expression
*const expr
= ir
->as_expression();
2160 if ((expr
!= NULL
) && (expr
->get_num_operands() == 2)) {
2161 bool zero_on_left
= false;
2163 if (expr
->operands
[0]->is_zero()) {
2164 src_ir
= expr
->operands
[1];
2165 zero_on_left
= true;
2166 } else if (expr
->operands
[1]->is_zero()) {
2167 src_ir
= expr
->operands
[0];
2168 zero_on_left
= false;
2172 * (a < 0) T F F ( a < 0) T F F
2173 * (0 < a) F F T (-a < 0) F F T
2174 * (a <= 0) T T F (-a < 0) F F T (swap order of other operands)
2175 * (0 <= a) F T T ( a < 0) T F F (swap order of other operands)
2176 * (a > 0) F F T (-a < 0) F F T
2177 * (0 > a) T F F ( a < 0) T F F
2178 * (a >= 0) F T T ( a < 0) T F F (swap order of other operands)
2179 * (0 >= a) T T F (-a < 0) F F T (swap order of other operands)
2181 * Note that exchanging the order of 0 and 'a' in the comparison simply
2182 * means that the value of 'a' should be negated.
2185 switch (expr
->operation
) {
2187 switch_order
= false;
2188 negate
= zero_on_left
;
2191 case ir_binop_greater
:
2192 switch_order
= false;
2193 negate
= !zero_on_left
;
2196 case ir_binop_lequal
:
2197 switch_order
= true;
2198 negate
= !zero_on_left
;
2201 case ir_binop_gequal
:
2202 switch_order
= true;
2203 negate
= zero_on_left
;
2207 /* This isn't the right kind of comparison afterall, so make sure
2208 * the whole condition is visited.
2216 src_ir
->accept(this);
2218 /* We use the TGSI_OPCODE_CMP (a < 0 ? b : c) for conditional moves, and the
2219 * condition we produced is 0.0 or 1.0. By flipping the sign, we can
2220 * choose which value TGSI_OPCODE_CMP produces without an extra instruction
2221 * computing the condition.
2224 this->result
.negate
= ~this->result
.negate
;
2226 return switch_order
;
2230 glsl_to_tgsi_visitor::visit(ir_assignment
*ir
)
2236 ir
->rhs
->accept(this);
2239 l
= get_assignment_lhs(ir
->lhs
, this);
2241 /* FINISHME: This should really set to the correct maximal writemask for each
2242 * FINISHME: component written (in the loops below). This case can only
2243 * FINISHME: occur for matrices, arrays, and structures.
2245 if (ir
->write_mask
== 0) {
2246 assert(!ir
->lhs
->type
->is_scalar() && !ir
->lhs
->type
->is_vector());
2247 l
.writemask
= WRITEMASK_XYZW
;
2248 } else if (ir
->lhs
->type
->is_scalar() &&
2249 ir
->lhs
->variable_referenced()->mode
== ir_var_out
) {
2250 /* FINISHME: This hack makes writing to gl_FragDepth, which lives in the
2251 * FINISHME: W component of fragment shader output zero, work correctly.
2253 l
.writemask
= WRITEMASK_XYZW
;
2256 int first_enabled_chan
= 0;
2259 l
.writemask
= ir
->write_mask
;
2261 for (int i
= 0; i
< 4; i
++) {
2262 if (l
.writemask
& (1 << i
)) {
2263 first_enabled_chan
= GET_SWZ(r
.swizzle
, i
);
2268 /* Swizzle a small RHS vector into the channels being written.
2270 * glsl ir treats write_mask as dictating how many channels are
2271 * present on the RHS while TGSI treats write_mask as just
2272 * showing which channels of the vec4 RHS get written.
2274 for (int i
= 0; i
< 4; i
++) {
2275 if (l
.writemask
& (1 << i
))
2276 swizzles
[i
] = GET_SWZ(r
.swizzle
, rhs_chan
++);
2278 swizzles
[i
] = first_enabled_chan
;
2280 r
.swizzle
= MAKE_SWIZZLE4(swizzles
[0], swizzles
[1],
2281 swizzles
[2], swizzles
[3]);
2284 assert(l
.file
!= PROGRAM_UNDEFINED
);
2285 assert(r
.file
!= PROGRAM_UNDEFINED
);
2287 if (ir
->condition
) {
2288 const bool switch_order
= this->process_move_condition(ir
->condition
);
2289 st_src_reg condition
= this->result
;
2291 for (i
= 0; i
< type_size(ir
->lhs
->type
); i
++) {
2292 st_src_reg l_src
= st_src_reg(l
);
2293 st_src_reg condition_temp
= condition
;
2294 l_src
.swizzle
= swizzle_for_size(ir
->lhs
->type
->vector_elements
);
2296 if (native_integers
) {
2297 /* This is necessary because TGSI's CMP instruction expects the
2298 * condition to be a float, and we store booleans as integers.
2299 * If TGSI had a UCMP instruction or similar, this extra
2300 * instruction would not be necessary.
2302 condition_temp
= get_temp(glsl_type::vec4_type
);
2303 condition
.negate
= 0;
2304 emit(ir
, TGSI_OPCODE_I2F
, st_dst_reg(condition_temp
), condition
);
2305 condition_temp
.swizzle
= condition
.swizzle
;
2309 emit(ir
, TGSI_OPCODE_CMP
, l
, condition_temp
, l_src
, r
);
2311 emit(ir
, TGSI_OPCODE_CMP
, l
, condition_temp
, r
, l_src
);
2317 } else if (ir
->rhs
->as_expression() &&
2318 this->instructions
.get_tail() &&
2319 ir
->rhs
== ((glsl_to_tgsi_instruction
*)this->instructions
.get_tail())->ir
&&
2320 type_size(ir
->lhs
->type
) == 1 &&
2321 l
.writemask
== ((glsl_to_tgsi_instruction
*)this->instructions
.get_tail())->dst
.writemask
) {
2322 /* To avoid emitting an extra MOV when assigning an expression to a
2323 * variable, emit the last instruction of the expression again, but
2324 * replace the destination register with the target of the assignment.
2325 * Dead code elimination will remove the original instruction.
2327 glsl_to_tgsi_instruction
*inst
, *new_inst
;
2328 inst
= (glsl_to_tgsi_instruction
*)this->instructions
.get_tail();
2329 new_inst
= emit(ir
, inst
->op
, l
, inst
->src
[0], inst
->src
[1], inst
->src
[2]);
2330 new_inst
->saturate
= inst
->saturate
;
2331 inst
->dead_mask
= inst
->dst
.writemask
;
2333 for (i
= 0; i
< type_size(ir
->lhs
->type
); i
++) {
2334 if (ir
->rhs
->type
->is_array())
2335 r
.type
= ir
->rhs
->type
->element_type()->base_type
;
2336 else if (ir
->rhs
->type
->is_record())
2337 r
.type
= ir
->rhs
->type
->fields
.structure
[i
].type
->base_type
;
2338 emit(ir
, TGSI_OPCODE_MOV
, l
, r
);
2347 glsl_to_tgsi_visitor::visit(ir_constant
*ir
)
2350 GLfloat stack_vals
[4] = { 0 };
2351 gl_constant_value
*values
= (gl_constant_value
*) stack_vals
;
2352 GLenum gl_type
= GL_NONE
;
2354 static int in_array
= 0;
2355 gl_register_file file
= in_array
? PROGRAM_CONSTANT
: PROGRAM_IMMEDIATE
;
2357 /* Unfortunately, 4 floats is all we can get into
2358 * _mesa_add_typed_unnamed_constant. So, make a temp to store an
2359 * aggregate constant and move each constant value into it. If we
2360 * get lucky, copy propagation will eliminate the extra moves.
2362 if (ir
->type
->base_type
== GLSL_TYPE_STRUCT
) {
2363 st_src_reg temp_base
= get_temp(ir
->type
);
2364 st_dst_reg temp
= st_dst_reg(temp_base
);
2366 foreach_iter(exec_list_iterator
, iter
, ir
->components
) {
2367 ir_constant
*field_value
= (ir_constant
*)iter
.get();
2368 int size
= type_size(field_value
->type
);
2372 field_value
->accept(this);
2375 for (i
= 0; i
< (unsigned int)size
; i
++) {
2376 emit(ir
, TGSI_OPCODE_MOV
, temp
, src
);
2382 this->result
= temp_base
;
2386 if (ir
->type
->is_array()) {
2387 st_src_reg temp_base
= get_temp(ir
->type
);
2388 st_dst_reg temp
= st_dst_reg(temp_base
);
2389 int size
= type_size(ir
->type
->fields
.array
);
2394 for (i
= 0; i
< ir
->type
->length
; i
++) {
2395 ir
->array_elements
[i
]->accept(this);
2397 for (int j
= 0; j
< size
; j
++) {
2398 emit(ir
, TGSI_OPCODE_MOV
, temp
, src
);
2404 this->result
= temp_base
;
2409 if (ir
->type
->is_matrix()) {
2410 st_src_reg mat
= get_temp(ir
->type
);
2411 st_dst_reg mat_column
= st_dst_reg(mat
);
2413 for (i
= 0; i
< ir
->type
->matrix_columns
; i
++) {
2414 assert(ir
->type
->base_type
== GLSL_TYPE_FLOAT
);
2415 values
= (gl_constant_value
*) &ir
->value
.f
[i
* ir
->type
->vector_elements
];
2417 src
= st_src_reg(file
, -1, ir
->type
->base_type
);
2418 src
.index
= add_constant(file
,
2420 ir
->type
->vector_elements
,
2423 emit(ir
, TGSI_OPCODE_MOV
, mat_column
, src
);
2432 switch (ir
->type
->base_type
) {
2433 case GLSL_TYPE_FLOAT
:
2435 for (i
= 0; i
< ir
->type
->vector_elements
; i
++) {
2436 values
[i
].f
= ir
->value
.f
[i
];
2439 case GLSL_TYPE_UINT
:
2440 gl_type
= native_integers
? GL_UNSIGNED_INT
: GL_FLOAT
;
2441 for (i
= 0; i
< ir
->type
->vector_elements
; i
++) {
2442 if (native_integers
)
2443 values
[i
].u
= ir
->value
.u
[i
];
2445 values
[i
].f
= ir
->value
.u
[i
];
2449 gl_type
= native_integers
? GL_INT
: GL_FLOAT
;
2450 for (i
= 0; i
< ir
->type
->vector_elements
; i
++) {
2451 if (native_integers
)
2452 values
[i
].i
= ir
->value
.i
[i
];
2454 values
[i
].f
= ir
->value
.i
[i
];
2457 case GLSL_TYPE_BOOL
:
2458 gl_type
= native_integers
? GL_BOOL
: GL_FLOAT
;
2459 for (i
= 0; i
< ir
->type
->vector_elements
; i
++) {
2460 if (native_integers
)
2461 values
[i
].u
= ir
->value
.b
[i
] ? ~0 : 0;
2463 values
[i
].f
= ir
->value
.b
[i
];
2467 assert(!"Non-float/uint/int/bool constant");
2470 this->result
= st_src_reg(file
, -1, ir
->type
);
2471 this->result
.index
= add_constant(file
,
2473 ir
->type
->vector_elements
,
2475 &this->result
.swizzle
);
2479 glsl_to_tgsi_visitor::get_function_signature(ir_function_signature
*sig
)
2481 function_entry
*entry
;
2483 foreach_iter(exec_list_iterator
, iter
, this->function_signatures
) {
2484 entry
= (function_entry
*)iter
.get();
2486 if (entry
->sig
== sig
)
2490 entry
= ralloc(mem_ctx
, function_entry
);
2492 entry
->sig_id
= this->next_signature_id
++;
2493 entry
->bgn_inst
= NULL
;
2495 /* Allocate storage for all the parameters. */
2496 foreach_iter(exec_list_iterator
, iter
, sig
->parameters
) {
2497 ir_variable
*param
= (ir_variable
*)iter
.get();
2498 variable_storage
*storage
;
2500 storage
= find_variable_storage(param
);
2503 storage
= new(mem_ctx
) variable_storage(param
, PROGRAM_TEMPORARY
,
2505 this->variables
.push_tail(storage
);
2507 this->next_temp
+= type_size(param
->type
);
2510 if (!sig
->return_type
->is_void()) {
2511 entry
->return_reg
= get_temp(sig
->return_type
);
2513 entry
->return_reg
= undef_src
;
2516 this->function_signatures
.push_tail(entry
);
2521 glsl_to_tgsi_visitor::visit(ir_call
*ir
)
2523 glsl_to_tgsi_instruction
*call_inst
;
2524 ir_function_signature
*sig
= ir
->callee
;
2525 function_entry
*entry
= get_function_signature(sig
);
2528 /* Process in parameters. */
2529 exec_list_iterator sig_iter
= sig
->parameters
.iterator();
2530 foreach_iter(exec_list_iterator
, iter
, *ir
) {
2531 ir_rvalue
*param_rval
= (ir_rvalue
*)iter
.get();
2532 ir_variable
*param
= (ir_variable
*)sig_iter
.get();
2534 if (param
->mode
== ir_var_in
||
2535 param
->mode
== ir_var_inout
) {
2536 variable_storage
*storage
= find_variable_storage(param
);
2539 param_rval
->accept(this);
2540 st_src_reg r
= this->result
;
2543 l
.file
= storage
->file
;
2544 l
.index
= storage
->index
;
2546 l
.writemask
= WRITEMASK_XYZW
;
2547 l
.cond_mask
= COND_TR
;
2549 for (i
= 0; i
< type_size(param
->type
); i
++) {
2550 emit(ir
, TGSI_OPCODE_MOV
, l
, r
);
2558 assert(!sig_iter
.has_next());
2560 /* Emit call instruction */
2561 call_inst
= emit(ir
, TGSI_OPCODE_CAL
);
2562 call_inst
->function
= entry
;
2564 /* Process out parameters. */
2565 sig_iter
= sig
->parameters
.iterator();
2566 foreach_iter(exec_list_iterator
, iter
, *ir
) {
2567 ir_rvalue
*param_rval
= (ir_rvalue
*)iter
.get();
2568 ir_variable
*param
= (ir_variable
*)sig_iter
.get();
2570 if (param
->mode
== ir_var_out
||
2571 param
->mode
== ir_var_inout
) {
2572 variable_storage
*storage
= find_variable_storage(param
);
2576 r
.file
= storage
->file
;
2577 r
.index
= storage
->index
;
2579 r
.swizzle
= SWIZZLE_NOOP
;
2582 param_rval
->accept(this);
2583 st_dst_reg l
= st_dst_reg(this->result
);
2585 for (i
= 0; i
< type_size(param
->type
); i
++) {
2586 emit(ir
, TGSI_OPCODE_MOV
, l
, r
);
2594 assert(!sig_iter
.has_next());
2596 /* Process return value. */
2597 this->result
= entry
->return_reg
;
2601 glsl_to_tgsi_visitor::visit(ir_texture
*ir
)
2603 st_src_reg result_src
, coord
, cube_sc
, lod_info
, projector
, dx
, dy
, offset
;
2604 st_dst_reg result_dst
, coord_dst
, cube_sc_dst
;
2605 glsl_to_tgsi_instruction
*inst
= NULL
;
2606 unsigned opcode
= TGSI_OPCODE_NOP
;
2607 const glsl_type
*sampler_type
= ir
->sampler
->type
;
2608 bool is_cube_array
= false;
2610 /* if we are a cube array sampler */
2611 if ((sampler_type
->sampler_dimensionality
== GLSL_SAMPLER_DIM_CUBE
&&
2612 sampler_type
->sampler_array
)) {
2613 is_cube_array
= true;
2616 if (ir
->coordinate
) {
2617 ir
->coordinate
->accept(this);
2619 /* Put our coords in a temp. We'll need to modify them for shadow,
2620 * projection, or LOD, so the only case we'd use it as is is if
2621 * we're doing plain old texturing. The optimization passes on
2622 * glsl_to_tgsi_visitor should handle cleaning up our mess in that case.
