2 * Copyright (C) 2005-2007 Brian Paul All Rights Reserved.
3 * Copyright (C) 2008 VMware, Inc. All Rights Reserved.
4 * Copyright © 2010 Intel Corporation
5 * Copyright © 2011 Bryan Cain
7 * Permission is hereby granted, free of charge, to any person obtaining a
8 * copy of this software and associated documentation files (the "Software"),
9 * to deal in the Software without restriction, including without limitation
10 * the rights to use, copy, modify, merge, publish, distribute, sublicense,
11 * and/or sell copies of the Software, and to permit persons to whom the
12 * Software is furnished to do so, subject to the following conditions:
14 * The above copyright notice and this permission notice (including the next
15 * paragraph) shall be included in all copies or substantial portions of the
18 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
19 * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
20 * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL
21 * THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
22 * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING
23 * FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER
24 * DEALINGS IN THE SOFTWARE.
28 * \file glsl_to_tgsi.cpp
30 * Translate GLSL IR to TGSI.
34 #include "main/compiler.h"
36 #include "ir_visitor.h"
37 #include "ir_print_visitor.h"
38 #include "ir_expression_flattening.h"
39 #include "glsl_types.h"
40 #include "glsl_parser_extras.h"
41 #include "../glsl/program.h"
42 #include "ir_optimization.h"
45 #include "main/mtypes.h"
46 #include "main/shaderobj.h"
47 #include "program/hash_table.h"
50 #include "main/shaderapi.h"
51 #include "main/uniforms.h"
52 #include "program/prog_instruction.h"
53 #include "program/prog_optimize.h"
54 #include "program/prog_print.h"
55 #include "program/program.h"
56 #include "program/prog_parameter.h"
57 #include "program/sampler.h"
59 #include "pipe/p_compiler.h"
60 #include "pipe/p_context.h"
61 #include "pipe/p_screen.h"
62 #include "pipe/p_shader_tokens.h"
63 #include "pipe/p_state.h"
64 #include "util/u_math.h"
65 #include "tgsi/tgsi_ureg.h"
66 #include "tgsi/tgsi_info.h"
67 #include "st_context.h"
68 #include "st_program.h"
69 #include "st_glsl_to_tgsi.h"
70 #include "st_mesa_to_tgsi.h"
73 #define PROGRAM_IMMEDIATE PROGRAM_FILE_MAX
74 #define PROGRAM_ANY_CONST ((1 << PROGRAM_LOCAL_PARAM) | \
75 (1 << PROGRAM_ENV_PARAM) | \
76 (1 << PROGRAM_STATE_VAR) | \
77 (1 << PROGRAM_NAMED_PARAM) | \
78 (1 << PROGRAM_CONSTANT) | \
79 (1 << PROGRAM_UNIFORM))
82 * Maximum number of temporary registers.
84 * It is too big for stack allocated arrays -- it will cause stack overflow on
85 * Windows and likely Mac OS X.
87 #define MAX_TEMPS 4096
89 /* will be 4 for GLSL 4.00 */
90 #define MAX_GLSL_TEXTURE_OFFSET 1
95 static int swizzle_for_size(int size
);
98 * This struct is a corresponding struct to TGSI ureg_src.
102 st_src_reg(gl_register_file file
, int index
, const glsl_type
*type
)
106 if (type
&& (type
->is_scalar() || type
->is_vector() || type
->is_matrix()))
107 this->swizzle
= swizzle_for_size(type
->vector_elements
);
109 this->swizzle
= SWIZZLE_XYZW
;
111 this->type
= type
? type
->base_type
: GLSL_TYPE_ERROR
;
112 this->reladdr
= NULL
;
115 st_src_reg(gl_register_file file
, int index
, int type
)
120 this->swizzle
= SWIZZLE_XYZW
;
122 this->reladdr
= NULL
;
127 this->type
= GLSL_TYPE_ERROR
;
128 this->file
= PROGRAM_UNDEFINED
;
132 this->reladdr
= NULL
;
135 explicit st_src_reg(st_dst_reg reg
);
137 gl_register_file file
; /**< PROGRAM_* from Mesa */
138 int index
; /**< temporary index, VERT_ATTRIB_*, FRAG_ATTRIB_*, etc. */
139 GLuint swizzle
; /**< SWIZZLE_XYZWONEZERO swizzles from Mesa. */
140 int negate
; /**< NEGATE_XYZW mask from mesa */
141 int type
; /** GLSL_TYPE_* from GLSL IR (enum glsl_base_type) */
142 /** Register index should be offset by the integer in this reg. */
148 st_dst_reg(gl_register_file file
, int writemask
, int type
)
152 this->writemask
= writemask
;
153 this->cond_mask
= COND_TR
;
154 this->reladdr
= NULL
;
160 this->type
= GLSL_TYPE_ERROR
;
161 this->file
= PROGRAM_UNDEFINED
;
164 this->cond_mask
= COND_TR
;
165 this->reladdr
= NULL
;
168 explicit st_dst_reg(st_src_reg reg
);
170 gl_register_file file
; /**< PROGRAM_* from Mesa */
171 int index
; /**< temporary index, VERT_ATTRIB_*, FRAG_ATTRIB_*, etc. */
172 int writemask
; /**< Bitfield of WRITEMASK_[XYZW] */
174 int type
; /** GLSL_TYPE_* from GLSL IR (enum glsl_base_type) */
175 /** Register index should be offset by the integer in this reg. */
179 st_src_reg::st_src_reg(st_dst_reg reg
)
181 this->type
= reg
.type
;
182 this->file
= reg
.file
;
183 this->index
= reg
.index
;
184 this->swizzle
= SWIZZLE_XYZW
;
186 this->reladdr
= reg
.reladdr
;
189 st_dst_reg::st_dst_reg(st_src_reg reg
)
191 this->type
= reg
.type
;
192 this->file
= reg
.file
;
193 this->index
= reg
.index
;
194 this->writemask
= WRITEMASK_XYZW
;
195 this->cond_mask
= COND_TR
;
196 this->reladdr
= reg
.reladdr
;
199 class glsl_to_tgsi_instruction
: public exec_node
{
201 /* Callers of this ralloc-based new need not call delete. It's
202 * easier to just ralloc_free 'ctx' (or any of its ancestors). */
203 static void* operator new(size_t size
, void *ctx
)
207 node
= rzalloc_size(ctx
, size
);
208 assert(node
!= NULL
);
216 /** Pointer to the ir source this tree came from for debugging */
218 GLboolean cond_update
;
220 int sampler
; /**< sampler index */
221 int tex_target
; /**< One of TEXTURE_*_INDEX */
222 GLboolean tex_shadow
;
223 struct tgsi_texture_offset tex_offsets
[MAX_GLSL_TEXTURE_OFFSET
];
224 unsigned tex_offset_num_offset
;
225 int dead_mask
; /**< Used in dead code elimination */
227 class function_entry
*function
; /* Set on TGSI_OPCODE_CAL or TGSI_OPCODE_BGNSUB */
230 class variable_storage
: public exec_node
{
232 variable_storage(ir_variable
*var
, gl_register_file file
, int index
)
233 : file(file
), index(index
), var(var
)
238 gl_register_file file
;
240 ir_variable
*var
; /* variable that maps to this, if any */
243 class immediate_storage
: public exec_node
{
245 immediate_storage(gl_constant_value
*values
, int size
, int type
)
247 memcpy(this->values
, values
, size
* sizeof(gl_constant_value
));
252 gl_constant_value values
[4];
253 int size
; /**< Number of components (1-4) */
254 int type
; /**< GL_FLOAT, GL_INT, GL_BOOL, or GL_UNSIGNED_INT */
257 class function_entry
: public exec_node
{
259 ir_function_signature
*sig
;
262 * identifier of this function signature used by the program.
264 * At the point that TGSI instructions for function calls are
265 * generated, we don't know the address of the first instruction of
266 * the function body. So we make the BranchTarget that is called a
267 * small integer and rewrite them during set_branchtargets().
272 * Pointer to first instruction of the function body.
274 * Set during function body emits after main() is processed.
276 glsl_to_tgsi_instruction
*bgn_inst
;
279 * Index of the first instruction of the function body in actual TGSI.
281 * Set after conversion from glsl_to_tgsi_instruction to TGSI.
285 /** Storage for the return value. */
286 st_src_reg return_reg
;
289 class glsl_to_tgsi_visitor
: public ir_visitor
{
291 glsl_to_tgsi_visitor();
292 ~glsl_to_tgsi_visitor();
294 function_entry
*current_function
;
296 struct gl_context
*ctx
;
297 struct gl_program
*prog
;
298 struct gl_shader_program
*shader_program
;
299 struct gl_shader_compiler_options
*options
;
303 int num_address_regs
;
305 bool indirect_addr_temps
;
306 bool indirect_addr_consts
;
309 bool native_integers
;
311 variable_storage
*find_variable_storage(ir_variable
*var
);
313 int add_constant(gl_register_file file
, gl_constant_value values
[4],
314 int size
, int datatype
, GLuint
*swizzle_out
);
316 function_entry
*get_function_signature(ir_function_signature
*sig
);
318 st_src_reg
get_temp(const glsl_type
*type
);
319 void reladdr_to_temp(ir_instruction
*ir
, st_src_reg
*reg
, int *num_reladdr
);
321 st_src_reg
st_src_reg_for_float(float val
);
322 st_src_reg
st_src_reg_for_int(int val
);
323 st_src_reg
st_src_reg_for_type(int type
, int val
);
326 * \name Visit methods
328 * As typical for the visitor pattern, there must be one \c visit method for
329 * each concrete subclass of \c ir_instruction. Virtual base classes within
330 * the hierarchy should not have \c visit methods.