2624 coord
= get_temp(glsl_type::vec4_type
);
2625 coord_dst
= st_dst_reg(coord
);
2626 emit(ir
, TGSI_OPCODE_MOV
, coord_dst
, this->result
);
2629 if (ir
->projector
) {
2630 ir
->projector
->accept(this);
2631 projector
= this->result
;
2634 /* Storage for our result. Ideally for an assignment we'd be using
2635 * the actual storage for the result here, instead.
2637 result_src
= get_temp(ir
->type
);
2638 result_dst
= st_dst_reg(result_src
);
2642 opcode
= (is_cube_array
&& ir
->shadow_comparitor
) ? TGSI_OPCODE_TEX2
: TGSI_OPCODE_TEX
;
2645 opcode
= is_cube_array
? TGSI_OPCODE_TXB2
: TGSI_OPCODE_TXB
;
2646 ir
->lod_info
.bias
->accept(this);
2647 lod_info
= this->result
;
2650 opcode
= is_cube_array
? TGSI_OPCODE_TXL2
: TGSI_OPCODE_TXL
;
2651 ir
->lod_info
.lod
->accept(this);
2652 lod_info
= this->result
;
2655 opcode
= TGSI_OPCODE_TXD
;
2656 ir
->lod_info
.grad
.dPdx
->accept(this);
2658 ir
->lod_info
.grad
.dPdy
->accept(this);
2662 opcode
= TGSI_OPCODE_TXQ
;
2663 ir
->lod_info
.lod
->accept(this);
2664 lod_info
= this->result
;
2667 opcode
= TGSI_OPCODE_TXF
;
2668 ir
->lod_info
.lod
->accept(this);
2669 lod_info
= this->result
;
2671 ir
->offset
->accept(this);
2672 offset
= this->result
;
2677 if (ir
->projector
) {
2678 if (opcode
== TGSI_OPCODE_TEX
) {
2679 /* Slot the projector in as the last component of the coord. */
2680 coord_dst
.writemask
= WRITEMASK_W
;
2681 emit(ir
, TGSI_OPCODE_MOV
, coord_dst
, projector
);
2682 coord_dst
.writemask
= WRITEMASK_XYZW
;
2683 opcode
= TGSI_OPCODE_TXP
;
2685 st_src_reg coord_w
= coord
;
2686 coord_w
.swizzle
= SWIZZLE_WWWW
;
2688 /* For the other TEX opcodes there's no projective version
2689 * since the last slot is taken up by LOD info. Do the
2690 * projective divide now.
2692 coord_dst
.writemask
= WRITEMASK_W
;
2693 emit(ir
, TGSI_OPCODE_RCP
, coord_dst
, projector
);
2695 /* In the case where we have to project the coordinates "by hand,"
2696 * the shadow comparator value must also be projected.
2698 st_src_reg tmp_src
= coord
;
2699 if (ir
->shadow_comparitor
) {
2700 /* Slot the shadow value in as the second to last component of the
2703 ir
->shadow_comparitor
->accept(this);
2705 tmp_src
= get_temp(glsl_type::vec4_type
);
2706 st_dst_reg tmp_dst
= st_dst_reg(tmp_src
);
2708 /* Projective division not allowed for array samplers. */
2709 assert(!sampler_type
->sampler_array
);
2711 tmp_dst
.writemask
= WRITEMASK_Z
;
2712 emit(ir
, TGSI_OPCODE_MOV
, tmp_dst
, this->result
);
2714 tmp_dst
.writemask
= WRITEMASK_XY
;
2715 emit(ir
, TGSI_OPCODE_MOV
, tmp_dst
, coord
);
2718 coord_dst
.writemask
= WRITEMASK_XYZ
;
2719 emit(ir
, TGSI_OPCODE_MUL
, coord_dst
, tmp_src
, coord_w
);
2721 coord_dst
.writemask
= WRITEMASK_XYZW
;
2722 coord
.swizzle
= SWIZZLE_XYZW
;
2726 /* If projection is done and the opcode is not TGSI_OPCODE_TXP, then the shadow
2727 * comparator was put in the correct place (and projected) by the code,
2728 * above, that handles by-hand projection.
2730 if (ir
->shadow_comparitor
&& (!ir
->projector
|| opcode
== TGSI_OPCODE_TXP
)) {
2731 /* Slot the shadow value in as the second to last component of the
2734 ir
->shadow_comparitor
->accept(this);
2736 if (is_cube_array
) {
2737 cube_sc
= get_temp(glsl_type::float_type
);
2738 cube_sc_dst
= st_dst_reg(cube_sc
);
2739 cube_sc_dst
.writemask
= WRITEMASK_X
;
2740 emit(ir
, TGSI_OPCODE_MOV
, cube_sc_dst
, this->result
);
2741 cube_sc_dst
.writemask
= WRITEMASK_X
;
2744 if ((sampler_type
->sampler_dimensionality
== GLSL_SAMPLER_DIM_2D
&&
2745 sampler_type
->sampler_array
) ||
2746 sampler_type
->sampler_dimensionality
== GLSL_SAMPLER_DIM_CUBE
) {
2747 coord_dst
.writemask
= WRITEMASK_W
;
2749 coord_dst
.writemask
= WRITEMASK_Z
;
2752 emit(ir
, TGSI_OPCODE_MOV
, coord_dst
, this->result
);
2753 coord_dst
.writemask
= WRITEMASK_XYZW
;
2757 if (opcode
== TGSI_OPCODE_TXL
|| opcode
== TGSI_OPCODE_TXB
||
2758 opcode
== TGSI_OPCODE_TXF
) {
2759 /* TGSI stores LOD or LOD bias in the last channel of the coords. */
2760 coord_dst
.writemask
= WRITEMASK_W
;
2761 emit(ir
, TGSI_OPCODE_MOV
, coord_dst
, lod_info
);
2762 coord_dst
.writemask
= WRITEMASK_XYZW
;
2765 if (opcode
== TGSI_OPCODE_TXD
)
2766 inst
= emit(ir
, opcode
, result_dst
, coord
, dx
, dy
);
2767 else if (opcode
== TGSI_OPCODE_TXQ
)
2768 inst
= emit(ir
, opcode
, result_dst
, lod_info
);
2769 else if (opcode
== TGSI_OPCODE_TXF
) {
2770 inst
= emit(ir
, opcode
, result_dst
, coord
);
2771 } else if (opcode
== TGSI_OPCODE_TXL2
|| opcode
== TGSI_OPCODE_TXB2
) {
2772 inst
= emit(ir
, opcode
, result_dst
, coord
, lod_info
);
2773 } else if (opcode
== TGSI_OPCODE_TEX2
) {
2774 inst
= emit(ir
, opcode
, result_dst
, coord
, cube_sc
);
2776 inst
= emit(ir
, opcode
, result_dst
, coord
);
2778 if (ir
->shadow_comparitor
)
2779 inst
->tex_shadow
= GL_TRUE
;
2781 inst
->sampler
= _mesa_get_sampler_uniform_value(ir
->sampler
,
2782 this->shader_program
,
2786 inst
->tex_offset_num_offset
= 1;
2787 inst
->tex_offsets
[0].Index
= offset
.index
;
2788 inst
->tex_offsets
[0].File
= offset
.file
;
2789 inst
->tex_offsets
[0].SwizzleX
= GET_SWZ(offset
.swizzle
, 0);
2790 inst
->tex_offsets
[0].SwizzleY
= GET_SWZ(offset
.swizzle
, 1);
2791 inst
->tex_offsets
[0].SwizzleZ
= GET_SWZ(offset
.swizzle
, 2);
2794 switch (sampler_type
->sampler_dimensionality
) {
2795 case GLSL_SAMPLER_DIM_1D
:
2796 inst
->tex_target
= (sampler_type
->sampler_array
)
2797 ? TEXTURE_1D_ARRAY_INDEX
: TEXTURE_1D_INDEX
;
2799 case GLSL_SAMPLER_DIM_2D
:
2800 inst
->tex_target
= (sampler_type
->sampler_array
)
2801 ? TEXTURE_2D_ARRAY_INDEX
: TEXTURE_2D_INDEX
;
2803 case GLSL_SAMPLER_DIM_3D
:
2804 inst
->tex_target
= TEXTURE_3D_INDEX
;
2806 case GLSL_SAMPLER_DIM_CUBE
:
2807 inst
->tex_target
= (sampler_type
->sampler_array
)
2808 ? TEXTURE_CUBE_ARRAY_INDEX
: TEXTURE_CUBE_INDEX
;
2810 case GLSL_SAMPLER_DIM_RECT
:
2811 inst
->tex_target
= TEXTURE_RECT_INDEX
;
2813 case GLSL_SAMPLER_DIM_BUF
:
2814 inst
->tex_target
= TEXTURE_BUFFER_INDEX
;
2816 case GLSL_SAMPLER_DIM_EXTERNAL
:
2817 inst
->tex_target
= TEXTURE_EXTERNAL_INDEX
;
2820 assert(!"Should not get here.");
2823 this->result
= result_src
;
2827 glsl_to_tgsi_visitor::visit(ir_return
*ir
)
2829 if (ir
->get_value()) {
2833 assert(current_function
);
2835 ir
->get_value()->accept(this);
2836 st_src_reg r
= this->result
;
2838 l
= st_dst_reg(current_function
->return_reg
);
2840 for (i
= 0; i
< type_size(current_function
->sig
->return_type
); i
++) {
2841 emit(ir
, TGSI_OPCODE_MOV
, l
, r
);
2847 emit(ir
, TGSI_OPCODE_RET
);
2851 glsl_to_tgsi_visitor::visit(ir_discard
*ir
)
2853 if (ir
->condition
) {
2854 ir
->condition
->accept(this);
2855 this->result
.negate
= ~this->result
.negate
;
2856 emit(ir
, TGSI_OPCODE_KIL
, undef_dst
, this->result
);
2858 emit(ir
, TGSI_OPCODE_KILP
);
2863 glsl_to_tgsi_visitor::visit(ir_if
*ir
)
2865 glsl_to_tgsi_instruction
*cond_inst
, *if_inst
;
2866 glsl_to_tgsi_instruction
*prev_inst
;
2868 prev_inst
= (glsl_to_tgsi_instruction
*)this->instructions
.get_tail();
2870 ir
->condition
->accept(this);
2871 assert(this->result
.file
!= PROGRAM_UNDEFINED
);
2873 if (this->options
->EmitCondCodes
) {
2874 cond_inst
= (glsl_to_tgsi_instruction
*)this->instructions
.get_tail();
2876 /* See if we actually generated any instruction for generating
2877 * the condition. If not, then cook up a move to a temp so we
2878 * have something to set cond_update on.