333 virtual void visit(ir_variable
*);
334 virtual void visit(ir_loop
*);
335 virtual void visit(ir_loop_jump
*);
336 virtual void visit(ir_function_signature
*);
337 virtual void visit(ir_function
*);
338 virtual void visit(ir_expression
*);
339 virtual void visit(ir_swizzle
*);
340 virtual void visit(ir_dereference_variable
*);
341 virtual void visit(ir_dereference_array
*);
342 virtual void visit(ir_dereference_record
*);
343 virtual void visit(ir_assignment
*);
344 virtual void visit(ir_constant
*);
345 virtual void visit(ir_call
*);
346 virtual void visit(ir_return
*);
347 virtual void visit(ir_discard
*);
348 virtual void visit(ir_texture
*);
349 virtual void visit(ir_if
*);
354 /** List of variable_storage */
357 /** List of immediate_storage */
358 exec_list immediates
;
361 /** List of function_entry */
362 exec_list function_signatures
;
363 int next_signature_id
;
365 /** List of glsl_to_tgsi_instruction */
366 exec_list instructions
;
368 glsl_to_tgsi_instruction
*emit(ir_instruction
*ir
, unsigned op
);
370 glsl_to_tgsi_instruction
*emit(ir_instruction
*ir
, unsigned op
,
371 st_dst_reg dst
, st_src_reg src0
);
373 glsl_to_tgsi_instruction
*emit(ir_instruction
*ir
, unsigned op
,
374 st_dst_reg dst
, st_src_reg src0
, st_src_reg src1
);
376 glsl_to_tgsi_instruction
*emit(ir_instruction
*ir
, unsigned op
,
378 st_src_reg src0
, st_src_reg src1
, st_src_reg src2
);
380 unsigned get_opcode(ir_instruction
*ir
, unsigned op
,
382 st_src_reg src0
, st_src_reg src1
);
385 * Emit the correct dot-product instruction for the type of arguments
387 glsl_to_tgsi_instruction
*emit_dp(ir_instruction
*ir
,
393 void emit_scalar(ir_instruction
*ir
, unsigned op
,
394 st_dst_reg dst
, st_src_reg src0
);
396 void emit_scalar(ir_instruction
*ir
, unsigned op
,
397 st_dst_reg dst
, st_src_reg src0
, st_src_reg src1
);
399 void try_emit_float_set(ir_instruction
*ir
, unsigned op
, st_dst_reg dst
);
401 void emit_arl(ir_instruction
*ir
, st_dst_reg dst
, st_src_reg src0
);
403 void emit_scs(ir_instruction
*ir
, unsigned op
,
404 st_dst_reg dst
, const st_src_reg
&src
);
406 bool try_emit_mad(ir_expression
*ir
,
408 bool try_emit_mad_for_and_not(ir_expression
*ir
,
410 bool try_emit_sat(ir_expression
*ir
);
412 void emit_swz(ir_expression
*ir
);
414 bool process_move_condition(ir_rvalue
*ir
);
416 void remove_output_reads(gl_register_file type
);
417 void simplify_cmp(void);
419 void rename_temp_register(int index
, int new_index
);
420 int get_first_temp_read(int index
);
421 int get_first_temp_write(int index
);
422 int get_last_temp_read(int index
);
423 int get_last_temp_write(int index
);
425 void copy_propagate(void);
426 void eliminate_dead_code(void);
427 int eliminate_dead_code_advanced(void);
428 void merge_registers(void);
429 void renumber_registers(void);
434 static st_src_reg undef_src
= st_src_reg(PROGRAM_UNDEFINED
, 0, GLSL_TYPE_ERROR
);
436 static st_dst_reg undef_dst
= st_dst_reg(PROGRAM_UNDEFINED
, SWIZZLE_NOOP
, GLSL_TYPE_ERROR
);
438 static st_dst_reg address_reg
= st_dst_reg(PROGRAM_ADDRESS
, WRITEMASK_X
, GLSL_TYPE_FLOAT
);
441 fail_link(struct gl_shader_program
*prog
, const char *fmt
, ...) PRINTFLIKE(2, 3);
444 fail_link(struct gl_shader_program
*prog
, const char *fmt
, ...)
448 ralloc_vasprintf_append(&prog
->InfoLog
, fmt
, args
);
451 prog
->LinkStatus
= GL_FALSE
;
455 swizzle_for_size(int size
)
457 int size_swizzles
[4] = {
458 MAKE_SWIZZLE4(SWIZZLE_X
, SWIZZLE_X
, SWIZZLE_X
, SWIZZLE_X
),
459 MAKE_SWIZZLE4(SWIZZLE_X
, SWIZZLE_Y
, SWIZZLE_Y
, SWIZZLE_Y
),
460 MAKE_SWIZZLE4(SWIZZLE_X
, SWIZZLE_Y
, SWIZZLE_Z
, SWIZZLE_Z
),
461 MAKE_SWIZZLE4(SWIZZLE_X
, SWIZZLE_Y
, SWIZZLE_Z
, SWIZZLE_W
),
464 assert((size
>= 1) && (size
<= 4));
465 return size_swizzles
[size
- 1];
469 is_tex_instruction(unsigned opcode
)
471 const tgsi_opcode_info
* info
= tgsi_get_opcode_info(opcode
);
476 num_inst_dst_regs(unsigned opcode
)
478 const tgsi_opcode_info
* info
= tgsi_get_opcode_info(opcode
);
479 return info
->num_dst
;
483 num_inst_src_regs(unsigned opcode
)
485 const tgsi_opcode_info
* info
= tgsi_get_opcode_info(opcode
);
486 return info
->is_tex
? info
->num_src
- 1 : info
->num_src
;
489 glsl_to_tgsi_instruction
*
490 glsl_to_tgsi_visitor::emit(ir_instruction
*ir
, unsigned op
,
492 st_src_reg src0
, st_src_reg src1
, st_src_reg src2
)
494 glsl_to_tgsi_instruction
*inst
= new(mem_ctx
) glsl_to_tgsi_instruction();
495 int num_reladdr
= 0, i
;
497 op
= get_opcode(ir
, op
, dst
, src0
, src1
);
499 /* If we have to do relative addressing, we want to load the ARL
500 * reg directly for one of the regs, and preload the other reladdr
501 * sources into temps.
503 num_reladdr
+= dst
.reladdr
!= NULL
;
504 num_reladdr
+= src0
.reladdr
!= NULL
;
505 num_reladdr
+= src1
.reladdr
!= NULL
;
506 num_reladdr
+= src2
.reladdr
!= NULL
;
508 reladdr_to_temp(ir
, &src2
, &num_reladdr
);
509 reladdr_to_temp(ir
, &src1
, &num_reladdr
);
510 reladdr_to_temp(ir
, &src0
, &num_reladdr
);
513 emit_arl(ir
, address_reg
, *dst
.reladdr
);
516 assert(num_reladdr
== 0);
526 inst
->function
= NULL
;
528 if (op
== TGSI_OPCODE_ARL
|| op
== TGSI_OPCODE_UARL
)
529 this->num_address_regs
= 1;
531 /* Update indirect addressing status used by TGSI */
534 case PROGRAM_TEMPORARY
:
535 this->indirect_addr_temps
= true;
537 case PROGRAM_LOCAL_PARAM
:
538 case PROGRAM_ENV_PARAM
:
539 case PROGRAM_STATE_VAR
:
540 case PROGRAM_NAMED_PARAM
:
541 case PROGRAM_CONSTANT
:
542 case PROGRAM_UNIFORM
:
543 this->indirect_addr_consts
= true;
545 case PROGRAM_IMMEDIATE
:
546 assert(!"immediates should not have indirect addressing");
553 for (i
=0; i
<3; i
++) {
554 if(inst
->src
[i
].reladdr
) {
555 switch(inst
->src
[i
].file
) {
556 case PROGRAM_TEMPORARY
:
557 this->indirect_addr_temps
= true;
559 case PROGRAM_LOCAL_PARAM
:
560 case PROGRAM_ENV_PARAM
:
561 case PROGRAM_STATE_VAR
:
562 case PROGRAM_NAMED_PARAM
:
563 case PROGRAM_CONSTANT
:
564 case PROGRAM_UNIFORM
:
565 this->indirect_addr_consts
= true;
567 case PROGRAM_IMMEDIATE
:
568 assert(!"immediates should not have indirect addressing");
577 this->instructions
.push_tail(inst
);
580 try_emit_float_set(ir
, op
, dst
);
586 glsl_to_tgsi_instruction
*
587 glsl_to_tgsi_visitor::emit(ir_instruction
*ir
, unsigned op
,
588 st_dst_reg dst
, st_src_reg src0
, st_src_reg src1
)
590 return emit(ir
, op
, dst
, src0
, src1
, undef_src
);
593 glsl_to_tgsi_instruction
*
594 glsl_to_tgsi_visitor::emit(ir_instruction
*ir
, unsigned op
,
595 st_dst_reg dst
, st_src_reg src0
)
597 assert(dst
.writemask
!= 0);
598 return emit(ir
, op
, dst
, src0
, undef_src
, undef_src
);
601 glsl_to_tgsi_instruction
*
602 glsl_to_tgsi_visitor::emit(ir_instruction
*ir
, unsigned op
)
604 return emit(ir
, op
, undef_dst
, undef_src
, undef_src
, undef_src
);
608 * Emits the code to convert the result of float SET instructions to integers.
611 glsl_to_tgsi_visitor::try_emit_float_set(ir_instruction
*ir
, unsigned op
,
614 if ((op
== TGSI_OPCODE_SEQ
||
615 op
== TGSI_OPCODE_SNE
||
616 op
== TGSI_OPCODE_SGE
||
617 op
== TGSI_OPCODE_SLT
))
619 st_src_reg src
= st_src_reg(dst
);
620 src
.negate
= ~src
.negate
;
621 dst
.type
= GLSL_TYPE_FLOAT
;
622 emit(ir
, TGSI_OPCODE_F2I
, dst
, src
);
627 * Determines whether to use an integer, unsigned integer, or float opcode
628 * based on the operands and input opcode, then emits the result.
631 glsl_to_tgsi_visitor::get_opcode(ir_instruction
*ir
, unsigned op
,
633 st_src_reg src0
, st_src_reg src1
)
635 int type
= GLSL_TYPE_FLOAT
;
637 if (src0
.type
== GLSL_TYPE_FLOAT
|| src1
.type
== GLSL_TYPE_FLOAT
)
638 type
= GLSL_TYPE_FLOAT
;
639 else if (native_integers
)
640 type
= src0
.type
== GLSL_TYPE_BOOL
? GLSL_TYPE_INT
: src0
.type
;
642 #define case4(c, f, i, u) \
643 case TGSI_OPCODE_##c: \
644 if (type == GLSL_TYPE_INT) op = TGSI_OPCODE_##i; \
645 else if (type == GLSL_TYPE_UINT) op = TGSI_OPCODE_##u; \
646 else op = TGSI_OPCODE_##f; \
648 #define case3(f, i, u) case4(f, f, i, u)
649 #define case2fi(f, i) case4(f, f, i, i)
650 #define case2iu(i, u) case4(i, LAST, i, u)
656 case3(DIV
, IDIV
, UDIV
);
657 case3(MAX
, IMAX
, UMAX
);
658 case3(MIN
, IMIN
, UMIN
);
663 case3(SGE
, ISGE
, USGE
);
664 case3(SLT
, ISLT
, USLT
);
671 assert(op
!= TGSI_OPCODE_LAST
);
675 glsl_to_tgsi_instruction
*
676 glsl_to_tgsi_visitor::emit_dp(ir_instruction
*ir
,
677 st_dst_reg dst
, st_src_reg src0
, st_src_reg src1
,
680 static const unsigned dot_opcodes
[] = {
681 TGSI_OPCODE_DP2
, TGSI_OPCODE_DP3
, TGSI_OPCODE_DP4
684 return emit(ir
, dot_opcodes
[elements
- 2], dst
, src0
, src1
);
688 * Emits TGSI scalar opcodes to produce unique answers across channels.
690 * Some TGSI opcodes are scalar-only, like ARB_fp/vp. The src X
691 * channel determines the result across all channels. So to do a vec4
692 * of this operation, we want to emit a scalar per source channel used
693 * to produce dest channels.
696 glsl_to_tgsi_visitor::emit_scalar(ir_instruction
*ir
, unsigned op
,
698 st_src_reg orig_src0
, st_src_reg orig_src1
)
701 int done_mask
= ~dst
.writemask
;
703 /* TGSI RCP is a scalar operation splatting results to all channels,
704 * like ARB_fp/vp. So emit as many RCPs as necessary to cover our
707 for (i
= 0; i
< 4; i
++) {
708 GLuint this_mask
= (1 << i
);
709 glsl_to_tgsi_instruction
*inst
;
710 st_src_reg src0
= orig_src0
;
711 st_src_reg src1
= orig_src1
;
713 if (done_mask
& this_mask
)
716 GLuint src0_swiz
= GET_SWZ(src0
.swizzle
, i
);
717 GLuint src1_swiz
= GET_SWZ(src1
.swizzle
, i
);
718 for (j
= i
+ 1; j
< 4; j
++) {
719 /* If there is another enabled component in the destination that is
720 * derived from the same inputs, generate its value on this pass as
723 if (!(done_mask
& (1 << j
)) &&
724 GET_SWZ(src0
.swizzle
, j
) == src0_swiz
&&
725 GET_SWZ(src1
.swizzle
, j
) == src1_swiz
) {
726 this_mask
|= (1 << j
);
729 src0
.swizzle
= MAKE_SWIZZLE4(src0_swiz
, src0_swiz
,
730 src0_swiz
, src0_swiz
);
731 src1
.swizzle
= MAKE_SWIZZLE4(src1_swiz
, src1_swiz
,
732 src1_swiz
, src1_swiz
);
734 inst
= emit(ir
, op
, dst
, src0
, src1
);
735 inst
->dst
.writemask
= this_mask
;
736 done_mask
|= this_mask
;
741 glsl_to_tgsi_visitor::emit_scalar(ir_instruction
*ir
, unsigned op
,
742 st_dst_reg dst
, st_src_reg src0
)
744 st_src_reg undef
= undef_src
;
746 undef
.swizzle
= SWIZZLE_XXXX
;
748 emit_scalar(ir
, op
, dst
, src0
, undef
);
752 glsl_to_tgsi_visitor::emit_arl(ir_instruction
*ir
,
753 st_dst_reg dst
, st_src_reg src0
)
755 int op
= TGSI_OPCODE_ARL
;
757 if (src0
.type
== GLSL_TYPE_INT
|| src0
.type
== GLSL_TYPE_UINT
)
758 op
= TGSI_OPCODE_UARL
;
760 emit(NULL
, op
, dst
, src0
);
764 * Emit an TGSI_OPCODE_SCS instruction
766 * The \c SCS opcode functions a bit differently than the other TGSI opcodes.
767 * Instead of splatting its result across all four components of the
768 * destination, it writes one value to the \c x component and another value to
769 * the \c y component.
771 * \param ir IR instruction being processed
772 * \param op Either \c TGSI_OPCODE_SIN or \c TGSI_OPCODE_COS depending
773 * on which value is desired.
774 * \param dst Destination register
775 * \param src Source register
778 glsl_to_tgsi_visitor::emit_scs(ir_instruction
*ir
, unsigned op
,
780 const st_src_reg
&src
)
782 /* Vertex programs cannot use the SCS opcode.
784 if (this->prog
->Target
== GL_VERTEX_PROGRAM_ARB
) {
785 emit_scalar(ir
, op
, dst
, src
);
789 const unsigned component
= (op
== TGSI_OPCODE_SIN
) ? 0 : 1;
790 const unsigned scs_mask
= (1U << component
);
791 int done_mask
= ~dst
.writemask
;
794 assert(op
== TGSI_OPCODE_SIN
|| op
== TGSI_OPCODE_COS
);
796 /* If there are compnents in the destination that differ from the component
797 * that will be written by the SCS instrution, we'll need a temporary.
799 if (scs_mask
!= unsigned(dst
.writemask
)) {
800 tmp
= get_temp(glsl_type::vec4_type
);
803 for (unsigned i
= 0; i
< 4; i
++) {
804 unsigned this_mask
= (1U << i
);
805 st_src_reg src0
= src
;
807 if ((done_mask
& this_mask
) != 0)
810 /* The source swizzle specified which component of the source generates
811 * sine / cosine for the current component in the destination. The SCS
812 * instruction requires that this value be swizzle to the X component.
813 * Replace the current swizzle with a swizzle that puts the source in
816 unsigned src0_swiz
= GET_SWZ(src
.swizzle
, i
);
818 src0
.swizzle
= MAKE_SWIZZLE4(src0_swiz
, src0_swiz
,
819 src0_swiz
, src0_swiz
);
820 for (unsigned j
= i
+ 1; j
< 4; j
++) {
821 /* If there is another enabled component in the destination that is
822 * derived from the same inputs, generate its value on this pass as
825 if (!(done_mask
& (1 << j
)) &&
826 GET_SWZ(src0
.swizzle
, j
) == src0_swiz
) {
827 this_mask
|= (1 << j
);
831 if (this_mask
!= scs_mask
) {
832 glsl_to_tgsi_instruction
*inst
;
833 st_dst_reg tmp_dst
= st_dst_reg(tmp
);
835 /* Emit the SCS instruction.
837 inst
= emit(ir
, TGSI_OPCODE_SCS
, tmp_dst
, src0
);
838 inst
->dst
.writemask
= scs_mask
;
840 /* Move the result of the SCS instruction to the desired location in
843 tmp
.swizzle
= MAKE_SWIZZLE4(component
, component
,
844 component
, component
);
845 inst
= emit(ir
, TGSI_OPCODE_SCS
, dst
, tmp
);
846 inst
->dst
.writemask
= this_mask
;
848 /* Emit the SCS instruction to write directly to the destination.
850 glsl_to_tgsi_instruction
*inst
= emit(ir
, TGSI_OPCODE_SCS
, dst
, src0
);
851 inst
->dst
.writemask
= scs_mask
;
854 done_mask
|= this_mask
;
859 glsl_to_tgsi_visitor::add_constant(gl_register_file file
,
860 gl_constant_value values
[4], int size
, int datatype
,
863 if (file
== PROGRAM_CONSTANT
) {
864 return _mesa_add_typed_unnamed_constant(this->prog
->Parameters
, values
,
865 size
, datatype
, swizzle_out
);
868 immediate_storage
*entry
;
869 assert(file
== PROGRAM_IMMEDIATE
);
871 /* Search immediate storage to see if we already have an identical
872 * immediate that we can use instead of adding a duplicate entry.
874 foreach_iter(exec_list_iterator
, iter
, this->immediates
) {
875 entry
= (immediate_storage
*)iter
.get();
877 if (entry
->size
== size
&&
878 entry
->type
== datatype
&&
879 !memcmp(entry
->values
, values
, size
* sizeof(gl_constant_value
))) {
885 /* Add this immediate to the list. */
886 entry
= new(mem_ctx
) immediate_storage(values
, size
, datatype
);
887 this->immediates
.push_tail(entry
);
888 this->num_immediates
++;
894 glsl_to_tgsi_visitor::st_src_reg_for_float(float val
)
896 st_src_reg
src(PROGRAM_IMMEDIATE
, -1, GLSL_TYPE_FLOAT
);
897 union gl_constant_value uval
;
900 src
.index
= add_constant(src
.file
, &uval
, 1, GL_FLOAT
, &src
.swizzle
);
906 glsl_to_tgsi_visitor::st_src_reg_for_int(int val
)
908 st_src_reg
src(PROGRAM_IMMEDIATE
, -1, GLSL_TYPE_INT
);
909 union gl_constant_value uval
;
911 assert(native_integers
);
914 src
.index
= add_constant(src
.file
, &uval
, 1, GL_INT
, &src
.swizzle
);
920 glsl_to_tgsi_visitor::st_src_reg_for_type(int type
, int val
)
923 return type
== GLSL_TYPE_FLOAT
? st_src_reg_for_float(val
) :
924 st_src_reg_for_int(val
);
926 return st_src_reg_for_float(val
);
930 type_size(const struct glsl_type
*type
)
935 switch (type
->base_type
) {
938 case GLSL_TYPE_FLOAT
:
940 if (type
->is_matrix()) {
941 return type
->matrix_columns
;
943 /* Regardless of size of vector, it gets a vec4. This is bad
944 * packing for things like floats, but otherwise arrays become a
945 * mess. Hopefully a later pass over the code can pack scalars
946 * down if appropriate.