2880 if (cond_inst
== prev_inst
) {
2881 st_src_reg temp
= get_temp(glsl_type::bool_type
);
2882 cond_inst
= emit(ir
->condition
, TGSI_OPCODE_MOV
, st_dst_reg(temp
), result
);
2884 cond_inst
->cond_update
= GL_TRUE
;
2886 if_inst
= emit(ir
->condition
, TGSI_OPCODE_IF
);
2887 if_inst
->dst
.cond_mask
= COND_NE
;
2889 if_inst
= emit(ir
->condition
, TGSI_OPCODE_IF
, undef_dst
, this->result
);
2892 this->instructions
.push_tail(if_inst
);
2894 visit_exec_list(&ir
->then_instructions
, this);
2896 if (!ir
->else_instructions
.is_empty()) {
2897 emit(ir
->condition
, TGSI_OPCODE_ELSE
);
2898 visit_exec_list(&ir
->else_instructions
, this);
2901 if_inst
= emit(ir
->condition
, TGSI_OPCODE_ENDIF
);
2904 glsl_to_tgsi_visitor::glsl_to_tgsi_visitor()
2906 result
.file
= PROGRAM_UNDEFINED
;
2908 next_signature_id
= 1;
2910 current_function
= NULL
;
2911 num_address_regs
= 0;
2913 indirect_addr_temps
= false;
2914 indirect_addr_consts
= false;
2916 native_integers
= false;
2917 mem_ctx
= ralloc_context(NULL
);
2920 shader_program
= NULL
;
2924 glsl_to_tgsi_visitor::~glsl_to_tgsi_visitor()
2926 ralloc_free(mem_ctx
);
2929 extern "C" void free_glsl_to_tgsi_visitor(glsl_to_tgsi_visitor
*v
)
2936 * Count resources used by the given gpu program (number of texture
2940 count_resources(glsl_to_tgsi_visitor
*v
, gl_program
*prog
)
2942 v
->samplers_used
= 0;
2944 foreach_iter(exec_list_iterator
, iter
, v
->instructions
) {
2945 glsl_to_tgsi_instruction
*inst
= (glsl_to_tgsi_instruction
*)iter
.get();
2947 if (is_tex_instruction(inst
->op
)) {
2948 v
->samplers_used
|= 1 << inst
->sampler
;
2950 if (inst
->tex_shadow
) {
2951 prog
->ShadowSamplers
|= 1 << inst
->sampler
;
2956 prog
->SamplersUsed
= v
->samplers_used
;
2958 if (v
->shader_program
!= NULL
)
2959 _mesa_update_shader_textures_used(v
->shader_program
, prog
);
2963 set_uniform_initializer(struct gl_context
*ctx
, void *mem_ctx
,
2964 struct gl_shader_program
*shader_program
,
2965 const char *name
, const glsl_type
*type
,
2968 if (type
->is_record()) {
2969 ir_constant
*field_constant
;
2971 field_constant
= (ir_constant
*)val
->components
.get_head();
2973 for (unsigned int i
= 0; i
< type
->length
; i
++) {
2974 const glsl_type
*field_type
= type
->fields
.structure
[i
].type
;
2975 const char *field_name
= ralloc_asprintf(mem_ctx
, "%s.%s", name
,
2976 type
->fields
.structure
[i
].name
);
2977 set_uniform_initializer(ctx
, mem_ctx
, shader_program
, field_name
,
2978 field_type
, field_constant
);
2979 field_constant
= (ir_constant
*)field_constant
->next
;
2985 unsigned index
= _mesa_get_uniform_location(ctx
, shader_program
, name
,
2987 if (offset
== GL_INVALID_INDEX
) {
2988 fail_link(shader_program
,
2989 "Couldn't find uniform for initializer %s\n", name
);
2992 int loc
= _mesa_uniform_merge_location_offset(index
, offset
);
2994 for (unsigned int i
= 0; i
< (type
->is_array() ? type
->length
: 1); i
++) {
2995 ir_constant
*element
;
2996 const glsl_type
*element_type
;
2997 if (type
->is_array()) {
2998 element
= val
->array_elements
[i
];
2999 element_type
= type
->fields
.array
;
3002 element_type
= type
;
3007 if (element_type
->base_type
== GLSL_TYPE_BOOL
) {
3008 int *conv
= ralloc_array(mem_ctx
, int, element_type
->components());
3009 for (unsigned int j
= 0; j
< element_type
->components(); j
++) {
3010 conv
[j
] = element
->value
.b
[j
];
3012 values
= (void *)conv
;
3013 element_type
= glsl_type::get_instance(GLSL_TYPE_INT
,
3014 element_type
->vector_elements
,
3017 values
= &element
->value
;
3020 if (element_type
->is_matrix()) {
3021 _mesa_uniform_matrix(ctx
, shader_program
,
3022 element_type
->matrix_columns
,
3023 element_type
->vector_elements
,
3024 loc
, 1, GL_FALSE
, (GLfloat
*)values
);
3026 _mesa_uniform(ctx
, shader_program
, loc
, element_type
->matrix_columns
,
3027 values
, element_type
->gl_type
);
3035 * Returns the mask of channels (bitmask of WRITEMASK_X,Y,Z,W) which
3036 * are read from the given src in this instruction
3039 get_src_arg_mask(st_dst_reg dst
, st_src_reg src
)
3041 int read_mask
= 0, comp
;
3043 /* Now, given the src swizzle and the written channels, find which
3044 * components are actually read
3046 for (comp
= 0; comp
< 4; ++comp
) {
3047 const unsigned coord
= GET_SWZ(src
.swizzle
, comp
);
3049 if (dst
.writemask
& (1 << comp
) && coord
<= SWIZZLE_W
)
3050 read_mask
|= 1 << coord
;
3057 * This pass replaces CMP T0, T1 T2 T0 with MOV T0, T2 when the CMP
3058 * instruction is the first instruction to write to register T0. There are
3059 * several lowering passes done in GLSL IR (e.g. branches and
3060 * relative addressing) that create a large number of conditional assignments
3061 * that ir_to_mesa converts to CMP instructions like the one mentioned above.
3063 * Here is why this conversion is safe:
3064 * CMP T0, T1 T2 T0 can be expanded to:
3070 * If (T1 < 0.0) evaluates to true then our replacement MOV T0, T2 is the same
3071 * as the original program. If (T1 < 0.0) evaluates to false, executing
3072 * MOV T0, T0 will store a garbage value in T0 since T0 is uninitialized.
3073 * Therefore, it doesn't matter that we are replacing MOV T0, T0 with MOV T0, T2
3074 * because any instruction that was going to read from T0 after this was going
3075 * to read a garbage value anyway.
3078 glsl_to_tgsi_visitor::simplify_cmp(void)
3080 unsigned *tempWrites
;
3081 unsigned outputWrites
[MAX_PROGRAM_OUTPUTS
];
3083 tempWrites
= new unsigned[MAX_TEMPS
];
3087 memset(tempWrites
, 0, sizeof(unsigned) * MAX_TEMPS
);
3088 memset(outputWrites
, 0, sizeof(outputWrites
));
3090 foreach_iter(exec_list_iterator
, iter
, this->instructions
) {
3091 glsl_to_tgsi_instruction
*inst
= (glsl_to_tgsi_instruction
*)iter
.get();
3092 unsigned prevWriteMask
= 0;
3094 /* Give up if we encounter relative addressing or flow control. */
3095 if (inst
->dst
.reladdr
||
3096 tgsi_get_opcode_info(inst
->op
)->is_branch
||
3097 inst
->op
== TGSI_OPCODE_BGNSUB
||
3098 inst
->op
== TGSI_OPCODE_CONT
||
3099 inst
->op
== TGSI_OPCODE_END
||
3100 inst
->op
== TGSI_OPCODE_ENDSUB
||
3101 inst
->op
== TGSI_OPCODE_RET
) {
3105 if (inst
->dst
.file
== PROGRAM_OUTPUT
) {
3106 assert(inst
->dst
.index
< MAX_PROGRAM_OUTPUTS
);
3107 prevWriteMask
= outputWrites
[inst
->dst
.index
];
3108 outputWrites
[inst
->dst
.index
] |= inst
->dst
.writemask
;
3109 } else if (inst
->dst
.file
== PROGRAM_TEMPORARY
) {
3110 assert(inst
->dst
.index
< MAX_TEMPS
);
3111 prevWriteMask
= tempWrites
[inst
->dst
.index
];
3112 tempWrites
[inst
->dst
.index
] |= inst
->dst
.writemask
;
3115 /* For a CMP to be considered a conditional write, the destination
3116 * register and source register two must be the same. */
3117 if (inst
->op
== TGSI_OPCODE_CMP
3118 && !(inst
->dst
.writemask
& prevWriteMask
)
3119 && inst
->src
[2].file
== inst
->dst
.file
3120 && inst
->src
[2].index
== inst
->dst
.index
3121 && inst
->dst
.writemask
== get_src_arg_mask(inst
->dst
, inst
->src
[2])) {
3123 inst
->op
= TGSI_OPCODE_MOV
;
3124 inst
->src
[0] = inst
->src
[1];
3128 delete [] tempWrites
;
3131 /* Replaces all references to a temporary register index with another index. */
3133 glsl_to_tgsi_visitor::rename_temp_register(int index
, int new_index
)
3135 foreach_iter(exec_list_iterator
, iter
, this->instructions
) {
3136 glsl_to_tgsi_instruction
*inst
= (glsl_to_tgsi_instruction
*)iter
.get();
3139 for (j
=0; j
< num_inst_src_regs(inst
->op
); j
++) {
3140 if (inst
->src
[j
].file
== PROGRAM_TEMPORARY
&&
3141 inst
->src
[j
].index
== index
) {
3142 inst
->src
[j
].index
= new_index
;
3146 if (inst
->dst
.file
== PROGRAM_TEMPORARY
&& inst
->dst
.index
== index
) {
3147 inst
->dst
.index
= new_index
;
3153 glsl_to_tgsi_visitor::get_first_temp_read(int index
)
3155 int depth
= 0; /* loop depth */
3156 int loop_start
= -1; /* index of the first active BGNLOOP (if any) */
3159 foreach_iter(exec_list_iterator
, iter
, this->instructions
) {
3160 glsl_to_tgsi_instruction
*inst
= (glsl_to_tgsi_instruction
*)iter
.get();
3162 for (j
=0; j
< num_inst_src_regs(inst
->op
); j
++) {
3163 if (inst
->src
[j
].file
== PROGRAM_TEMPORARY
&&
3164 inst
->src
[j
].index
== index
) {
3165 return (depth
== 0) ? i
: loop_start
;
3169 if (inst
->op
== TGSI_OPCODE_BGNLOOP
) {
3172 } else if (inst
->op
== TGSI_OPCODE_ENDLOOP
) {
3185 glsl_to_tgsi_visitor::get_first_temp_write(int index
)
3187 int depth
= 0; /* loop depth */
3188 int loop_start
= -1; /* index of the first active BGNLOOP (if any) */
3191 foreach_iter(exec_list_iterator
, iter
, this->instructions
) {
3192 glsl_to_tgsi_instruction
*inst
= (glsl_to_tgsi_instruction
*)iter
.get();
3194 if (inst
->dst
.file
== PROGRAM_TEMPORARY
&& inst
->dst
.index
== index
) {
3195 return (depth
== 0) ? i
: loop_start
;
3198 if (inst
->op
== TGSI_OPCODE_BGNLOOP
) {
3201 } else if (inst
->op
== TGSI_OPCODE_ENDLOOP
) {
3214 glsl_to_tgsi_visitor::get_last_temp_read(int index
)
3216 int depth
= 0; /* loop depth */
3217 int last
= -1; /* index of last instruction that reads the temporary */
3220 foreach_iter(exec_list_iterator
, iter
, this->instructions
) {
3221 glsl_to_tgsi_instruction
*inst
= (glsl_to_tgsi_instruction
*)iter
.get();
3223 for (j
=0; j
< num_inst_src_regs(inst
->op
); j
++) {
3224 if (inst
->src
[j
].file
== PROGRAM_TEMPORARY
&&
3225 inst
->src
[j
].index
== index
) {
3226 last
= (depth
== 0) ? i
: -2;
3230 if (inst
->op
== TGSI_OPCODE_BGNLOOP
)
3232 else if (inst
->op
== TGSI_OPCODE_ENDLOOP
)
3233 if (--depth
== 0 && last
== -2)
3245 glsl_to_tgsi_visitor::get_last_temp_write(int index
)
3247 int depth
= 0; /* loop depth */
3248 int last
= -1; /* index of last instruction that writes to the temporary */
3251 foreach_iter(exec_list_iterator
, iter
, this->instructions
) {
3252 glsl_to_tgsi_instruction
*inst
= (glsl_to_tgsi_instruction
*)iter
.get();
3254 if (inst
->dst
.file
== PROGRAM_TEMPORARY
&& inst
->dst
.index
== index
)
3255 last
= (depth
== 0) ? i
: -2;
3257 if (inst
->op
== TGSI_OPCODE_BGNLOOP
)
3259 else if (inst
->op
== TGSI_OPCODE_ENDLOOP
)
3260 if (--depth
== 0 && last
== -2)
3272 * On a basic block basis, tracks available PROGRAM_TEMPORARY register
3273 * channels for copy propagation and updates following instructions to
3274 * use the original versions.
3276 * The glsl_to_tgsi_visitor lazily produces code assuming that this pass
3277 * will occur. As an example, a TXP production before this pass:
3279 * 0: MOV TEMP[1], INPUT[4].xyyy;
3280 * 1: MOV TEMP[1].w, INPUT[4].wwww;
3281 * 2: TXP TEMP[2], TEMP[1], texture[0], 2D;
3285 * 0: MOV TEMP[1], INPUT[4].xyyy;
3286 * 1: MOV TEMP[1].w, INPUT[4].wwww;
3287 * 2: TXP TEMP[2], INPUT[4].xyyw, texture[0], 2D;
3289 * which allows for dead code elimination on TEMP[1]'s writes.
3292 glsl_to_tgsi_visitor::copy_propagate(void)
3294 glsl_to_tgsi_instruction
**acp
= rzalloc_array(mem_ctx
,
3295 glsl_to_tgsi_instruction
*,
3296 this->next_temp
* 4);
3297 int *acp_level
= rzalloc_array(mem_ctx
, int, this->next_temp
* 4);
3300 foreach_iter(exec_list_iterator
, iter
, this->instructions
) {
3301 glsl_to_tgsi_instruction
*inst
= (glsl_to_tgsi_instruction
*)iter
.get();
3303 assert(inst
->dst
.file
!= PROGRAM_TEMPORARY
3304 || inst
->dst
.index
< this->next_temp
);
3306 /* First, do any copy propagation possible into the src regs. */
3307 for (int r
= 0; r
< 3; r
++) {
3308 glsl_to_tgsi_instruction
*first
= NULL
;
3310 int acp_base
= inst
->src
[r
].index
* 4;
3312 if (inst
->src
[r
].file
!= PROGRAM_TEMPORARY
||
3313 inst
->src
[r
].reladdr
)
3316 /* See if we can find entries in the ACP consisting of MOVs
3317 * from the same src register for all the swizzled channels
3318 * of this src register reference.
3320 for (int i
= 0; i
< 4; i
++) {
3321 int src_chan
= GET_SWZ(inst
->src
[r
].swizzle
, i
);
3322 glsl_to_tgsi_instruction
*copy_chan
= acp
[acp_base
+ src_chan
];
3329 assert(acp_level
[acp_base
+ src_chan
] <= level
);
3334 if (first
->src
[0].file
!= copy_chan
->src
[0].file
||
3335 first
->src
[0].index
!= copy_chan
->src
[0].index
) {
3343 /* We've now validated that we can copy-propagate to
3344 * replace this src register reference. Do it.