950 case GLSL_TYPE_ARRAY
:
951 assert(type
->length
> 0);
952 return type_size(type
->fields
.array
) * type
->length
;
953 case GLSL_TYPE_STRUCT
:
955 for (i
= 0; i
< type
->length
; i
++) {
956 size
+= type_size(type
->fields
.structure
[i
].type
);
959 case GLSL_TYPE_SAMPLER
:
960 /* Samplers take up one slot in UNIFORMS[], but they're baked in
971 * In the initial pass of codegen, we assign temporary numbers to
972 * intermediate results. (not SSA -- variable assignments will reuse
976 glsl_to_tgsi_visitor::get_temp(const glsl_type
*type
)
980 src
.type
= native_integers
? type
->base_type
: GLSL_TYPE_FLOAT
;
981 src
.file
= PROGRAM_TEMPORARY
;
982 src
.index
= next_temp
;
984 next_temp
+= type_size(type
);
986 if (type
->is_array() || type
->is_record()) {
987 src
.swizzle
= SWIZZLE_NOOP
;
989 src
.swizzle
= swizzle_for_size(type
->vector_elements
);
997 glsl_to_tgsi_visitor::find_variable_storage(ir_variable
*var
)
1000 variable_storage
*entry
;
1002 foreach_iter(exec_list_iterator
, iter
, this->variables
) {
1003 entry
= (variable_storage
*)iter
.get();
1005 if (entry
->var
== var
)
1013 glsl_to_tgsi_visitor::visit(ir_variable
*ir
)
1015 if (strcmp(ir
->name
, "gl_FragCoord") == 0) {
1016 struct gl_fragment_program
*fp
= (struct gl_fragment_program
*)this->prog
;
1018 fp
->OriginUpperLeft
= ir
->origin_upper_left
;
1019 fp
->PixelCenterInteger
= ir
->pixel_center_integer
;
1021 } else if (strcmp(ir
->name
, "gl_FragDepth") == 0) {
1022 struct gl_fragment_program
*fp
= (struct gl_fragment_program
*)this->prog
;
1023 switch (ir
->depth_layout
) {
1024 case ir_depth_layout_none
:
1025 fp
->FragDepthLayout
= FRAG_DEPTH_LAYOUT_NONE
;
1027 case ir_depth_layout_any
:
1028 fp
->FragDepthLayout
= FRAG_DEPTH_LAYOUT_ANY
;
1030 case ir_depth_layout_greater
:
1031 fp
->FragDepthLayout
= FRAG_DEPTH_LAYOUT_GREATER
;
1033 case ir_depth_layout_less
:
1034 fp
->FragDepthLayout
= FRAG_DEPTH_LAYOUT_LESS
;
1036 case ir_depth_layout_unchanged
:
1037 fp
->FragDepthLayout
= FRAG_DEPTH_LAYOUT_UNCHANGED
;
1045 if (ir
->mode
== ir_var_uniform
&& strncmp(ir
->name
, "gl_", 3) == 0) {
1047 const ir_state_slot
*const slots
= ir
->state_slots
;
1048 assert(ir
->state_slots
!= NULL
);
1050 /* Check if this statevar's setup in the STATE file exactly
1051 * matches how we'll want to reference it as a
1052 * struct/array/whatever. If not, then we need to move it into
1053 * temporary storage and hope that it'll get copy-propagated
1056 for (i
= 0; i
< ir
->num_state_slots
; i
++) {
1057 if (slots
[i
].swizzle
!= SWIZZLE_XYZW
) {
1062 variable_storage
*storage
;
1064 if (i
== ir
->num_state_slots
) {
1065 /* We'll set the index later. */
1066 storage
= new(mem_ctx
) variable_storage(ir
, PROGRAM_STATE_VAR
, -1);
1067 this->variables
.push_tail(storage
);
1071 /* The variable_storage constructor allocates slots based on the size
1072 * of the type. However, this had better match the number of state
1073 * elements that we're going to copy into the new temporary.
1075 assert((int) ir
->num_state_slots
== type_size(ir
->type
));
1077 storage
= new(mem_ctx
) variable_storage(ir
, PROGRAM_TEMPORARY
,
1079 this->variables
.push_tail(storage
);
1080 this->next_temp
+= type_size(ir
->type
);
1082 dst
= st_dst_reg(st_src_reg(PROGRAM_TEMPORARY
, storage
->index
,
1083 native_integers
? ir
->type
->base_type
: GLSL_TYPE_FLOAT
));
1087 for (unsigned int i
= 0; i
< ir
->num_state_slots
; i
++) {
1088 int index
= _mesa_add_state_reference(this->prog
->Parameters
,
1089 (gl_state_index
*)slots
[i
].tokens
);
1091 if (storage
->file
== PROGRAM_STATE_VAR
) {
1092 if (storage
->index
== -1) {
1093 storage
->index
= index
;
1095 assert(index
== storage
->index
+ (int)i
);
1098 st_src_reg
src(PROGRAM_STATE_VAR
, index
,
1099 native_integers
? ir
->type
->base_type
: GLSL_TYPE_FLOAT
);
1100 src
.swizzle
= slots
[i
].swizzle
;
1101 emit(ir
, TGSI_OPCODE_MOV
, dst
, src
);
1102 /* even a float takes up a whole vec4 reg in a struct/array. */
1107 if (storage
->file
== PROGRAM_TEMPORARY
&&
1108 dst
.index
!= storage
->index
+ (int) ir
->num_state_slots
) {
1109 fail_link(this->shader_program
,
1110 "failed to load builtin uniform `%s' (%d/%d regs loaded)\n",
1111 ir
->name
, dst
.index
- storage
->index
,
1112 type_size(ir
->type
));
1118 glsl_to_tgsi_visitor::visit(ir_loop
*ir
)
1120 ir_dereference_variable
*counter
= NULL
;
1122 if (ir
->counter
!= NULL
)
1123 counter
= new(ir
) ir_dereference_variable(ir
->counter
);
1125 if (ir
->from
!= NULL
) {
1126 assert(ir
->counter
!= NULL
);
1128 ir_assignment
*a
= new(ir
) ir_assignment(counter
, ir
->from
, NULL
);
1134 emit(NULL
, TGSI_OPCODE_BGNLOOP
);
1138 new(ir
) ir_expression(ir
->cmp
, glsl_type::bool_type
,
1140 ir_if
*if_stmt
= new(ir
) ir_if(e
);
1142 ir_loop_jump
*brk
= new(ir
) ir_loop_jump(ir_loop_jump::jump_break
);
1144 if_stmt
->then_instructions
.push_tail(brk
);
1146 if_stmt
->accept(this);
1153 visit_exec_list(&ir
->body_instructions
, this);
1155 if (ir
->increment
) {
1157 new(ir
) ir_expression(ir_binop_add
, counter
->type
,
1158 counter
, ir
->increment
);
1160 ir_assignment
*a
= new(ir
) ir_assignment(counter
, e
, NULL
);
1167 emit(NULL
, TGSI_OPCODE_ENDLOOP
);
1171 glsl_to_tgsi_visitor::visit(ir_loop_jump
*ir
)
1174 case ir_loop_jump::jump_break
:
1175 emit(NULL
, TGSI_OPCODE_BRK
);
1177 case ir_loop_jump::jump_continue
:
1178 emit(NULL
, TGSI_OPCODE_CONT
);
1185 glsl_to_tgsi_visitor::visit(ir_function_signature
*ir
)
1192 glsl_to_tgsi_visitor::visit(ir_function
*ir
)
1194 /* Ignore function bodies other than main() -- we shouldn't see calls to
1195 * them since they should all be inlined before we get to glsl_to_tgsi.
1197 if (strcmp(ir
->name
, "main") == 0) {
1198 const ir_function_signature
*sig
;
1201 sig
= ir
->matching_signature(&empty
);
1205 foreach_iter(exec_list_iterator
, iter
, sig
->body
) {
1206 ir_instruction
*ir
= (ir_instruction
*)iter
.get();
1214 glsl_to_tgsi_visitor::try_emit_mad(ir_expression
*ir
, int mul_operand
)
1216 int nonmul_operand
= 1 - mul_operand
;
1218 st_dst_reg result_dst
;
1220 ir_expression
*expr
= ir
->operands
[mul_operand
]->as_expression();
1221 if (!expr
|| expr
->operation
!= ir_binop_mul
)
1224 expr
->operands
[0]->accept(this);
1226 expr
->operands
[1]->accept(this);
1228 ir
->operands
[nonmul_operand
]->accept(this);
1231 this->result
= get_temp(ir
->type
);
1232 result_dst
= st_dst_reg(this->result
);
1233 result_dst
.writemask
= (1 << ir
->type
->vector_elements
) - 1;
1234 emit(ir
, TGSI_OPCODE_MAD
, result_dst
, a
, b
, c
);
1240 * Emit MAD(a, -b, a) instead of AND(a, NOT(b))
1242 * The logic values are 1.0 for true and 0.0 for false. Logical-and is
1243 * implemented using multiplication, and logical-or is implemented using
1244 * addition. Logical-not can be implemented as (true - x), or (1.0 - x).
1245 * As result, the logical expression (a & !b) can be rewritten as:
1249 * - (a * 1) - (a * b)
1253 * This final expression can be implemented as a single MAD(a, -b, a)
1257 glsl_to_tgsi_visitor::try_emit_mad_for_and_not(ir_expression
*ir
, int try_operand
)
1259 const int other_operand
= 1 - try_operand
;
1262 ir_expression
*expr
= ir
->operands
[try_operand
]->as_expression();
1263 if (!expr
|| expr
->operation
!= ir_unop_logic_not
)
1266 ir
->operands
[other_operand
]->accept(this);
1268 expr
->operands
[0]->accept(this);
1271 b
.negate
= ~b
.negate
;
1273 this->result
= get_temp(ir
->type
);
1274 emit(ir
, TGSI_OPCODE_MAD
, st_dst_reg(this->result
), a
, b
, a
);
1280 glsl_to_tgsi_visitor::try_emit_sat(ir_expression
*ir
)
1282 /* Saturates were only introduced to vertex programs in
1283 * NV_vertex_program3, so don't give them to drivers in the VP.
1285 if (this->prog
->Target
== GL_VERTEX_PROGRAM_ARB
)
1288 ir_rvalue
*sat_src
= ir
->as_rvalue_to_saturate();
1292 sat_src
->accept(this);
1293 st_src_reg src
= this->result
;
1295 /* If we generated an expression instruction into a temporary in
1296 * processing the saturate's operand, apply the saturate to that
1297 * instruction. Otherwise, generate a MOV to do the saturate.
1299 * Note that we have to be careful to only do this optimization if
1300 * the instruction in question was what generated src->result. For
1301 * example, ir_dereference_array might generate a MUL instruction
1302 * to create the reladdr, and return us a src reg using that
1303 * reladdr. That MUL result is not the value we're trying to
1306 ir_expression
*sat_src_expr
= sat_src
->as_expression();
1307 if (sat_src_expr
&& (sat_src_expr
->operation
== ir_binop_mul
||
1308 sat_src_expr
->operation
== ir_binop_add
||
1309 sat_src_expr
->operation
== ir_binop_dot
)) {
1310 glsl_to_tgsi_instruction
*new_inst
;
1311 new_inst
= (glsl_to_tgsi_instruction
*)this->instructions
.get_tail();
1312 new_inst
->saturate
= true;
1314 this->result
= get_temp(ir
->type
);
1315 st_dst_reg result_dst
= st_dst_reg(this->result
);
1316 result_dst
.writemask
= (1 << ir
->type
->vector_elements
) - 1;
1317 glsl_to_tgsi_instruction
*inst
;
1318 inst
= emit(ir
, TGSI_OPCODE_MOV
, result_dst
, src
);
1319 inst
->saturate
= true;
1326 glsl_to_tgsi_visitor::reladdr_to_temp(ir_instruction
*ir
,
1327 st_src_reg
*reg
, int *num_reladdr
)
1332 emit_arl(ir
, address_reg
, *reg
->reladdr
);
1334 if (*num_reladdr
!= 1) {
1335 st_src_reg temp
= get_temp(glsl_type::vec4_type
);
1337 emit(ir
, TGSI_OPCODE_MOV
, st_dst_reg(temp
), *reg
);
1345 glsl_to_tgsi_visitor::visit(ir_expression
*ir
)
1347 unsigned int operand
;
1348 st_src_reg op
[Elements(ir
->operands
)];
1349 st_src_reg result_src
;
1350 st_dst_reg result_dst
;
1352 /* Quick peephole: Emit MAD(a, b, c) instead of ADD(MUL(a, b), c)
1354 if (ir
->operation
== ir_binop_add
) {
1355 if (try_emit_mad(ir
, 1))
1357 if (try_emit_mad(ir
, 0))
1361 /* Quick peephole: Emit OPCODE_MAD(-a, -b, a) instead of AND(a, NOT(b))
1363 if (ir
->operation
== ir_binop_logic_and
) {
1364 if (try_emit_mad_for_and_not(ir
, 1))
1366 if (try_emit_mad_for_and_not(ir
, 0))
1370 if (try_emit_sat(ir
))
1373 if (ir
->operation
== ir_quadop_vector
)
1374 assert(!"ir_quadop_vector should have been lowered");
1376 for (operand
= 0; operand
< ir
->get_num_operands(); operand
++) {
1377 this->result
.file
= PROGRAM_UNDEFINED
;
1378 ir
->operands
[operand
]->accept(this);
1379 if (this->result
.file
== PROGRAM_UNDEFINED
) {
1381 printf("Failed to get tree for expression operand:\n");
1382 ir
->operands
[operand
]->accept(&v
);
1385 op
[operand
] = this->result
;
1387 /* Matrix expression operands should have been broken down to vector
1388 * operations already.
1390 assert(!ir
->operands
[operand
]->type
->is_matrix());
1393 int vector_elements
= ir
->operands
[0]->type
->vector_elements
;
1394 if (ir
->operands
[1]) {
1395 vector_elements
= MAX2(vector_elements
,
1396 ir
->operands
[1]->type
->vector_elements
);
1399 this->result
.file
= PROGRAM_UNDEFINED
;
1401 /* Storage for our result. Ideally for an assignment we'd be using
1402 * the actual storage for the result here, instead.
1404 result_src
= get_temp(ir
->type
);
1405 /* convenience for the emit functions below. */
1406 result_dst
= st_dst_reg(result_src
);
1407 /* Limit writes to the channels that will be used by result_src later.
1408 * This does limit this temp's use as a temporary for multi-instruction
1411 result_dst
.writemask
= (1 << ir
->type
->vector_elements
) - 1;
1413 switch (ir
->operation
) {
1414 case ir_unop_logic_not
:
1415 if (result_dst
.type
!= GLSL_TYPE_FLOAT
)
1416 emit(ir
, TGSI_OPCODE_NOT
, result_dst
, op
[0]);
1418 /* Previously 'SEQ dst, src, 0.0' was used for this. However, many
1419 * older GPUs implement SEQ using multiple instructions (i915 uses two
1420 * SGE instructions and a MUL instruction). Since our logic values are
1421 * 0.0 and 1.0, 1-x also implements !x.
1423 op
[0].negate
= ~op
[0].negate
;
1424 emit(ir
, TGSI_OPCODE_ADD
, result_dst
, op
[0], st_src_reg_for_float(1.0));
1428 assert(result_dst
.type
== GLSL_TYPE_FLOAT
|| result_dst
.type
== GLSL_TYPE_INT
);
1429 if (result_dst
.type
== GLSL_TYPE_INT
)
1430 emit(ir
, TGSI_OPCODE_INEG
, result_dst
, op
[0]);
1432 op
[0].negate
= ~op
[0].negate
;
1437 assert(result_dst
.type
== GLSL_TYPE_FLOAT
);
1438 emit(ir
, TGSI_OPCODE_ABS
, result_dst
, op
[0]);
1441 emit(ir
, TGSI_OPCODE_SSG
, result_dst
, op
[0]);
1444 emit_scalar(ir
, TGSI_OPCODE_RCP
, result_dst
, op
[0]);
1448 emit_scalar(ir
, TGSI_OPCODE_EX2
, result_dst
, op
[0]);
1452 assert(!"not reached: should be handled by ir_explog_to_explog2");
1455 emit_scalar(ir
, TGSI_OPCODE_LG2
, result_dst
, op
[0]);
1458 emit_scalar(ir
, TGSI_OPCODE_SIN
, result_dst
, op
[0]);
1461 emit_scalar(ir
, TGSI_OPCODE_COS
, result_dst
, op
[0]);
1463 case ir_unop_sin_reduced
:
1464 emit_scs(ir
, TGSI_OPCODE_SIN
, result_dst
, op
[0]);
1466 case ir_unop_cos_reduced
:
1467 emit_scs(ir
, TGSI_OPCODE_COS
, result_dst
, op
[0]);
1471 emit(ir
, TGSI_OPCODE_DDX
, result_dst
, op
[0]);
1474 op
[0].negate
= ~op
[0].negate
;
1475 emit(ir
, TGSI_OPCODE_DDY
, result_dst
, op
[0]);
1478 case ir_unop_noise
: {
1479 /* At some point, a motivated person could add a better
1480 * implementation of noise. Currently not even the nvidia
1481 * binary drivers do anything more than this. In any case, the
1482 * place to do this is in the GL state tracker, not the poor
1485 emit(ir
, TGSI_OPCODE_MOV
, result_dst
, st_src_reg_for_float(0.5));
1490 emit(ir
, TGSI_OPCODE_ADD
, result_dst
, op
[0], op
[1]);
1493 emit(ir
, TGSI_OPCODE_SUB
, result_dst
, op
[0], op
[1]);
1497 emit(ir
, TGSI_OPCODE_MUL
, result_dst
, op
[0], op
[1]);
1500 if (result_dst
.type
== GLSL_TYPE_FLOAT
)
1501 assert(!"not reached: should be handled by ir_div_to_mul_rcp");
1503 emit(ir
, TGSI_OPCODE_DIV
, result_dst
, op
[0], op
[1]);
1506 if (result_dst
.type
== GLSL_TYPE_FLOAT
)
1507 assert(!"ir_binop_mod should have been converted to b * fract(a/b)");
1509 emit(ir
, TGSI_OPCODE_MOD
, result_dst
, op
[0], op
[1]);
1513 emit(ir
, TGSI_OPCODE_SLT
, result_dst
, op
[0], op
[1]);
1515 case ir_binop_greater
:
1516 emit(ir
, TGSI_OPCODE_SLT
, result_dst
, op
[1], op
[0]);
1518 case ir_binop_lequal
:
1519 emit(ir
, TGSI_OPCODE_SGE
, result_dst
, op
[1], op
[0]);
1521 case ir_binop_gequal
:
1522 emit(ir
, TGSI_OPCODE_SGE
, result_dst
, op
[0], op
[1]);
1524 case ir_binop_equal
:
1525 emit(ir
, TGSI_OPCODE_SEQ
, result_dst
, op
[0], op
[1]);
1527 case ir_binop_nequal
:
1528 emit(ir
, TGSI_OPCODE_SNE
, result_dst
, op
[0], op
[1]);
1530 case ir_binop_all_equal
:
1531 /* "==" operator producing a scalar boolean. */
1532 if (ir
->operands
[0]->type
->is_vector() ||
1533 ir
->operands
[1]->type
->is_vector()) {
1534 st_src_reg temp
= get_temp(native_integers
?