3346 inst
->src
[r
].file
= first
->src
[0].file
;
3347 inst
->src
[r
].index
= first
->src
[0].index
;
3350 for (int i
= 0; i
< 4; i
++) {
3351 int src_chan
= GET_SWZ(inst
->src
[r
].swizzle
, i
);
3352 glsl_to_tgsi_instruction
*copy_inst
= acp
[acp_base
+ src_chan
];
3353 swizzle
|= (GET_SWZ(copy_inst
->src
[0].swizzle
, src_chan
) <<
3356 inst
->src
[r
].swizzle
= swizzle
;
3361 case TGSI_OPCODE_BGNLOOP
:
3362 case TGSI_OPCODE_ENDLOOP
:
3363 /* End of a basic block, clear the ACP entirely. */
3364 memset(acp
, 0, sizeof(*acp
) * this->next_temp
* 4);
3367 case TGSI_OPCODE_IF
:
3371 case TGSI_OPCODE_ENDIF
:
3372 case TGSI_OPCODE_ELSE
:
3373 /* Clear all channels written inside the block from the ACP, but
3374 * leaving those that were not touched.
3376 for (int r
= 0; r
< this->next_temp
; r
++) {
3377 for (int c
= 0; c
< 4; c
++) {
3378 if (!acp
[4 * r
+ c
])
3381 if (acp_level
[4 * r
+ c
] >= level
)
3382 acp
[4 * r
+ c
] = NULL
;
3385 if (inst
->op
== TGSI_OPCODE_ENDIF
)
3390 /* Continuing the block, clear any written channels from
3393 if (inst
->dst
.file
== PROGRAM_TEMPORARY
&& inst
->dst
.reladdr
) {
3394 /* Any temporary might be written, so no copy propagation
3395 * across this instruction.
3397 memset(acp
, 0, sizeof(*acp
) * this->next_temp
* 4);
3398 } else if (inst
->dst
.file
== PROGRAM_OUTPUT
&&
3399 inst
->dst
.reladdr
) {
3400 /* Any output might be written, so no copy propagation
3401 * from outputs across this instruction.
3403 for (int r
= 0; r
< this->next_temp
; r
++) {
3404 for (int c
= 0; c
< 4; c
++) {
3405 if (!acp
[4 * r
+ c
])
3408 if (acp
[4 * r
+ c
]->src
[0].file
== PROGRAM_OUTPUT
)
3409 acp
[4 * r
+ c
] = NULL
;
3412 } else if (inst
->dst
.file
== PROGRAM_TEMPORARY
||
3413 inst
->dst
.file
== PROGRAM_OUTPUT
) {
3414 /* Clear where it's used as dst. */
3415 if (inst
->dst
.file
== PROGRAM_TEMPORARY
) {
3416 for (int c
= 0; c
< 4; c
++) {
3417 if (inst
->dst
.writemask
& (1 << c
)) {
3418 acp
[4 * inst
->dst
.index
+ c
] = NULL
;
3423 /* Clear where it's used as src. */
3424 for (int r
= 0; r
< this->next_temp
; r
++) {
3425 for (int c
= 0; c
< 4; c
++) {
3426 if (!acp
[4 * r
+ c
])
3429 int src_chan
= GET_SWZ(acp
[4 * r
+ c
]->src
[0].swizzle
, c
);
3431 if (acp
[4 * r
+ c
]->src
[0].file
== inst
->dst
.file
&&
3432 acp
[4 * r
+ c
]->src
[0].index
== inst
->dst
.index
&&
3433 inst
->dst
.writemask
& (1 << src_chan
))
3435 acp
[4 * r
+ c
] = NULL
;
3443 /* If this is a copy, add it to the ACP. */
3444 if (inst
->op
== TGSI_OPCODE_MOV
&&
3445 inst
->dst
.file
== PROGRAM_TEMPORARY
&&
3446 !inst
->dst
.reladdr
&&
3448 !inst
->src
[0].reladdr
&&
3449 !inst
->src
[0].negate
) {
3450 for (int i
= 0; i
< 4; i
++) {
3451 if (inst
->dst
.writemask
& (1 << i
)) {
3452 acp
[4 * inst
->dst
.index
+ i
] = inst
;
3453 acp_level
[4 * inst
->dst
.index
+ i
] = level
;
3459 ralloc_free(acp_level
);
3464 * Tracks available PROGRAM_TEMPORARY registers for dead code elimination.
3466 * The glsl_to_tgsi_visitor lazily produces code assuming that this pass
3467 * will occur. As an example, a TXP production after copy propagation but
3470 * 0: MOV TEMP[1], INPUT[4].xyyy;
3471 * 1: MOV TEMP[1].w, INPUT[4].wwww;
3472 * 2: TXP TEMP[2], INPUT[4].xyyw, texture[0], 2D;
3474 * and after this pass:
3476 * 0: TXP TEMP[2], INPUT[4].xyyw, texture[0], 2D;
3478 * FIXME: assumes that all functions are inlined (no support for BGNSUB/ENDSUB)
3479 * FIXME: doesn't eliminate all dead code inside of loops; it steps around them
3482 glsl_to_tgsi_visitor::eliminate_dead_code(void)
3486 for (i
=0; i
< this->next_temp
; i
++) {
3487 int last_read
= get_last_temp_read(i
);
3490 foreach_iter(exec_list_iterator
, iter
, this->instructions
) {
3491 glsl_to_tgsi_instruction
*inst
= (glsl_to_tgsi_instruction
*)iter
.get();
3493 if (inst
->dst
.file
== PROGRAM_TEMPORARY
&& inst
->dst
.index
== i
&&
3506 * On a basic block basis, tracks available PROGRAM_TEMPORARY registers for dead
3507 * code elimination. This is less primitive than eliminate_dead_code(), as it
3508 * is per-channel and can detect consecutive writes without a read between them
3509 * as dead code. However, there is some dead code that can be eliminated by
3510 * eliminate_dead_code() but not this function - for example, this function
3511 * cannot eliminate an instruction writing to a register that is never read and
3512 * is the only instruction writing to that register.
3514 * The glsl_to_tgsi_visitor lazily produces code assuming that this pass
3518 glsl_to_tgsi_visitor::eliminate_dead_code_advanced(void)
3520 glsl_to_tgsi_instruction
**writes
= rzalloc_array(mem_ctx
,
3521 glsl_to_tgsi_instruction
*,
3522 this->next_temp
* 4);
3523 int *write_level
= rzalloc_array(mem_ctx
, int, this->next_temp
* 4);
3527 foreach_iter(exec_list_iterator
, iter
, this->instructions
) {
3528 glsl_to_tgsi_instruction
*inst
= (glsl_to_tgsi_instruction
*)iter
.get();
3530 assert(inst
->dst
.file
!= PROGRAM_TEMPORARY
3531 || inst
->dst
.index
< this->next_temp
);
3534 case TGSI_OPCODE_BGNLOOP
:
3535 case TGSI_OPCODE_ENDLOOP
:
3536 case TGSI_OPCODE_CONT
:
3537 case TGSI_OPCODE_BRK
:
3538 /* End of a basic block, clear the write array entirely.
3540 * This keeps us from killing dead code when the writes are
3541 * on either side of a loop, even when the register isn't touched
3542 * inside the loop. However, glsl_to_tgsi_visitor doesn't seem to emit
3543 * dead code of this type, so it shouldn't make a difference as long as
3544 * the dead code elimination pass in the GLSL compiler does its job.
3546 memset(writes
, 0, sizeof(*writes
) * this->next_temp
* 4);
3549 case TGSI_OPCODE_ENDIF
:
3550 case TGSI_OPCODE_ELSE
:
3551 /* Promote the recorded level of all channels written inside the
3552 * preceding if or else block to the level above the if/else block.
3554 for (int r
= 0; r
< this->next_temp
; r
++) {
3555 for (int c
= 0; c
< 4; c
++) {
3556 if (!writes
[4 * r
+ c
])
3559 if (write_level
[4 * r
+ c
] == level
)
3560 write_level
[4 * r
+ c
] = level
-1;
3564 if(inst
->op
== TGSI_OPCODE_ENDIF
)
3569 case TGSI_OPCODE_IF
:
3571 /* fallthrough to default case to mark the condition as read */
3574 /* Continuing the block, clear any channels from the write array that
3575 * are read by this instruction.
3577 for (unsigned i
= 0; i
< Elements(inst
->src
); i
++) {
3578 if (inst
->src
[i
].file
== PROGRAM_TEMPORARY
&& inst
->src
[i
].reladdr
){
3579 /* Any temporary might be read, so no dead code elimination
3580 * across this instruction.
3582 memset(writes
, 0, sizeof(*writes
) * this->next_temp
* 4);
3583 } else if (inst
->src
[i
].file
== PROGRAM_TEMPORARY
) {
3584 /* Clear where it's used as src. */
3585 int src_chans
= 1 << GET_SWZ(inst
->src
[i
].swizzle
, 0);
3586 src_chans
|= 1 << GET_SWZ(inst
->src
[i
].swizzle
, 1);
3587 src_chans
|= 1 << GET_SWZ(inst
->src
[i
].swizzle
, 2);
3588 src_chans
|= 1 << GET_SWZ(inst
->src
[i
].swizzle
, 3);
3590 for (int c
= 0; c
< 4; c
++) {
3591 if (src_chans
& (1 << c
)) {
3592 writes
[4 * inst
->src
[i
].index
+ c
] = NULL
;
3600 /* If this instruction writes to a temporary, add it to the write array.
3601 * If there is already an instruction in the write array for one or more
3602 * of the channels, flag that channel write as dead.
3604 if (inst
->dst
.file
== PROGRAM_TEMPORARY
&&
3605 !inst
->dst
.reladdr
&&
3607 for (int c
= 0; c
< 4; c
++) {
3608 if (inst
->dst
.writemask
& (1 << c
)) {
3609 if (writes
[4 * inst
->dst
.index
+ c
]) {
3610 if (write_level
[4 * inst
->dst
.index
+ c
] < level
)
3613 writes
[4 * inst
->dst
.index
+ c
]->dead_mask
|= (1 << c
);
3615 writes
[4 * inst
->dst
.index
+ c
] = inst
;
3616 write_level
[4 * inst
->dst
.index
+ c
] = level
;
3622 /* Anything still in the write array at this point is dead code. */
3623 for (int r
= 0; r
< this->next_temp
; r
++) {
3624 for (int c
= 0; c
< 4; c
++) {
3625 glsl_to_tgsi_instruction
*inst
= writes
[4 * r
+ c
];
3627 inst
->dead_mask
|= (1 << c
);
3631 /* Now actually remove the instructions that are completely dead and update
3632 * the writemask of other instructions with dead channels.
3634 foreach_iter(exec_list_iterator
, iter
, this->instructions
) {
3635 glsl_to_tgsi_instruction
*inst
= (glsl_to_tgsi_instruction
*)iter
.get();
3637 if (!inst
->dead_mask
|| !inst
->dst
.writemask
)
3639 else if ((inst
->dst
.writemask
& ~inst
->dead_mask
) == 0) {
3644 inst
->dst
.writemask
&= ~(inst
->dead_mask
);
3647 ralloc_free(write_level
);
3648 ralloc_free(writes
);
3653 /* Merges temporary registers together where possible to reduce the number of
3654 * registers needed to run a program.
3656 * Produces optimal code only after copy propagation and dead code elimination
3659 glsl_to_tgsi_visitor::merge_registers(void)
3661 int *last_reads
= rzalloc_array(mem_ctx
, int, this->next_temp
);
3662 int *first_writes
= rzalloc_array(mem_ctx
, int, this->next_temp
);
3665 /* Read the indices of the last read and first write to each temp register
3666 * into an array so that we don't have to traverse the instruction list as
3668 for (i
=0; i
< this->next_temp
; i
++) {
3669 last_reads
[i
] = get_last_temp_read(i
);
3670 first_writes
[i
] = get_first_temp_write(i
);
3673 /* Start looking for registers with non-overlapping usages that can be
3674 * merged together. */
3675 for (i
=0; i
< this->next_temp
; i
++) {
3676 /* Don't touch unused registers. */
3677 if (last_reads
[i
] < 0 || first_writes
[i
] < 0) continue;
3679 for (j
=0; j
< this->next_temp
; j
++) {
3680 /* Don't touch unused registers. */
3681 if (last_reads
[j
] < 0 || first_writes
[j
] < 0) continue;
3683 /* We can merge the two registers if the first write to j is after or
3684 * in the same instruction as the last read from i. Note that the
3685 * register at index i will always be used earlier or at the same time
3686 * as the register at index j. */
3687 if (first_writes
[i
] <= first_writes
[j
] &&
3688 last_reads
[i
] <= first_writes
[j
])
3690 rename_temp_register(j
, i
); /* Replace all references to j with i.*/
3692 /* Update the first_writes and last_reads arrays with the new
3693 * values for the merged register index, and mark the newly unused
3694 * register index as such. */
3695 last_reads
[i
] = last_reads
[j
];
3696 first_writes
[j
] = -1;
3702 ralloc_free(last_reads
);
3703 ralloc_free(first_writes
);
3706 /* Reassign indices to temporary registers by reusing unused indices created
3707 * by optimization passes. */
3709 glsl_to_tgsi_visitor::renumber_registers(void)
3714 for (i
=0; i
< this->next_temp
; i
++) {
3715 if (get_first_temp_read(i
) < 0) continue;
3717 rename_temp_register(i
, new_index
);
3721 this->next_temp
= new_index
;
3725 * Returns a fragment program which implements the current pixel transfer ops.
3726 * Based on get_pixel_transfer_program in st_atom_pixeltransfer.c.
3729 get_pixel_transfer_visitor(struct st_fragment_program
*fp
,
3730 glsl_to_tgsi_visitor
*original
,
3731 int scale_and_bias
, int pixel_maps
)
3733 glsl_to_tgsi_visitor
*v
= new glsl_to_tgsi_visitor();
3734 struct st_context
*st
= st_context(original
->ctx
);
3735 struct gl_program
*prog
= &fp
->Base
.Base
;
3736 struct gl_program_parameter_list
*params
= _mesa_new_parameter_list();
3737 st_src_reg coord
, src0
;
3739 glsl_to_tgsi_instruction
*inst
;
3741 /* Copy attributes of the glsl_to_tgsi_visitor in the original shader. */
3742 v
->ctx
= original
->ctx
;
3744 v
->shader_program
= NULL
;
3745 v
->glsl_version
= original
->glsl_version
;
3746 v
->native_integers
= original
->native_integers
;
3747 v
->options
= original
->options
;
3748 v
->next_temp
= original
->next_temp
;
3749 v
->num_address_regs
= original
->num_address_regs
;
3750 v
->samplers_used
= prog
->SamplersUsed
= original
->samplers_used
;
3751 v
->indirect_addr_temps
= original
->indirect_addr_temps
;
3752 v
->indirect_addr_consts
= original
->indirect_addr_consts
;
3753 memcpy(&v
->immediates
, &original
->immediates
, sizeof(v
->immediates
));
3754 v
->num_immediates
= original
->num_immediates
;
3757 * Get initial pixel color from the texture.