1535 glsl_type::get_instance(ir
->operands
[0]->type
->base_type
, 4, 1) :
1536 glsl_type::vec4_type
);
1538 if (native_integers
) {
1539 st_dst_reg temp_dst
= st_dst_reg(temp
);
1540 st_src_reg temp1
= st_src_reg(temp
), temp2
= st_src_reg(temp
);
1542 emit(ir
, TGSI_OPCODE_SEQ
, st_dst_reg(temp
), op
[0], op
[1]);
1544 /* Emit 1-3 AND operations to combine the SEQ results. */
1545 switch (ir
->operands
[0]->type
->vector_elements
) {
1549 temp_dst
.writemask
= WRITEMASK_Y
;
1550 temp1
.swizzle
= SWIZZLE_YYYY
;
1551 temp2
.swizzle
= SWIZZLE_ZZZZ
;
1552 emit(ir
, TGSI_OPCODE_AND
, temp_dst
, temp1
, temp2
);
1555 temp_dst
.writemask
= WRITEMASK_X
;
1556 temp1
.swizzle
= SWIZZLE_XXXX
;
1557 temp2
.swizzle
= SWIZZLE_YYYY
;
1558 emit(ir
, TGSI_OPCODE_AND
, temp_dst
, temp1
, temp2
);
1559 temp_dst
.writemask
= WRITEMASK_Y
;
1560 temp1
.swizzle
= SWIZZLE_ZZZZ
;
1561 temp2
.swizzle
= SWIZZLE_WWWW
;
1562 emit(ir
, TGSI_OPCODE_AND
, temp_dst
, temp1
, temp2
);
1565 temp1
.swizzle
= SWIZZLE_XXXX
;
1566 temp2
.swizzle
= SWIZZLE_YYYY
;
1567 emit(ir
, TGSI_OPCODE_AND
, result_dst
, temp1
, temp2
);
1569 emit(ir
, TGSI_OPCODE_SNE
, st_dst_reg(temp
), op
[0], op
[1]);
1571 /* After the dot-product, the value will be an integer on the
1572 * range [0,4]. Zero becomes 1.0, and positive values become zero.
1574 emit_dp(ir
, result_dst
, temp
, temp
, vector_elements
);
1576 /* Negating the result of the dot-product gives values on the range
1577 * [-4, 0]. Zero becomes 1.0, and negative values become zero.
1578 * This is achieved using SGE.
1580 st_src_reg sge_src
= result_src
;
1581 sge_src
.negate
= ~sge_src
.negate
;
1582 emit(ir
, TGSI_OPCODE_SGE
, result_dst
, sge_src
, st_src_reg_for_float(0.0));
1585 emit(ir
, TGSI_OPCODE_SEQ
, result_dst
, op
[0], op
[1]);
1588 case ir_binop_any_nequal
:
1589 /* "!=" operator producing a scalar boolean. */
1590 if (ir
->operands
[0]->type
->is_vector() ||
1591 ir
->operands
[1]->type
->is_vector()) {
1592 st_src_reg temp
= get_temp(native_integers
?
1593 glsl_type::get_instance(ir
->operands
[0]->type
->base_type
, 4, 1) :
1594 glsl_type::vec4_type
);
1595 emit(ir
, TGSI_OPCODE_SNE
, st_dst_reg(temp
), op
[0], op
[1]);
1597 if (native_integers
) {
1598 st_dst_reg temp_dst
= st_dst_reg(temp
);
1599 st_src_reg temp1
= st_src_reg(temp
), temp2
= st_src_reg(temp
);
1601 /* Emit 1-3 OR operations to combine the SNE results. */
1602 switch (ir
->operands
[0]->type
->vector_elements
) {
1606 temp_dst
.writemask
= WRITEMASK_Y
;
1607 temp1
.swizzle
= SWIZZLE_YYYY
;
1608 temp2
.swizzle
= SWIZZLE_ZZZZ
;
1609 emit(ir
, TGSI_OPCODE_OR
, temp_dst
, temp1
, temp2
);
1612 temp_dst
.writemask
= WRITEMASK_X
;
1613 temp1
.swizzle
= SWIZZLE_XXXX
;
1614 temp2
.swizzle
= SWIZZLE_YYYY
;
1615 emit(ir
, TGSI_OPCODE_OR
, temp_dst
, temp1
, temp2
);
1616 temp_dst
.writemask
= WRITEMASK_Y
;
1617 temp1
.swizzle
= SWIZZLE_ZZZZ
;
1618 temp2
.swizzle
= SWIZZLE_WWWW
;
1619 emit(ir
, TGSI_OPCODE_OR
, temp_dst
, temp1
, temp2
);
1622 temp1
.swizzle
= SWIZZLE_XXXX
;
1623 temp2
.swizzle
= SWIZZLE_YYYY
;
1624 emit(ir
, TGSI_OPCODE_OR
, result_dst
, temp1
, temp2
);
1626 /* After the dot-product, the value will be an integer on the
1627 * range [0,4]. Zero stays zero, and positive values become 1.0.
1629 glsl_to_tgsi_instruction
*const dp
=
1630 emit_dp(ir
, result_dst
, temp
, temp
, vector_elements
);
1631 if (this->prog
->Target
== GL_FRAGMENT_PROGRAM_ARB
) {
1632 /* The clamping to [0,1] can be done for free in the fragment
1633 * shader with a saturate.
1635 dp
->saturate
= true;
1637 /* Negating the result of the dot-product gives values on the range
1638 * [-4, 0]. Zero stays zero, and negative values become 1.0. This
1639 * achieved using SLT.
1641 st_src_reg slt_src
= result_src
;
1642 slt_src
.negate
= ~slt_src
.negate
;
1643 emit(ir
, TGSI_OPCODE_SLT
, result_dst
, slt_src
, st_src_reg_for_float(0.0));
1647 emit(ir
, TGSI_OPCODE_SNE
, result_dst
, op
[0], op
[1]);
1652 assert(ir
->operands
[0]->type
->is_vector());
1654 /* After the dot-product, the value will be an integer on the
1655 * range [0,4]. Zero stays zero, and positive values become 1.0.
1657 glsl_to_tgsi_instruction
*const dp
=
1658 emit_dp(ir
, result_dst
, op
[0], op
[0],
1659 ir
->operands
[0]->type
->vector_elements
);
1660 if (this->prog
->Target
== GL_FRAGMENT_PROGRAM_ARB
&&
1661 result_dst
.type
== GLSL_TYPE_FLOAT
) {
1662 /* The clamping to [0,1] can be done for free in the fragment
1663 * shader with a saturate.
1665 dp
->saturate
= true;
1666 } else if (result_dst
.type
== GLSL_TYPE_FLOAT
) {
1667 /* Negating the result of the dot-product gives values on the range
1668 * [-4, 0]. Zero stays zero, and negative values become 1.0. This
1669 * is achieved using SLT.
1671 st_src_reg slt_src
= result_src
;
1672 slt_src
.negate
= ~slt_src
.negate
;
1673 emit(ir
, TGSI_OPCODE_SLT
, result_dst
, slt_src
, st_src_reg_for_float(0.0));
1676 /* Use SNE 0 if integers are being used as boolean values. */
1677 emit(ir
, TGSI_OPCODE_SNE
, result_dst
, result_src
, st_src_reg_for_int(0));
1682 case ir_binop_logic_xor
:
1683 if (native_integers
)
1684 emit(ir
, TGSI_OPCODE_XOR
, result_dst
, op
[0], op
[1]);
1686 emit(ir
, TGSI_OPCODE_SNE
, result_dst
, op
[0], op
[1]);
1689 case ir_binop_logic_or
: {
1690 if (native_integers
) {
1691 /* If integers are used as booleans, we can use an actual "or"
1694 assert(native_integers
);
1695 emit(ir
, TGSI_OPCODE_OR
, result_dst
, op
[0], op
[1]);
1697 /* After the addition, the value will be an integer on the
1698 * range [0,2]. Zero stays zero, and positive values become 1.0.
1700 glsl_to_tgsi_instruction
*add
=
1701 emit(ir
, TGSI_OPCODE_ADD
, result_dst
, op
[0], op
[1]);
1702 if (this->prog
->Target
== GL_FRAGMENT_PROGRAM_ARB
) {
1703 /* The clamping to [0,1] can be done for free in the fragment
1704 * shader with a saturate if floats are being used as boolean values.
1706 add
->saturate
= true;
1708 /* Negating the result of the addition gives values on the range
1709 * [-2, 0]. Zero stays zero, and negative values become 1.0. This
1710 * is achieved using SLT.
1712 st_src_reg slt_src
= result_src
;
1713 slt_src
.negate
= ~slt_src
.negate
;
1714 emit(ir
, TGSI_OPCODE_SLT
, result_dst
, slt_src
, st_src_reg_for_float(0.0));
1720 case ir_binop_logic_and
:
1721 /* If native integers are disabled, the bool args are stored as float 0.0
1722 * or 1.0, so "mul" gives us "and". If they're enabled, just use the
1723 * actual AND opcode.
1725 if (native_integers
)
1726 emit(ir
, TGSI_OPCODE_AND
, result_dst
, op
[0], op
[1]);
1728 emit(ir
, TGSI_OPCODE_MUL
, result_dst
, op
[0], op
[1]);
1732 assert(ir
->operands
[0]->type
->is_vector());
1733 assert(ir
->operands
[0]->type
== ir
->operands
[1]->type
);
1734 emit_dp(ir
, result_dst
, op
[0], op
[1],
1735 ir
->operands
[0]->type
->vector_elements
);
1739 /* sqrt(x) = x * rsq(x). */
1740 emit_scalar(ir
, TGSI_OPCODE_RSQ
, result_dst
, op
[0]);
1741 emit(ir
, TGSI_OPCODE_MUL
, result_dst
, result_src
, op
[0]);
1742 /* For incoming channels <= 0, set the result to 0. */
1743 op
[0].negate
= ~op
[0].negate
;
1744 emit(ir
, TGSI_OPCODE_CMP
, result_dst
,
1745 op
[0], result_src
, st_src_reg_for_float(0.0));
1748 emit_scalar(ir
, TGSI_OPCODE_RSQ
, result_dst
, op
[0]);
1751 if (native_integers
) {
1752 emit(ir
, TGSI_OPCODE_I2F
, result_dst
, op
[0]);
1755 /* fallthrough to next case otherwise */
1757 if (native_integers
) {
1758 emit(ir
, TGSI_OPCODE_AND
, result_dst
, op
[0], st_src_reg_for_float(1.0));
1761 /* fallthrough to next case otherwise */
1764 /* Converting between signed and unsigned integers is a no-op. */
1768 if (native_integers
) {
1769 /* Booleans are stored as integers using ~0 for true and 0 for false.
1770 * GLSL requires that int(bool) return 1 for true and 0 for false.
1771 * This conversion is done with AND, but it could be done with NEG.
1773 emit(ir
, TGSI_OPCODE_AND
, result_dst
, op
[0], st_src_reg_for_int(1));
1775 /* Booleans and integers are both stored as floats when native
1776 * integers are disabled.
1782 if (native_integers
)
1783 emit(ir
, TGSI_OPCODE_F2I
, result_dst
, op
[0]);
1785 emit(ir
, TGSI_OPCODE_TRUNC
, result_dst
, op
[0]);
1788 emit(ir
, TGSI_OPCODE_SNE
, result_dst
, op
[0], st_src_reg_for_float(0.0));
1791 if (native_integers
)
1792 emit(ir
, TGSI_OPCODE_INEG
, result_dst
, op
[0]);
1794 emit(ir
, TGSI_OPCODE_SNE
, result_dst
, op
[0], st_src_reg_for_float(0.0));
1797 emit(ir
, TGSI_OPCODE_TRUNC
, result_dst
, op
[0]);
1800 op
[0].negate
= ~op
[0].negate
;
1801 emit(ir
, TGSI_OPCODE_FLR
, result_dst
, op
[0]);
1802 result_src
.negate
= ~result_src
.negate
;
1805 emit(ir
, TGSI_OPCODE_FLR
, result_dst
, op
[0]);
1808 emit(ir
, TGSI_OPCODE_FRC
, result_dst
, op
[0]);
1812 emit(ir
, TGSI_OPCODE_MIN
, result_dst
, op
[0], op
[1]);
1815 emit(ir
, TGSI_OPCODE_MAX
, result_dst
, op
[0], op
[1]);
1818 emit_scalar(ir
, TGSI_OPCODE_POW
, result_dst
, op
[0], op
[1]);
1821 case ir_unop_bit_not
:
1822 if (native_integers
) {
1823 emit(ir
, TGSI_OPCODE_NOT
, result_dst
, op
[0]);
1827 if (native_integers
) {
1828 emit(ir
, TGSI_OPCODE_U2F
, result_dst
, op
[0]);
1831 case ir_binop_lshift
:
1832 if (native_integers
) {
1833 emit(ir
, TGSI_OPCODE_SHL
, result_dst
, op
[0]);
1836 case ir_binop_rshift
:
1837 if (native_integers
) {
1838 emit(ir
, TGSI_OPCODE_ISHR
, result_dst
, op
[0]);
1841 case ir_binop_bit_and
:
1842 if (native_integers
) {
1843 emit(ir
, TGSI_OPCODE_AND
, result_dst
, op
[0]);
1846 case ir_binop_bit_xor
:
1847 if (native_integers
) {
1848 emit(ir
, TGSI_OPCODE_XOR
, result_dst
, op
[0]);
1851 case ir_binop_bit_or
:
1852 if (native_integers
) {
1853 emit(ir
, TGSI_OPCODE_OR
, result_dst
, op
[0]);
1856 case ir_unop_round_even
:
1857 assert(!"GLSL 1.30 features unsupported");
1860 case ir_quadop_vector
:
1861 /* This operation should have already been handled.
1863 assert(!"Should not get here.");
1867 this->result
= result_src
;
1872 glsl_to_tgsi_visitor::visit(ir_swizzle
*ir
)
1878 /* Note that this is only swizzles in expressions, not those on the left
1879 * hand side of an assignment, which do write masking. See ir_assignment
1883 ir
->val
->accept(this);
1885 assert(src
.file
!= PROGRAM_UNDEFINED
);
1887 for (i
= 0; i
< 4; i
++) {
1888 if (i
< ir
->type
->vector_elements
) {
1891 swizzle
[i
] = GET_SWZ(src
.swizzle
, ir
->mask
.x
);
1894 swizzle
[i
] = GET_SWZ(src
.swizzle
, ir
->mask
.y
);
1897 swizzle
[i
] = GET_SWZ(src
.swizzle
, ir
->mask
.z
);
1900 swizzle
[i
] = GET_SWZ(src
.swizzle
, ir
->mask
.w
);
1904 /* If the type is smaller than a vec4, replicate the last
1907 swizzle
[i
] = swizzle
[ir
->type
->vector_elements
- 1];
1911 src
.swizzle
= MAKE_SWIZZLE4(swizzle
[0], swizzle
[1], swizzle
[2], swizzle
[3]);
1917 glsl_to_tgsi_visitor::visit(ir_dereference_variable
*ir
)
1919 variable_storage
*entry
= find_variable_storage(ir
->var
);
1920 ir_variable
*var
= ir
->var
;
1923 switch (var
->mode
) {
1924 case ir_var_uniform
:
1925 entry
= new(mem_ctx
) variable_storage(var
, PROGRAM_UNIFORM
,
1927 this->variables
.push_tail(entry
);
1931 /* The linker assigns locations for varyings and attributes,
1932 * including deprecated builtins (like gl_Color), user-assign
1933 * generic attributes (glBindVertexLocation), and
1934 * user-defined varyings.
1936 * FINISHME: We would hit this path for function arguments. Fix!
1938 assert(var
->location
!= -1);
1939 entry
= new(mem_ctx
) variable_storage(var
,
1944 assert(var
->location
!= -1);
1945 entry
= new(mem_ctx
) variable_storage(var
,
1949 case ir_var_system_value
:
1950 entry
= new(mem_ctx
) variable_storage(var
,
1951 PROGRAM_SYSTEM_VALUE
,
1955 case ir_var_temporary
:
1956 entry
= new(mem_ctx
) variable_storage(var
, PROGRAM_TEMPORARY
,
1958 this->variables
.push_tail(entry
);
1960 next_temp
+= type_size(var
->type
);
1965 printf("Failed to make storage for %s\n", var
->name
);
1970 this->result
= st_src_reg(entry
->file
, entry
->index
, var
->type
);
1971 if (!native_integers
)
1972 this->result
.type
= GLSL_TYPE_FLOAT
;
1976 glsl_to_tgsi_visitor::visit(ir_dereference_array
*ir
)
1980 int element_size
= type_size(ir
->type
);
1982 index
= ir
->array_index
->constant_expression_value();
1984 ir
->array
->accept(this);
1988 src
.index
+= index
->value
.i
[0] * element_size
;
1990 /* Variable index array dereference. It eats the "vec4" of the
1991 * base of the array and an index that offsets the TGSI register
1994 ir
->array_index
->accept(this);
1996 st_src_reg index_reg
;
1998 if (element_size
== 1) {
1999 index_reg
= this->result
;
2001 index_reg
= get_temp(native_integers
?
2002 glsl_type::int_type
: glsl_type::float_type
);
2004 emit(ir
, TGSI_OPCODE_MUL
, st_dst_reg(index_reg
),
2005 this->result
, st_src_reg_for_type(index_reg
.type
, element_size
));
2008 /* If there was already a relative address register involved, add the
2009 * new and the old together to get the new offset.
2011 if (src
.reladdr
!= NULL
) {
2012 st_src_reg accum_reg
= get_temp(native_integers
?
2013 glsl_type::int_type
: glsl_type::float_type
);
2015 emit(ir
, TGSI_OPCODE_ADD
, st_dst_reg(accum_reg
),
2016 index_reg
, *src
.reladdr
);
2018 index_reg
= accum_reg
;
2021 src
.reladdr
= ralloc(mem_ctx
, st_src_reg
);
2022 memcpy(src
.reladdr
, &index_reg
, sizeof(index_reg
));
2025 /* If the type is smaller than a vec4, replicate the last channel out. */
2026 if (ir
->type
->is_scalar() || ir
->type
->is_vector())
2027 src
.swizzle
= swizzle_for_size(ir
->type
->vector_elements
);
2029 src
.swizzle
= SWIZZLE_NOOP
;
2035 glsl_to_tgsi_visitor::visit(ir_dereference_record
*ir
)
2038 const glsl_type
*struct_type
= ir
->record
->type
;
2041 ir
->record
->accept(this);
2043 for (i
= 0; i
< struct_type
->length
; i
++) {
2044 if (strcmp(struct_type
->fields
.structure
[i
].name
, ir
->field
) == 0)
2046 offset
+= type_size(struct_type
->fields
.structure
[i
].type
);
2049 /* If the type is smaller than a vec4, replicate the last channel out. */
2050 if (ir
->type
->is_scalar() || ir
->type
->is_vector())
2051 this->result
.swizzle
= swizzle_for_size(ir
->type
->vector_elements
);
2053 this->result
.swizzle
= SWIZZLE_NOOP
;
2055 this->result
.index
+= offset
;
2059 * We want to be careful in assignment setup to hit the actual storage
2060 * instead of potentially using a temporary like we might with the
2061 * ir_dereference handler.
2064 get_assignment_lhs(ir_dereference
*ir
, glsl_to_tgsi_visitor
*v
)
2066 /* The LHS must be a dereference. If the LHS is a variable indexed array
2067 * access of a vector, it must be separated into a series conditional moves
2068 * before reaching this point (see ir_vec_index_to_cond_assign).
2070 assert(ir
->as_dereference());
2071 ir_dereference_array
*deref_array
= ir
->as_dereference_array();
2073 assert(!deref_array
->array
->type
->is_vector());
2076 /* Use the rvalue deref handler for the most part. We'll ignore
2077 * swizzles in it and write swizzles using writemask, though.