3758 * TEX colorTemp, fragment.texcoord[0], texture[0], 2D;
3760 coord
= st_src_reg(PROGRAM_INPUT
, FRAG_ATTRIB_TEX0
, glsl_type::vec2_type
);
3761 src0
= v
->get_temp(glsl_type::vec4_type
);
3762 dst0
= st_dst_reg(src0
);
3763 inst
= v
->emit(NULL
, TGSI_OPCODE_TEX
, dst0
, coord
);
3765 inst
->tex_target
= TEXTURE_2D_INDEX
;
3767 prog
->InputsRead
|= FRAG_BIT_TEX0
;
3768 prog
->SamplersUsed
|= (1 << 0); /* mark sampler 0 as used */
3769 v
->samplers_used
|= (1 << 0);
3771 if (scale_and_bias
) {
3772 static const gl_state_index scale_state
[STATE_LENGTH
] =
3773 { STATE_INTERNAL
, STATE_PT_SCALE
,
3774 (gl_state_index
) 0, (gl_state_index
) 0, (gl_state_index
) 0 };
3775 static const gl_state_index bias_state
[STATE_LENGTH
] =
3776 { STATE_INTERNAL
, STATE_PT_BIAS
,
3777 (gl_state_index
) 0, (gl_state_index
) 0, (gl_state_index
) 0 };
3778 GLint scale_p
, bias_p
;
3779 st_src_reg scale
, bias
;
3781 scale_p
= _mesa_add_state_reference(params
, scale_state
);
3782 bias_p
= _mesa_add_state_reference(params
, bias_state
);
3784 /* MAD colorTemp, colorTemp, scale, bias; */
3785 scale
= st_src_reg(PROGRAM_STATE_VAR
, scale_p
, GLSL_TYPE_FLOAT
);
3786 bias
= st_src_reg(PROGRAM_STATE_VAR
, bias_p
, GLSL_TYPE_FLOAT
);
3787 inst
= v
->emit(NULL
, TGSI_OPCODE_MAD
, dst0
, src0
, scale
, bias
);
3791 st_src_reg temp
= v
->get_temp(glsl_type::vec4_type
);
3792 st_dst_reg temp_dst
= st_dst_reg(temp
);
3794 assert(st
->pixel_xfer
.pixelmap_texture
);
3796 /* With a little effort, we can do four pixel map look-ups with
3797 * two TEX instructions:
3800 /* TEX temp.rg, colorTemp.rgba, texture[1], 2D; */
3801 temp_dst
.writemask
= WRITEMASK_XY
; /* write R,G */
3802 inst
= v
->emit(NULL
, TGSI_OPCODE_TEX
, temp_dst
, src0
);
3804 inst
->tex_target
= TEXTURE_2D_INDEX
;
3806 /* TEX temp.ba, colorTemp.baba, texture[1], 2D; */
3807 src0
.swizzle
= MAKE_SWIZZLE4(SWIZZLE_Z
, SWIZZLE_W
, SWIZZLE_Z
, SWIZZLE_W
);
3808 temp_dst
.writemask
= WRITEMASK_ZW
; /* write B,A */
3809 inst
= v
->emit(NULL
, TGSI_OPCODE_TEX
, temp_dst
, src0
);
3811 inst
->tex_target
= TEXTURE_2D_INDEX
;
3813 prog
->SamplersUsed
|= (1 << 1); /* mark sampler 1 as used */
3814 v
->samplers_used
|= (1 << 1);
3816 /* MOV colorTemp, temp; */
3817 inst
= v
->emit(NULL
, TGSI_OPCODE_MOV
, dst0
, temp
);
3820 /* Now copy the instructions from the original glsl_to_tgsi_visitor into the
3822 foreach_iter(exec_list_iterator
, iter
, original
->instructions
) {
3823 glsl_to_tgsi_instruction
*inst
= (glsl_to_tgsi_instruction
*)iter
.get();
3824 glsl_to_tgsi_instruction
*newinst
;
3825 st_src_reg src_regs
[3];
3827 if (inst
->dst
.file
== PROGRAM_OUTPUT
)
3828 prog
->OutputsWritten
|= BITFIELD64_BIT(inst
->dst
.index
);
3830 for (int i
=0; i
<3; i
++) {
3831 src_regs
[i
] = inst
->src
[i
];
3832 if (src_regs
[i
].file
== PROGRAM_INPUT
&&
3833 src_regs
[i
].index
== FRAG_ATTRIB_COL0
)
3835 src_regs
[i
].file
= PROGRAM_TEMPORARY
;
3836 src_regs
[i
].index
= src0
.index
;
3838 else if (src_regs
[i
].file
== PROGRAM_INPUT
)
3839 prog
->InputsRead
|= BITFIELD64_BIT(src_regs
[i
].index
);
3842 newinst
= v
->emit(NULL
, inst
->op
, inst
->dst
, src_regs
[0], src_regs
[1], src_regs
[2]);
3843 newinst
->tex_target
= inst
->tex_target
;
3846 /* Make modifications to fragment program info. */
3847 prog
->Parameters
= _mesa_combine_parameter_lists(params
,
3848 original
->prog
->Parameters
);
3849 _mesa_free_parameter_list(params
);
3850 count_resources(v
, prog
);
3851 fp
->glsl_to_tgsi
= v
;
3855 * Make fragment program for glBitmap:
3856 * Sample the texture and kill the fragment if the bit is 0.
3857 * This program will be combined with the user's fragment program.
3859 * Based on make_bitmap_fragment_program in st_cb_bitmap.c.
3862 get_bitmap_visitor(struct st_fragment_program
*fp
,
3863 glsl_to_tgsi_visitor
*original
, int samplerIndex
)
3865 glsl_to_tgsi_visitor
*v
= new glsl_to_tgsi_visitor();
3866 struct st_context
*st
= st_context(original
->ctx
);
3867 struct gl_program
*prog
= &fp
->Base
.Base
;
3868 st_src_reg coord
, src0
;
3870 glsl_to_tgsi_instruction
*inst
;
3872 /* Copy attributes of the glsl_to_tgsi_visitor in the original shader. */
3873 v
->ctx
= original
->ctx
;
3875 v
->shader_program
= NULL
;
3876 v
->glsl_version
= original
->glsl_version
;
3877 v
->native_integers
= original
->native_integers
;
3878 v
->options
= original
->options
;
3879 v
->next_temp
= original
->next_temp
;
3880 v
->num_address_regs
= original
->num_address_regs
;
3881 v
->samplers_used
= prog
->SamplersUsed
= original
->samplers_used
;
3882 v
->indirect_addr_temps
= original
->indirect_addr_temps
;
3883 v
->indirect_addr_consts
= original
->indirect_addr_consts
;
3884 memcpy(&v
->immediates
, &original
->immediates
, sizeof(v
->immediates
));
3885 v
->num_immediates
= original
->num_immediates
;
3887 /* TEX tmp0, fragment.texcoord[0], texture[0], 2D; */
3888 coord
= st_src_reg(PROGRAM_INPUT
, FRAG_ATTRIB_TEX0
, glsl_type::vec2_type
);
3889 src0
= v
->get_temp(glsl_type::vec4_type
);
3890 dst0
= st_dst_reg(src0
);
3891 inst
= v
->emit(NULL
, TGSI_OPCODE_TEX
, dst0
, coord
);
3892 inst
->sampler
= samplerIndex
;
3893 inst
->tex_target
= TEXTURE_2D_INDEX
;
3895 prog
->InputsRead
|= FRAG_BIT_TEX0
;
3896 prog
->SamplersUsed
|= (1 << samplerIndex
); /* mark sampler as used */
3897 v
->samplers_used
|= (1 << samplerIndex
);
3899 /* KIL if -tmp0 < 0 # texel=0 -> keep / texel=0 -> discard */
3900 src0
.negate
= NEGATE_XYZW
;
3901 if (st
->bitmap
.tex_format
== PIPE_FORMAT_L8_UNORM
)
3902 src0
.swizzle
= SWIZZLE_XXXX
;
3903 inst
= v
->emit(NULL
, TGSI_OPCODE_KIL
, undef_dst
, src0
);
3905 /* Now copy the instructions from the original glsl_to_tgsi_visitor into the
3907 foreach_iter(exec_list_iterator
, iter
, original
->instructions
) {
3908 glsl_to_tgsi_instruction
*inst
= (glsl_to_tgsi_instruction
*)iter
.get();
3909 glsl_to_tgsi_instruction
*newinst
;
3910 st_src_reg src_regs
[3];
3912 if (inst
->dst
.file
== PROGRAM_OUTPUT
)
3913 prog
->OutputsWritten
|= BITFIELD64_BIT(inst
->dst
.index
);
3915 for (int i
=0; i
<3; i
++) {
3916 src_regs
[i
] = inst
->src
[i
];
3917 if (src_regs
[i
].file
== PROGRAM_INPUT
)
3918 prog
->InputsRead
|= BITFIELD64_BIT(src_regs
[i
].index
);
3921 newinst
= v
->emit(NULL
, inst
->op
, inst
->dst
, src_regs
[0], src_regs
[1], src_regs
[2]);
3922 newinst
->tex_target
= inst
->tex_target
;
3925 /* Make modifications to fragment program info. */
3926 prog
->Parameters
= _mesa_clone_parameter_list(original
->prog
->Parameters
);
3927 count_resources(v
, prog
);
3928 fp
->glsl_to_tgsi
= v
;
3931 /* ------------------------- TGSI conversion stuff -------------------------- */
3933 unsigned branch_target
;
3938 * Intermediate state used during shader translation.
3940 struct st_translate
{
3941 struct ureg_program
*ureg
;
3943 struct ureg_dst temps
[MAX_TEMPS
];
3944 struct ureg_src
*constants
;
3945 struct ureg_src
*immediates
;
3946 struct ureg_dst outputs
[PIPE_MAX_SHADER_OUTPUTS
];
3947 struct ureg_src inputs
[PIPE_MAX_SHADER_INPUTS
];
3948 struct ureg_dst address
[1];
3949 struct ureg_src samplers
[PIPE_MAX_SAMPLERS
];
3950 struct ureg_src systemValues
[SYSTEM_VALUE_MAX
];
3952 const GLuint
*inputMapping
;
3953 const GLuint
*outputMapping
;
3955 /* For every instruction that contains a label (eg CALL), keep
3956 * details so that we can go back afterwards and emit the correct
3957 * tgsi instruction number for each label.
3959 struct label
*labels
;
3960 unsigned labels_size
;
3961 unsigned labels_count
;
3963 /* Keep a record of the tgsi instruction number that each mesa
3964 * instruction starts at, will be used to fix up labels after
3969 unsigned insn_count
;
3971 unsigned procType
; /**< TGSI_PROCESSOR_VERTEX/FRAGMENT */
3976 /** Map Mesa's SYSTEM_VALUE_x to TGSI_SEMANTIC_x */
3977 static unsigned mesa_sysval_to_semantic
[SYSTEM_VALUE_MAX
] = {
3979 TGSI_SEMANTIC_VERTEXID
,
3980 TGSI_SEMANTIC_INSTANCEID
3984 * Make note of a branch to a label in the TGSI code.
3985 * After we've emitted all instructions, we'll go over the list
3986 * of labels built here and patch the TGSI code with the actual
3987 * location of each label.
3989 static unsigned *get_label(struct st_translate
*t
, unsigned branch_target
)
3993 if (t
->labels_count
+ 1 >= t
->labels_size
) {
3994 t
->labels_size
= 1 << (util_logbase2(t
->labels_size
) + 1);
3995 t
->labels
= (struct label
*)realloc(t
->labels
,
3996 t
->labels_size
* sizeof(struct label
));
3997 if (t
->labels
== NULL
) {
3998 static unsigned dummy
;
4004 i
= t
->labels_count
++;
4005 t
->labels
[i
].branch_target
= branch_target
;
4006 return &t
->labels
[i
].token
;
4010 * Called prior to emitting the TGSI code for each instruction.
4011 * Allocate additional space for instructions if needed.
4012 * Update the insn[] array so the next glsl_to_tgsi_instruction points to
4013 * the next TGSI instruction.
4015 static void set_insn_start(struct st_translate
*t
, unsigned start
)
4017 if (t
->insn_count
+ 1 >= t
->insn_size
) {
4018 t
->insn_size
= 1 << (util_logbase2(t
->insn_size
) + 1);
4019 t
->insn
= (unsigned *)realloc(t
->insn
, t
->insn_size
* sizeof(t
->insn
[0]));
4020 if (t
->insn
== NULL
) {
4026 t
->insn
[t
->insn_count
++] = start
;
4030 * Map a glsl_to_tgsi constant/immediate to a TGSI immediate.
4032 static struct ureg_src
4033 emit_immediate(struct st_translate
*t
,
4034 gl_constant_value values
[4],
4037 struct ureg_program
*ureg
= t
->ureg
;
4042 return ureg_DECL_immediate(ureg
, &values
[0].f
, size
);
4044 return ureg_DECL_immediate_int(ureg
, &values
[0].i
, size
);
4045 case GL_UNSIGNED_INT
:
4047 return ureg_DECL_immediate_uint(ureg
, &values
[0].u
, size
);
4049 assert(!"should not get here - type must be float, int, uint, or bool");
4050 return ureg_src_undef();
4055 * Map a glsl_to_tgsi dst register to a TGSI ureg_dst register.
4057 static struct ureg_dst
4058 dst_register(struct st_translate
*t
,
4059 gl_register_file file
,
4063 case PROGRAM_UNDEFINED
:
4064 return ureg_dst_undef();
4066 case PROGRAM_TEMPORARY
:
4067 if (ureg_dst_is_undef(t
->temps
[index
]))
4068 t
->temps
[index
] = ureg_DECL_local_temporary(t
->ureg
);
4070 return t
->temps
[index
];
4072 case PROGRAM_OUTPUT
:
4073 if (t
->procType
== TGSI_PROCESSOR_VERTEX
)
4074 assert(index
< VERT_RESULT_MAX
);
4075 else if (t
->procType
== TGSI_PROCESSOR_FRAGMENT
)
4076 assert(index
< FRAG_RESULT_MAX
);
4078 assert(index
< GEOM_RESULT_MAX
);
4080 assert(t
->outputMapping
[index
] < Elements(t
->outputs
));
4082 return t
->outputs
[t
->outputMapping
[index
]];
4084 case PROGRAM_ADDRESS
:
4085 return t
->address
[index
];
4088 assert(!"unknown dst register file");
4089 return ureg_dst_undef();
4094 * Map a glsl_to_tgsi src register to a TGSI ureg_src register.
4096 static struct ureg_src
4097 src_register(struct st_translate
*t
,
4098 gl_register_file file
,
4099 GLint index
, GLint index2D
)
4102 case PROGRAM_UNDEFINED
:
4103 return ureg_src_undef();
4105 case PROGRAM_TEMPORARY
:
4107 assert(index
< (int) Elements(t
->temps
));
4108 if (ureg_dst_is_undef(t
->temps
[index
]))
4109 t
->temps
[index
] = ureg_DECL_local_temporary(t
->ureg
);
4110 return ureg_src(t
->temps
[index
]);
4112 case PROGRAM_ENV_PARAM
:
4113 case PROGRAM_LOCAL_PARAM
:
4114 case PROGRAM_UNIFORM
:
4116 return t
->constants
[index
];
4117 case PROGRAM_STATE_VAR
:
4118 case PROGRAM_CONSTANT
: /* ie, immediate */
4120 struct ureg_src src
;
4121 src
= ureg_src_register(TGSI_FILE_CONSTANT
, 0);
4123 src
.DimensionIndex
= index2D
;
4125 } else if (index
< 0)
4126 return ureg_DECL_constant(t
->ureg
, 0);
4128 return t
->constants
[index
];
4130 case PROGRAM_IMMEDIATE
:
4131 return t
->immediates
[index
];
4134 assert(t
->inputMapping
[index
] < Elements(t
->inputs
));
4135 return t
->inputs
[t
->inputMapping
[index
]];
4137 case PROGRAM_OUTPUT
:
4138 assert(t
->outputMapping
[index
] < Elements(t
->outputs
));
4139 return ureg_src(t
->outputs
[t
->outputMapping
[index
]]); /* not needed? */
4141 case PROGRAM_ADDRESS
:
4142 return ureg_src(t
->address
[index
]);
4144 case PROGRAM_SYSTEM_VALUE
:
4145 assert(index
< (int) Elements(t
->systemValues
));
4146 return t
->systemValues
[index
];
4149 assert(!"unknown src register file");
4150 return ureg_src_undef();
4155 * Create a TGSI ureg_dst register from an st_dst_reg.
4157 static struct ureg_dst
4158 translate_dst(struct st_translate
*t
,
4159 const st_dst_reg
*dst_reg
,
4160 bool saturate
, bool clamp_color
)
4162 struct ureg_dst dst
= dst_register(t
,
4166 dst
= ureg_writemask(dst
, dst_reg
->writemask
);
4169 dst
= ureg_saturate(dst
);
4170 else if (clamp_color
&& dst_reg
->file
== PROGRAM_OUTPUT
) {
4171 /* Clamp colors for ARB_color_buffer_float. */
4172 switch (t
->procType
) {
4173 case TGSI_PROCESSOR_VERTEX
:
4174 /* XXX if the geometry shader is present, this must be done there
4175 * instead of here. */
4176 if (dst_reg
->index
== VERT_RESULT_COL0
||
4177 dst_reg
->index
== VERT_RESULT_COL1
||
4178 dst_reg
->index
== VERT_RESULT_BFC0
||
4179 dst_reg
->index
== VERT_RESULT_BFC1
) {
4180 dst
= ureg_saturate(dst
);
4184 case TGSI_PROCESSOR_FRAGMENT
:
4185 if (dst_reg
->index
>= FRAG_RESULT_COLOR
) {
4186 dst
= ureg_saturate(dst
);
4192 if (dst_reg
->reladdr
!= NULL
)
4193 dst
= ureg_dst_indirect(dst
, ureg_src(t
->address
[0]));
4199 * Create a TGSI ureg_src register from an st_src_reg.
4201 static struct ureg_src
4202 translate_src(struct st_translate
*t
, const st_src_reg
*src_reg
)
4204 struct ureg_src src
= src_register(t
, src_reg
->file
, src_reg
->index
, src_reg
->index2D
);
4206 src
= ureg_swizzle(src
,
4207 GET_SWZ(src_reg
->swizzle
, 0) & 0x3,
4208 GET_SWZ(src_reg
->swizzle
, 1) & 0x3,
4209 GET_SWZ(src_reg
->swizzle
, 2) & 0x3,
4210 GET_SWZ(src_reg
->swizzle
, 3) & 0x3);
4212 if ((src_reg
->negate
& 0xf) == NEGATE_XYZW
)
4213 src
= ureg_negate(src
);
4215 if (src_reg
->reladdr
!= NULL
) {
4216 /* Normally ureg_src_indirect() would be used here, but a stupid compiler
4217 * bug in g++ makes ureg_src_indirect (an inline C function) erroneously
4218 * set the bit for src.Negate. So we have to do the operation manually
4219 * here to work around the compiler's problems. */
4220 /*src = ureg_src_indirect(src, ureg_src(t->address[0]));*/
4221 struct ureg_src addr
= ureg_src(t
->address
[0]);
4223 src
.IndirectFile
= addr
.File
;
4224 src
.IndirectIndex
= addr
.Index
;
4225 src
.IndirectSwizzle
= addr
.SwizzleX
;
4227 if (src_reg
->file
!= PROGRAM_INPUT
&&
4228 src_reg
->file
!= PROGRAM_OUTPUT
) {
4229 /* If src_reg->index was negative, it was set to zero in
4230 * src_register(). Reassign it now. But don't do this
4231 * for input/output regs since they get remapped while
4232 * const buffers don't.
4234 src
.Index
= src_reg
->index
;
4241 static struct tgsi_texture_offset
4242 translate_tex_offset(struct st_translate
*t
,
4243 const struct tgsi_texture_offset
*in_offset
)
4245 struct tgsi_texture_offset offset
;
4246 struct ureg_src imm_src
;
4248 assert(in_offset
->File
== PROGRAM_IMMEDIATE
);
4249 imm_src
= t
->immediates
[in_offset
->Index
];
4251 offset
.File
= imm_src
.File
;
4252 offset
.Index
= imm_src
.Index
;
4253 offset
.SwizzleX
= imm_src
.SwizzleX
;
4254 offset
.SwizzleY
= imm_src
.SwizzleY
;
4255 offset
.SwizzleZ
= imm_src
.SwizzleZ
;
4256 offset
.File
= TGSI_FILE_IMMEDIATE
;
4263 compile_tgsi_instruction(struct st_translate
*t
,
4264 const glsl_to_tgsi_instruction
*inst
,
4265 bool clamp_dst_color_output
)
4267 struct ureg_program
*ureg
= t
->ureg
;
4269 struct ureg_dst dst
[1];
4270 struct ureg_src src
[4];
4271 struct tgsi_texture_offset texoffsets
[MAX_GLSL_TEXTURE_OFFSET
];
4275 unsigned tex_target
;
4277 num_dst
= num_inst_dst_regs(inst
->op
);
4278 num_src
= num_inst_src_regs(inst
->op
);
4281 dst
[0] = translate_dst(t
,
4284 clamp_dst_color_output
);
4286 for (i
= 0; i
< num_src
; i
++)
4287 src
[i
] = translate_src(t
, &inst
->src
[i
]);
4290 case TGSI_OPCODE_BGNLOOP
:
4291 case TGSI_OPCODE_CAL
:
4292 case TGSI_OPCODE_ELSE
:
4293 case TGSI_OPCODE_ENDLOOP
:
4294 case TGSI_OPCODE_IF
:
4295 assert(num_dst
== 0);
4296 ureg_label_insn(ureg
,
4300 inst
->op
== TGSI_OPCODE_CAL
? inst
->function
->sig_id
: 0));
4303 case TGSI_OPCODE_TEX
:
4304 case TGSI_OPCODE_TXB
:
4305 case TGSI_OPCODE_TXD
:
4306 case TGSI_OPCODE_TXL
:
4307 case TGSI_OPCODE_TXP
:
4308 case TGSI_OPCODE_TXQ
:
4309 case TGSI_OPCODE_TXF
:
4310 case TGSI_OPCODE_TEX2
:
4311 case TGSI_OPCODE_TXB2
:
4312 case TGSI_OPCODE_TXL2
:
4313 src
[num_src
++] = t
->samplers
[inst
->sampler
];
4314 for (i
= 0; i
< inst
->tex_offset_num_offset
; i
++) {
4315 texoffsets
[i
] = translate_tex_offset(t
, &inst
->tex_offsets
[i
]);
4317 tex_target
= st_translate_texture_target(inst
->tex_target
, inst
->tex_shadow
);
4323 texoffsets
, inst
->tex_offset_num_offset
,
4327 case TGSI_OPCODE_SCS
:
4328 dst
[0] = ureg_writemask(dst
[0], TGSI_WRITEMASK_XY
);
4329 ureg_insn(ureg
, inst
->op
, dst
, num_dst
, src
, num_src
);
4342 * Emit the TGSI instructions for inverting and adjusting WPOS.
4343 * This code is unavoidable because it also depends on whether
4344 * a FBO is bound (STATE_FB_WPOS_Y_TRANSFORM).
4347 emit_wpos_adjustment( struct st_translate
*t
,
4348 const struct gl_program
*program
,
4350 GLfloat adjX
, GLfloat adjY
[2])
4352 struct ureg_program
*ureg
= t
->ureg
;
4354 /* Fragment program uses fragment position input.
4355 * Need to replace instances of INPUT[WPOS] with temp T
4356 * where T = INPUT[WPOS] by y is inverted.
4358 static const gl_state_index wposTransformState
[STATE_LENGTH
]
4359 = { STATE_INTERNAL
, STATE_FB_WPOS_Y_TRANSFORM
,
4360 (gl_state_index
)0, (gl_state_index
)0, (gl_state_index
)0 };
4362 /* XXX: note we are modifying the incoming shader here! Need to
4363 * do this before emitting the constant decls below, or this
4366 unsigned wposTransConst
= _mesa_add_state_reference(program
->Parameters
,
4367 wposTransformState
);
4369 struct ureg_src wpostrans
= ureg_DECL_constant( ureg
, wposTransConst
);
4370 struct ureg_dst wpos_temp
= ureg_DECL_temporary( ureg
);
4371 struct ureg_src wpos_input
= t
->inputs
[t
->inputMapping
[FRAG_ATTRIB_WPOS
]];
4373 /* First, apply the coordinate shift: */
4374 if (adjX
|| adjY
[0] || adjY
[1]) {
4375 if (adjY
[0] != adjY
[1]) {
4376 /* Adjust the y coordinate by adjY[1] or adjY[0] respectively
4377 * depending on whether inversion is actually going to be applied
4378 * or not, which is determined by testing against the inversion
4379 * state variable used below, which will be either +1 or -1.
4381 struct ureg_dst adj_temp
= ureg_DECL_local_temporary(ureg
);
4383 ureg_CMP(ureg
, adj_temp
,
4384 ureg_scalar(wpostrans
, invert
? 2 : 0),
4385 ureg_imm4f(ureg
, adjX
, adjY
[0], 0.0f
, 0.0f
),
4386 ureg_imm4f(ureg
, adjX
, adjY
[1], 0.0f
, 0.0f
));
4387 ureg_ADD(ureg
, wpos_temp
, wpos_input
, ureg_src(adj_temp
));
4389 ureg_ADD(ureg
, wpos_temp
, wpos_input
,
4390 ureg_imm4f(ureg
, adjX
, adjY
[0], 0.0f
, 0.0f
));
4392 wpos_input
= ureg_src(wpos_temp
);
4394 /* MOV wpos_temp, input[wpos]
4396 ureg_MOV( ureg
, wpos_temp
, wpos_input
);
4399 /* Now the conditional y flip: STATE_FB_WPOS_Y_TRANSFORM.xy/zw will be
4400 * inversion/identity, or the other way around if we're drawing to an FBO.
4403 /* MAD wpos_temp.y, wpos_input, wpostrans.xxxx, wpostrans.yyyy
4406 ureg_writemask(wpos_temp
, TGSI_WRITEMASK_Y
),
4408 ureg_scalar(wpostrans
, 0),
4409 ureg_scalar(wpostrans
, 1));
4411 /* MAD wpos_temp.y, wpos_input, wpostrans.zzzz, wpostrans.wwww
4414 ureg_writemask(wpos_temp
, TGSI_WRITEMASK_Y
),
4416 ureg_scalar(wpostrans
, 2),
4417 ureg_scalar(wpostrans
, 3));
4420 /* Use wpos_temp as position input from here on:
4422 t
->inputs
[t
->inputMapping
[FRAG_ATTRIB_WPOS
]] = ureg_src(wpos_temp
);
4427 * Emit fragment position/ooordinate code.
4430 emit_wpos(struct st_context
*st
,
4431 struct st_translate
*t
,
4432 const struct gl_program
*program
,
4433 struct ureg_program
*ureg
)
4435 const struct gl_fragment_program
*fp
=
4436 (const struct gl_fragment_program
*) program
;
4437 struct pipe_screen
*pscreen
= st
->pipe
->screen
;
4438 GLfloat adjX
= 0.0f
;
4439 GLfloat adjY
[2] = { 0.0f
, 0.0f
};
4440 boolean invert
= FALSE
;
4442 /* Query the pixel center conventions supported by the pipe driver and set
4443 * adjX, adjY to help out if it cannot handle the requested one internally.
4445 * The bias of the y-coordinate depends on whether y-inversion takes place
4446 * (adjY[1]) or not (adjY[0]), which is in turn dependent on whether we are
4447 * drawing to an FBO (causes additional inversion), and whether the the pipe
4448 * driver origin and the requested origin differ (the latter condition is
4449 * stored in the 'invert' variable).
4451 * For height = 100 (i = integer, h = half-integer, l = lower, u = upper):
4453 * center shift only:
4458 * l,i -> u,i: ( 0.0 + 1.0) * -1 + 100 = 99
4459 * l,h -> u,h: ( 0.5 + 0.0) * -1 + 100 = 99.5
4460 * u,i -> l,i: (99.0 + 1.0) * -1 + 100 = 0
4461 * u,h -> l,h: (99.5 + 0.0) * -1 + 100 = 0.5
4463 * inversion and center shift:
4464 * l,i -> u,h: ( 0.0 + 0.5) * -1 + 100 = 99.5
4465 * l,h -> u,i: ( 0.5 + 0.5) * -1 + 100 = 99
4466 * u,i -> l,h: (99.0 + 0.5) * -1 + 100 = 0.5
4467 * u,h -> l,i: (99.5 + 0.5) * -1 + 100 = 0
4469 if (fp
->OriginUpperLeft
) {
4470 /* Fragment shader wants origin in upper-left */
4471 if (pscreen
->get_param(pscreen
, PIPE_CAP_TGSI_FS_COORD_ORIGIN_UPPER_LEFT
)) {
4472 /* the driver supports upper-left origin */
4474 else if (pscreen
->get_param(pscreen
, PIPE_CAP_TGSI_FS_COORD_ORIGIN_LOWER_LEFT
)) {
4475 /* the driver supports lower-left origin, need to invert Y */
4476 ureg_property_fs_coord_origin(ureg
, TGSI_FS_COORD_ORIGIN_LOWER_LEFT
);
4483 /* Fragment shader wants origin in lower-left */
4484 if (pscreen
->get_param(pscreen
, PIPE_CAP_TGSI_FS_COORD_ORIGIN_LOWER_LEFT
))
4485 /* the driver supports lower-left origin */
4486 ureg_property_fs_coord_origin(ureg
, TGSI_FS_COORD_ORIGIN_LOWER_LEFT
);
4487 else if (pscreen
->get_param(pscreen
, PIPE_CAP_TGSI_FS_COORD_ORIGIN_UPPER_LEFT
))
4488 /* the driver supports upper-left origin, need to invert Y */
4494 if (fp
->PixelCenterInteger
) {
4495 /* Fragment shader wants pixel center integer */
4496 if (pscreen
->get_param(pscreen
, PIPE_CAP_TGSI_FS_COORD_PIXEL_CENTER_INTEGER
)) {
4497 /* the driver supports pixel center integer */
4499 ureg_property_fs_coord_pixel_center(ureg
, TGSI_FS_COORD_PIXEL_CENTER_INTEGER
);
4501 else if (pscreen
->get_param(pscreen
, PIPE_CAP_TGSI_FS_COORD_PIXEL_CENTER_HALF_INTEGER
)) {
4502 /* the driver supports pixel center half integer, need to bias X,Y */
4511 /* Fragment shader wants pixel center half integer */
4512 if (pscreen
->get_param(pscreen
, PIPE_CAP_TGSI_FS_COORD_PIXEL_CENTER_HALF_INTEGER
)) {
4513 /* the driver supports pixel center half integer */
4515 else if (pscreen
->get_param(pscreen
, PIPE_CAP_TGSI_FS_COORD_PIXEL_CENTER_INTEGER
)) {
4516 /* the driver supports pixel center integer, need to bias X,Y */
4517 adjX
= adjY
[0] = adjY
[1] = 0.5f
;
4518 ureg_property_fs_coord_pixel_center(ureg
, TGSI_FS_COORD_PIXEL_CENTER_INTEGER
);
4524 /* we invert after adjustment so that we avoid the MOV to temporary,
4525 * and reuse the adjustment ADD instead */
4526 emit_wpos_adjustment(t
, program
, invert
, adjX
, adjY
);
4530 * OpenGL's fragment gl_FrontFace input is 1 for front-facing, 0 for back.
4531 * TGSI uses +1 for front, -1 for back.
4532 * This function converts the TGSI value to the GL value. Simply clamping/
4533 * saturating the value to [0,1] does the job.
4536 emit_face_var(struct st_translate
*t
)
4538 struct ureg_program
*ureg
= t
->ureg
;
4539 struct ureg_dst face_temp
= ureg_DECL_temporary(ureg
);
4540 struct ureg_src face_input
= t
->inputs
[t
->inputMapping
[FRAG_ATTRIB_FACE
]];
4542 /* MOV_SAT face_temp, input[face] */
4543 face_temp
= ureg_saturate(face_temp
);
4544 ureg_MOV(ureg
, face_temp
, face_input
);
4546 /* Use face_temp as face input from here on: */
4547 t
->inputs
[t
->inputMapping
[FRAG_ATTRIB_FACE
]] = ureg_src(face_temp
);
4551 emit_edgeflags(struct st_translate
*t
)
4553 struct ureg_program
*ureg
= t
->ureg
;
4554 struct ureg_dst edge_dst
= t
->outputs
[t
->outputMapping
[VERT_RESULT_EDGE
]];
4555 struct ureg_src edge_src
= t
->inputs
[t
->inputMapping
[VERT_ATTRIB_EDGEFLAG
]];
4557 ureg_MOV(ureg
, edge_dst
, edge_src
);
4561 * Translate intermediate IR (glsl_to_tgsi_instruction) to TGSI format.
4562 * \param program the program to translate
4563 * \param numInputs number of input registers used
4564 * \param inputMapping maps Mesa fragment program inputs to TGSI generic
4566 * \param inputSemanticName the TGSI_SEMANTIC flag for each input
4567 * \param inputSemanticIndex the semantic index (ex: which texcoord) for
4569 * \param interpMode the TGSI_INTERPOLATE_LINEAR/PERSP mode for each input
4570 * \param numOutputs number of output registers used
4571 * \param outputMapping maps Mesa fragment program outputs to TGSI
4573 * \param outputSemanticName the TGSI_SEMANTIC flag for each output
4574 * \param outputSemanticIndex the semantic index (ex: which texcoord) for
4577 * \return PIPE_OK or PIPE_ERROR_OUT_OF_MEMORY
4579 extern "C" enum pipe_error
4580 st_translate_program(
4581 struct gl_context
*ctx
,
4583 struct ureg_program
*ureg
,
4584 glsl_to_tgsi_visitor
*program
,
4585 const struct gl_program
*proginfo
,
4587 const GLuint inputMapping
[],
4588 const ubyte inputSemanticName
[],
4589 const ubyte inputSemanticIndex
[],
4590 const GLuint interpMode
[],
4591 const GLboolean is_centroid
[],
4593 const GLuint outputMapping
[],
4594 const ubyte outputSemanticName
[],
4595 const ubyte outputSemanticIndex
[],
4596 boolean passthrough_edgeflags
,
4597 boolean clamp_color
)
4599 struct st_translate
*t
;
4601 enum pipe_error ret
= PIPE_OK
;
4603 assert(numInputs
<= Elements(t
->inputs
));
4604 assert(numOutputs
<= Elements(t
->outputs
));
4606 t
= CALLOC_STRUCT(st_translate
);
4608 ret
= PIPE_ERROR_OUT_OF_MEMORY
;
4612 memset(t
, 0, sizeof *t
);
4614 t
->procType
= procType
;
4615 t
->inputMapping
= inputMapping
;
4616 t
->outputMapping
= outputMapping
;
4619 if (program
->shader_program
) {
4620 for (i
= 0; i
< program
->shader_program
->NumUserUniformStorage
; i
++) {
4621 struct gl_uniform_storage
*const storage
=
4622 &program
->shader_program
->UniformStorage
[i
];
4624 _mesa_uniform_detach_all_driver_storage(storage
);
4629 * Declare input attributes.
4631 if (procType
== TGSI_PROCESSOR_FRAGMENT
) {
4632 for (i
= 0; i
< numInputs
; i
++) {
4633 t
->inputs
[i
] = ureg_DECL_fs_input_cyl_centroid(ureg
,
4634 inputSemanticName
[i
],
4635 inputSemanticIndex
[i
],
4640 if (proginfo
->InputsRead
& FRAG_BIT_WPOS
) {
4641 /* Must do this after setting up t->inputs, and before
4642 * emitting constant references, below:
4644 emit_wpos(st_context(ctx
), t
, proginfo
, ureg
);
4647 if (proginfo
->InputsRead
& FRAG_BIT_FACE
)
4651 * Declare output attributes.
4653 for (i
= 0; i
< numOutputs
; i
++) {
4654 switch (outputSemanticName
[i
]) {
4655 case TGSI_SEMANTIC_POSITION
:
4656 t
->outputs
[i
] = ureg_DECL_output(ureg
,
4657 TGSI_SEMANTIC_POSITION
, /* Z/Depth */
4658 outputSemanticIndex
[i
]);
4659 t
->outputs
[i
] = ureg_writemask(t
->outputs
[i
], TGSI_WRITEMASK_Z
);
4661 case TGSI_SEMANTIC_STENCIL
:
4662 t
->outputs
[i
] = ureg_DECL_output(ureg
,
4663 TGSI_SEMANTIC_STENCIL
, /* Stencil */
4664 outputSemanticIndex
[i
]);
4665 t
->outputs
[i
] = ureg_writemask(t
->outputs
[i
], TGSI_WRITEMASK_Y
);
4667 case TGSI_SEMANTIC_COLOR
:
4668 t
->outputs
[i
] = ureg_DECL_output(ureg
,
4669 TGSI_SEMANTIC_COLOR
,
4670 outputSemanticIndex
[i
]);
4673 assert(!"fragment shader outputs must be POSITION/STENCIL/COLOR");
4674 ret
= PIPE_ERROR_BAD_INPUT
;
4679 else if (procType
== TGSI_PROCESSOR_GEOMETRY
) {
4680 for (i
= 0; i
< numInputs
; i
++) {
4681 t
->inputs
[i
] = ureg_DECL_gs_input(ureg
,
4683 inputSemanticName
[i
],
4684 inputSemanticIndex
[i
]);
4687 for (i
= 0; i
< numOutputs
; i
++) {
4688 t
->outputs
[i
] = ureg_DECL_output(ureg
,
4689 outputSemanticName
[i
],
4690 outputSemanticIndex
[i
]);
4694 assert(procType
== TGSI_PROCESSOR_VERTEX
);
4696 for (i
= 0; i
< numInputs
; i
++) {
4697 t
->inputs
[i
] = ureg_DECL_vs_input(ureg
, i
);
4700 for (i
= 0; i
< numOutputs
; i
++) {
4701 t
->outputs
[i
] = ureg_DECL_output(ureg
,
4702 outputSemanticName
[i
],
4703 outputSemanticIndex
[i
]);
4705 if (passthrough_edgeflags
)
4709 /* Declare address register.
4711 if (program
->num_address_regs
> 0) {
4712 assert(program
->num_address_regs
== 1);
4713 t
->address
[0] = ureg_DECL_address(ureg
);
4716 /* Declare misc input registers
4719 GLbitfield sysInputs
= proginfo
->SystemValuesRead
;
4720 unsigned numSys
= 0;
4721 for (i
= 0; sysInputs
; i
++) {
4722 if (sysInputs
& (1 << i
)) {
4723 unsigned semName
= mesa_sysval_to_semantic
[i
];
4724 t
->systemValues
[i
] = ureg_DECL_system_value(ureg
, numSys
, semName
, 0);
4725 if (semName
== TGSI_SEMANTIC_INSTANCEID
||
4726 semName
== TGSI_SEMANTIC_VERTEXID
) {
4727 /* From Gallium perspective, these system values are always
4728 * integer, and require native integer support. However, if
4729 * native integer is supported on the vertex stage but not the
4730 * pixel stage (e.g, i915g + draw), Mesa will generate IR that
4731 * assumes these system values are floats. To resolve the
4732 * inconsistency, we insert a U2F.
4734 struct st_context
*st
= st_context(ctx
);
4735 struct pipe_screen
*pscreen
= st
->pipe
->screen
;
4736 assert(procType
== TGSI_PROCESSOR_VERTEX
);
4737 assert(pscreen
->get_shader_param(pscreen
, PIPE_SHADER_VERTEX
, PIPE_SHADER_CAP_INTEGERS
));
4738 if (!ctx
->Const
.NativeIntegers
) {
4739 struct ureg_dst temp
= ureg_DECL_local_temporary(t
->ureg
);
4740 ureg_U2F( t
->ureg
, ureg_writemask(temp
, TGSI_WRITEMASK_X
), t
->systemValues
[i
]);
4741 t
->systemValues
[i
] = ureg_scalar(ureg_src(temp
), 0);
4745 sysInputs
&= ~(1 << i
);
4750 if (program
->indirect_addr_temps
) {
4751 /* If temps are accessed with indirect addressing, declare temporaries
4752 * in sequential order. Else, we declare them on demand elsewhere.
4753 * (Note: the number of temporaries is equal to program->next_temp)
4755 for (i
= 0; i
< (unsigned)program
->next_temp
; i
++) {
4756 /* XXX use TGSI_FILE_TEMPORARY_ARRAY when it's supported by ureg */
4757 t
->temps
[i
] = ureg_DECL_local_temporary(t
->ureg
);
4761 /* Emit constants and uniforms. TGSI uses a single index space for these,
4762 * so we put all the translated regs in t->constants.
4764 if (proginfo
->Parameters
) {
4765 t
->constants
= (struct ureg_src
*)
4766 calloc(proginfo
->Parameters
->NumParameters
, sizeof(t
->constants
[0]));
4767 if (t
->constants
== NULL
) {
4768 ret
= PIPE_ERROR_OUT_OF_MEMORY
;
4772 for (i
= 0; i
< proginfo
->Parameters
->NumParameters
; i
++) {
4773 switch (proginfo
->Parameters
->Parameters
[i
].Type
) {
4774 case PROGRAM_ENV_PARAM
:
4775 case PROGRAM_LOCAL_PARAM
:
4776 case PROGRAM_STATE_VAR
:
4777 case PROGRAM_UNIFORM
:
4778 t
->constants
[i
] = ureg_DECL_constant(ureg
, i
);
4781 /* Emit immediates for PROGRAM_CONSTANT only when there's no indirect
4782 * addressing of the const buffer.
4783 * FIXME: Be smarter and recognize param arrays:
4784 * indirect addressing is only valid within the referenced
4787 case PROGRAM_CONSTANT
:
4788 if (program
->indirect_addr_consts
)
4789 t
->constants
[i
] = ureg_DECL_constant(ureg
, i
);
4791 t
->constants
[i
] = emit_immediate(t
,
4792 proginfo
->Parameters
->ParameterValues
[i
],
4793 proginfo
->Parameters
->Parameters
[i
].DataType
,
4802 if (program
->shader_program
) {
4803 unsigned num_ubos
= program
->shader_program
->NumUniformBlocks
;
4805 for (i
= 0; i
< num_ubos
; i
++) {
4806 ureg_DECL_constant2D(t
->ureg
, 0, program
->shader_program
->UniformBlocks
[i
].UniformBufferSize
/ 4, i
+ 1);
4810 /* Emit immediate values.
4812 t
->immediates
= (struct ureg_src
*)
4813 calloc(program
->num_immediates
, sizeof(struct ureg_src
));
4814 if (t
->immediates
== NULL
) {
4815 ret
= PIPE_ERROR_OUT_OF_MEMORY
;
4819 foreach_iter(exec_list_iterator
, iter
, program
->immediates
) {
4820 immediate_storage
*imm
= (immediate_storage
*)iter
.get();
4821 assert(i
< program
->num_immediates
);
4822 t
->immediates
[i
++] = emit_immediate(t
, imm
->values
, imm
->type
, imm
->size
);
4824 assert(i
== program
->num_immediates
);
4826 /* texture samplers */
4827 for (i
= 0; i
< ctx
->Const
.MaxTextureImageUnits
; i
++) {
4828 if (program
->samplers_used
& (1 << i
)) {
4829 t
->samplers
[i
] = ureg_DECL_sampler(ureg
, i
);
4833 /* Emit each instruction in turn:
4835 foreach_iter(exec_list_iterator
, iter
, program
->instructions
) {
4836 set_insn_start(t
, ureg_get_instruction_number(ureg
));
4837 compile_tgsi_instruction(t
, (glsl_to_tgsi_instruction
*)iter
.get(),
4841 /* Fix up all emitted labels:
4843 for (i
= 0; i
< t
->labels_count
; i
++) {
4844 ureg_fixup_label(ureg
, t
->labels
[i
].token
,
4845 t
->insn
[t
->labels
[i
].branch_target
]);
4848 if (program
->shader_program
) {
4849 /* This has to be done last. Any operation the can cause
4850 * prog->ParameterValues to get reallocated (e.g., anything that adds a
4851 * program constant) has to happen before creating this linkage.
4853 for (unsigned i
= 0; i
< MESA_SHADER_TYPES
; i
++) {
4854 if (program
->shader_program
->_LinkedShaders
[i
] == NULL
)
4857 _mesa_associate_uniform_storage(ctx
, program
->shader_program
,
4858 program
->shader_program
->_LinkedShaders
[i
]->Program
->Parameters
);
4867 free(t
->immediates
);
4870 debug_printf("%s: translate error flag set\n", __FUNCTION__
);
4878 /* ----------------------------- End TGSI code ------------------------------ */
4881 * Convert a shader's GLSL IR into a Mesa gl_program, although without
4882 * generating Mesa IR.
4884 static struct gl_program
*
4885 get_mesa_program(struct gl_context
*ctx
,
4886 struct gl_shader_program
*shader_program
,
4887 struct gl_shader
*shader
)
4889 glsl_to_tgsi_visitor
* v
;
4890 struct gl_program
*prog
;
4892 const char *target_string
;
4894 struct gl_shader_compiler_options
*options
=
4895 &ctx
->ShaderCompilerOptions
[_mesa_shader_type_to_index(shader
->Type
)];
4897 switch (shader
->Type
) {
4898 case GL_VERTEX_SHADER
:
4899 target
= GL_VERTEX_PROGRAM_ARB
;
4900 target_string
= "vertex";
4902 case GL_FRAGMENT_SHADER
:
4903 target
= GL_FRAGMENT_PROGRAM_ARB
;
4904 target_string
= "fragment";
4906 case GL_GEOMETRY_SHADER
:
4907 target
= GL_GEOMETRY_PROGRAM_NV
;
4908 target_string
= "geometry";
4911 assert(!"should not be reached");
4915 validate_ir_tree(shader
->ir
);
4917 prog
= ctx
->Driver
.NewProgram(ctx
, target
, shader_program
->Name
);
4920 prog
->Parameters
= _mesa_new_parameter_list();
4921 v
= new glsl_to_tgsi_visitor();
4924 v
->shader_program
= shader_program
;
4925 v
->options
= options
;
4926 v
->glsl_version
= ctx
->Const
.GLSLVersion
;
4927 v
->native_integers
= ctx
->Const
.NativeIntegers
;
4929 _mesa_generate_parameters_list_for_uniforms(shader_program
, shader
,
4932 /* Remove reads from output registers. */
4933 lower_output_reads(shader
->ir
);
4935 /* Emit intermediate IR for main(). */
4936 visit_exec_list(shader
->ir
, v
);
4938 /* Now emit bodies for any functions that were used. */
4940 progress
= GL_FALSE
;
4942 foreach_iter(exec_list_iterator
, iter
, v
->function_signatures
) {
4943 function_entry
*entry
= (function_entry
*)iter
.get();
4945 if (!entry
->bgn_inst
) {
4946 v
->current_function
= entry
;
4948 entry
->bgn_inst
= v
->emit(NULL
, TGSI_OPCODE_BGNSUB
);
4949 entry
->bgn_inst
->function
= entry
;
4951 visit_exec_list(&entry
->sig
->body
, v
);
4953 glsl_to_tgsi_instruction
*last
;
4954 last
= (glsl_to_tgsi_instruction
*)v
->instructions
.get_tail();
4955 if (last
->op
!= TGSI_OPCODE_RET
)
4956 v
->emit(NULL
, TGSI_OPCODE_RET
);
4958 glsl_to_tgsi_instruction
*end
;
4959 end
= v
->emit(NULL
, TGSI_OPCODE_ENDSUB
);
4960 end
->function
= entry
;
4968 /* Print out some information (for debugging purposes) used by the
4969 * optimization passes. */
4970 for (i
=0; i
< v
->next_temp
; i
++) {
4971 int fr
= v
->get_first_temp_read(i
);
4972 int fw
= v
->get_first_temp_write(i
);
4973 int lr
= v
->get_last_temp_read(i
);
4974 int lw
= v
->get_last_temp_write(i
);
4976 printf("Temp %d: FR=%3d FW=%3d LR=%3d LW=%3d\n", i
, fr
, fw
, lr
, lw
);
4981 /* Perform optimizations on the instructions in the glsl_to_tgsi_visitor. */
4983 v
->copy_propagate();
4984 while (v
->eliminate_dead_code_advanced());
4986 /* FIXME: These passes to optimize temporary registers don't work when there
4987 * is indirect addressing of the temporary register space. We need proper
4988 * array support so that we don't have to give up these passes in every
4989 * shader that uses arrays.
4991 if (!v
->indirect_addr_temps
) {
4992 v
->eliminate_dead_code();
4993 v
->merge_registers();
4994 v
->renumber_registers();
4997 /* Write the END instruction. */
4998 v
->emit(NULL
, TGSI_OPCODE_END
);
5000 if (ctx
->Shader
.Flags
& GLSL_DUMP
) {
5002 printf("GLSL IR for linked %s program %d:\n", target_string
,
5003 shader_program
->Name
);
5004 _mesa_print_ir(shader
->ir
, NULL
);
5010 prog
->Instructions
= NULL
;
5011 prog
->NumInstructions
= 0;
5013 do_set_program_inouts(shader
->ir
, prog
, shader
->Type
== GL_FRAGMENT_SHADER
);
5014 count_resources(v
, prog
);
5016 _mesa_reference_program(ctx
, &shader
->Program
, prog
);
5018 /* This has to be done last. Any operation the can cause
5019 * prog->ParameterValues to get reallocated (e.g., anything that adds a
5020 * program constant) has to happen before creating this linkage.
5022 _mesa_associate_uniform_storage(ctx
, shader_program
, prog
->Parameters
);
5023 if (!shader_program
->LinkStatus
) {
5027 struct st_vertex_program
*stvp
;
5028 struct st_fragment_program
*stfp
;
5029 struct st_geometry_program
*stgp
;
5031 switch (shader
->Type
) {
5032 case GL_VERTEX_SHADER
:
5033 stvp
= (struct st_vertex_program
*)prog
;
5034 stvp
->glsl_to_tgsi
= v
;
5036 case GL_FRAGMENT_SHADER
:
5037 stfp
= (struct st_fragment_program
*)prog
;
5038 stfp
->glsl_to_tgsi
= v
;
5040 case GL_GEOMETRY_SHADER
:
5041 stgp
= (struct st_geometry_program
*)prog
;
5042 stgp
->glsl_to_tgsi
= v
;
5045 assert(!"should not be reached");
5055 st_new_shader(struct gl_context
*ctx
, GLuint name
, GLuint type
)
5057 struct gl_shader
*shader
;
5058 assert(type
== GL_FRAGMENT_SHADER
|| type
== GL_VERTEX_SHADER
||
5059 type
== GL_GEOMETRY_SHADER_ARB
);
5060 shader
= rzalloc(NULL
, struct gl_shader
);
5062 shader
->Type
= type
;
5063 shader
->Name
= name
;
5064 _mesa_init_shader(ctx
, shader
);
5069 struct gl_shader_program
*
5070 st_new_shader_program(struct gl_context
*ctx
, GLuint name
)
5072 struct gl_shader_program
*shProg
;
5073 shProg
= rzalloc(NULL
, struct gl_shader_program
);
5075 shProg
->Name
= name
;
5076 _mesa_init_shader_program(ctx
, shProg
);
5083 * Called via ctx->Driver.LinkShader()
5084 * This actually involves converting GLSL IR into an intermediate TGSI-like IR
5085 * with code lowering and other optimizations.
5088 st_link_shader(struct gl_context
*ctx
, struct gl_shader_program
*prog
)
5090 assert(prog
->LinkStatus
);
5092 for (unsigned i
= 0; i
< MESA_SHADER_TYPES
; i
++) {
5093 if (prog
->_LinkedShaders
[i
] == NULL
)
5097 exec_list
*ir
= prog
->_LinkedShaders
[i
]->ir
;
5098 const struct gl_shader_compiler_options
*options
=
5099 &ctx
->ShaderCompilerOptions
[_mesa_shader_type_to_index(prog
->_LinkedShaders
[i
]->Type
)];
5102 unsigned what_to_lower
= MOD_TO_FRACT
| DIV_TO_MUL_RCP
|
5103 EXP_TO_EXP2
| LOG_TO_LOG2
;
5104 if (options
->EmitNoPow
)
5105 what_to_lower
|= POW_TO_EXP2
;
5106 if (!ctx
->Const
.NativeIntegers
)
5107 what_to_lower
|= INT_DIV_TO_MUL_RCP
;
5112 do_mat_op_to_vec(ir
);
5113 lower_instructions(ir
, what_to_lower
);
5115 progress
= do_lower_jumps(ir
, true, true, options
->EmitNoMainReturn
, options
->EmitNoCont
, options
->EmitNoLoops
) || progress
;
5117 progress
= do_common_optimization(ir
, true, true,
5118 options
->MaxUnrollIterations
)
5121 progress
= lower_quadop_vector(ir
, false) || progress
;
5123 if (options
->MaxIfDepth
== 0)
5124 progress
= lower_discard(ir
) || progress
;
5126 progress
= lower_if_to_cond_assign(ir
, options
->MaxIfDepth
) || progress
;
5128 if (options
->EmitNoNoise
)
5129 progress
= lower_noise(ir
) || progress
;
5131 /* If there are forms of indirect addressing that the driver
5132 * cannot handle, perform the lowering pass.
5134 if (options
->EmitNoIndirectInput
|| options
->EmitNoIndirectOutput
5135 || options
->EmitNoIndirectTemp
|| options
->EmitNoIndirectUniform
)
5137 lower_variable_index_to_cond_assign(ir
,
5138 options
->EmitNoIndirectInput
,
5139 options
->EmitNoIndirectOutput
,
5140 options
->EmitNoIndirectTemp
,
5141 options
->EmitNoIndirectUniform
)
5144 progress
= do_vec_index_to_cond_assign(ir
) || progress
;
5146 lower_ubo_reference(prog
->_LinkedShaders
[i
], ir
);
5149 validate_ir_tree(ir
);
5152 for (unsigned i
= 0; i
< MESA_SHADER_TYPES
; i
++) {
5153 struct gl_program
*linked_prog
;
5155 if (prog
->_LinkedShaders
[i
] == NULL
)
5158 linked_prog
= get_mesa_program(ctx
, prog
, prog
->_LinkedShaders
[i
]);
5161 static const GLenum targets
[] = {
5162 GL_VERTEX_PROGRAM_ARB
,
5163 GL_FRAGMENT_PROGRAM_ARB
,
5164 GL_GEOMETRY_PROGRAM_NV
5167 _mesa_reference_program(ctx
, &prog
->_LinkedShaders
[i
]->Program
,
5169 if (!ctx
->Driver
.ProgramStringNotify(ctx
, targets
[i
], linked_prog
)) {
5170 _mesa_reference_program(ctx
, &prog
->_LinkedShaders
[i
]->Program
,
5172 _mesa_reference_program(ctx
, &linked_prog
, NULL
);
5177 _mesa_reference_program(ctx
, &linked_prog
, NULL
);
5184 st_translate_stream_output_info(glsl_to_tgsi_visitor
*glsl_to_tgsi
,
5185 const GLuint outputMapping
[],
5186 struct pipe_stream_output_info
*so
)
5189 struct gl_transform_feedback_info
*info
=
5190 &glsl_to_tgsi
->shader_program
->LinkedTransformFeedback
;
5192 for (i
= 0; i
< info
->NumOutputs
; i
++) {
5193 so
->output
[i
].register_index
=
5194 outputMapping
[info
->Outputs
[i
].OutputRegister
];
5195 so
->output
[i
].start_component
= info
->Outputs
[i
].ComponentOffset
;
5196 so
->output
[i
].num_components
= info
->Outputs
[i
].NumComponents
;
5197 so
->output
[i
].output_buffer
= info
->Outputs
[i
].OutputBuffer
;
5198 so
->output
[i
].dst_offset
= info
->Outputs
[i
].DstOffset
;
5201 for (i
= 0; i
< PIPE_MAX_SO_BUFFERS
; i
++) {
5202 so
->stride
[i
] = info
->BufferStride
[i
];
5204 so
->num_outputs
= info
->NumOutputs
;