2080 return st_dst_reg(v
->result
);
2084 * Process the condition of a conditional assignment
2086 * Examines the condition of a conditional assignment to generate the optimal
2087 * first operand of a \c CMP instruction. If the condition is a relational
2088 * operator with 0 (e.g., \c ir_binop_less), the value being compared will be
2089 * used as the source for the \c CMP instruction. Otherwise the comparison
2090 * is processed to a boolean result, and the boolean result is used as the
2091 * operand to the CMP instruction.
2094 glsl_to_tgsi_visitor::process_move_condition(ir_rvalue
*ir
)
2096 ir_rvalue
*src_ir
= ir
;
2098 bool switch_order
= false;
2100 ir_expression
*const expr
= ir
->as_expression();
2101 if ((expr
!= NULL
) && (expr
->get_num_operands() == 2)) {
2102 bool zero_on_left
= false;
2104 if (expr
->operands
[0]->is_zero()) {
2105 src_ir
= expr
->operands
[1];
2106 zero_on_left
= true;
2107 } else if (expr
->operands
[1]->is_zero()) {
2108 src_ir
= expr
->operands
[0];
2109 zero_on_left
= false;
2113 * (a < 0) T F F ( a < 0) T F F
2114 * (0 < a) F F T (-a < 0) F F T
2115 * (a <= 0) T T F (-a < 0) F F T (swap order of other operands)
2116 * (0 <= a) F T T ( a < 0) T F F (swap order of other operands)
2117 * (a > 0) F F T (-a < 0) F F T
2118 * (0 > a) T F F ( a < 0) T F F
2119 * (a >= 0) F T T ( a < 0) T F F (swap order of other operands)
2120 * (0 >= a) T T F (-a < 0) F F T (swap order of other operands)
2122 * Note that exchanging the order of 0 and 'a' in the comparison simply
2123 * means that the value of 'a' should be negated.
2126 switch (expr
->operation
) {
2128 switch_order
= false;
2129 negate
= zero_on_left
;
2132 case ir_binop_greater
:
2133 switch_order
= false;
2134 negate
= !zero_on_left
;
2137 case ir_binop_lequal
:
2138 switch_order
= true;
2139 negate
= !zero_on_left
;
2142 case ir_binop_gequal
:
2143 switch_order
= true;
2144 negate
= zero_on_left
;
2148 /* This isn't the right kind of comparison afterall, so make sure
2149 * the whole condition is visited.
2157 src_ir
->accept(this);
2159 /* We use the TGSI_OPCODE_CMP (a < 0 ? b : c) for conditional moves, and the
2160 * condition we produced is 0.0 or 1.0. By flipping the sign, we can
2161 * choose which value TGSI_OPCODE_CMP produces without an extra instruction
2162 * computing the condition.
2165 this->result
.negate
= ~this->result
.negate
;
2167 return switch_order
;
2171 glsl_to_tgsi_visitor::visit(ir_assignment
*ir
)
2177 ir
->rhs
->accept(this);
2180 l
= get_assignment_lhs(ir
->lhs
, this);
2182 /* FINISHME: This should really set to the correct maximal writemask for each
2183 * FINISHME: component written (in the loops below). This case can only
2184 * FINISHME: occur for matrices, arrays, and structures.
2186 if (ir
->write_mask
== 0) {
2187 assert(!ir
->lhs
->type
->is_scalar() && !ir
->lhs
->type
->is_vector());
2188 l
.writemask
= WRITEMASK_XYZW
;
2189 } else if (ir
->lhs
->type
->is_scalar() &&
2190 ir
->lhs
->variable_referenced()->mode
== ir_var_out
) {
2191 /* FINISHME: This hack makes writing to gl_FragDepth, which lives in the
2192 * FINISHME: W component of fragment shader output zero, work correctly.
2194 l
.writemask
= WRITEMASK_XYZW
;
2197 int first_enabled_chan
= 0;
2200 l
.writemask
= ir
->write_mask
;
2202 for (int i
= 0; i
< 4; i
++) {
2203 if (l
.writemask
& (1 << i
)) {
2204 first_enabled_chan
= GET_SWZ(r
.swizzle
, i
);
2209 /* Swizzle a small RHS vector into the channels being written.
2211 * glsl ir treats write_mask as dictating how many channels are
2212 * present on the RHS while TGSI treats write_mask as just
2213 * showing which channels of the vec4 RHS get written.
2215 for (int i
= 0; i
< 4; i
++) {
2216 if (l
.writemask
& (1 << i
))
2217 swizzles
[i
] = GET_SWZ(r
.swizzle
, rhs_chan
++);
2219 swizzles
[i
] = first_enabled_chan
;
2221 r
.swizzle
= MAKE_SWIZZLE4(swizzles
[0], swizzles
[1],
2222 swizzles
[2], swizzles
[3]);
2225 assert(l
.file
!= PROGRAM_UNDEFINED
);
2226 assert(r
.file
!= PROGRAM_UNDEFINED
);
2228 if (ir
->condition
) {
2229 const bool switch_order
= this->process_move_condition(ir
->condition
);
2230 st_src_reg condition
= this->result
;
2232 for (i
= 0; i
< type_size(ir
->lhs
->type
); i
++) {
2233 st_src_reg l_src
= st_src_reg(l
);
2234 st_src_reg condition_temp
= condition
;
2235 l_src
.swizzle
= swizzle_for_size(ir
->lhs
->type
->vector_elements
);
2237 if (native_integers
) {
2238 /* This is necessary because TGSI's CMP instruction expects the
2239 * condition to be a float, and we store booleans as integers.
2240 * If TGSI had a UCMP instruction or similar, this extra
2241 * instruction would not be necessary.
2243 condition_temp
= get_temp(glsl_type::vec4_type
);
2244 condition
.negate
= 0;
2245 emit(ir
, TGSI_OPCODE_I2F
, st_dst_reg(condition_temp
), condition
);
2246 condition_temp
.swizzle
= condition
.swizzle
;
2250 emit(ir
, TGSI_OPCODE_CMP
, l
, condition_temp
, l_src
, r
);
2252 emit(ir
, TGSI_OPCODE_CMP
, l
, condition_temp
, r
, l_src
);
2258 } else if (ir
->rhs
->as_expression() &&
2259 this->instructions
.get_tail() &&
2260 ir
->rhs
== ((glsl_to_tgsi_instruction
*)this->instructions
.get_tail())->ir
&&
2261 type_size(ir
->lhs
->type
) == 1 &&
2262 l
.writemask
== ((glsl_to_tgsi_instruction
*)this->instructions
.get_tail())->dst
.writemask
) {
2263 /* To avoid emitting an extra MOV when assigning an expression to a
2264 * variable, emit the last instruction of the expression again, but
2265 * replace the destination register with the target of the assignment.
2266 * Dead code elimination will remove the original instruction.
2268 glsl_to_tgsi_instruction
*inst
, *new_inst
;
2269 inst
= (glsl_to_tgsi_instruction
*)this->instructions
.get_tail();
2270 new_inst
= emit(ir
, inst
->op
, l
, inst
->src
[0], inst
->src
[1], inst
->src
[2]);
2271 new_inst
->saturate
= inst
->saturate
;
2272 inst
->dead_mask
= inst
->dst
.writemask
;
2274 for (i
= 0; i
< type_size(ir
->lhs
->type
); i
++) {
2275 emit(ir
, TGSI_OPCODE_MOV
, l
, r
);
2284 glsl_to_tgsi_visitor::visit(ir_constant
*ir
)
2287 GLfloat stack_vals
[4] = { 0 };
2288 gl_constant_value
*values
= (gl_constant_value
*) stack_vals
;
2289 GLenum gl_type
= GL_NONE
;
2291 static int in_array
= 0;
2292 gl_register_file file
= in_array
? PROGRAM_CONSTANT
: PROGRAM_IMMEDIATE
;
2294 /* Unfortunately, 4 floats is all we can get into
2295 * _mesa_add_typed_unnamed_constant. So, make a temp to store an
2296 * aggregate constant and move each constant value into it. If we
2297 * get lucky, copy propagation will eliminate the extra moves.
2299 if (ir
->type
->base_type
== GLSL_TYPE_STRUCT
) {
2300 st_src_reg temp_base
= get_temp(ir
->type
);
2301 st_dst_reg temp
= st_dst_reg(temp_base
);
2303 foreach_iter(exec_list_iterator
, iter
, ir
->components
) {
2304 ir_constant
*field_value
= (ir_constant
*)iter
.get();
2305 int size
= type_size(field_value
->type
);
2309 field_value
->accept(this);
2312 for (i
= 0; i
< (unsigned int)size
; i
++) {
2313 emit(ir
, TGSI_OPCODE_MOV
, temp
, src
);
2319 this->result
= temp_base
;
2323 if (ir
->type
->is_array()) {
2324 st_src_reg temp_base
= get_temp(ir
->type
);
2325 st_dst_reg temp
= st_dst_reg(temp_base
);
2326 int size
= type_size(ir
->type
->fields
.array
);
2331 for (i
= 0; i
< ir
->type
->length
; i
++) {
2332 ir
->array_elements
[i
]->accept(this);
2334 for (int j
= 0; j
< size
; j
++) {
2335 emit(ir
, TGSI_OPCODE_MOV
, temp
, src
);
2341 this->result
= temp_base
;
2346 if (ir
->type
->is_matrix()) {
2347 st_src_reg mat
= get_temp(ir
->type
);
2348 st_dst_reg mat_column
= st_dst_reg(mat
);
2350 for (i
= 0; i
< ir
->type
->matrix_columns
; i
++) {
2351 assert(ir
->type
->base_type
== GLSL_TYPE_FLOAT
);
2352 values
= (gl_constant_value
*) &ir
->value
.f
[i
* ir
->type
->vector_elements
];
2354 src
= st_src_reg(file
, -1, ir
->type
->base_type
);
2355 src
.index
= add_constant(file
,
2357 ir
->type
->vector_elements
,
2360 emit(ir
, TGSI_OPCODE_MOV
, mat_column
, src
);
2369 switch (ir
->type
->base_type
) {
2370 case GLSL_TYPE_FLOAT
:
2372 for (i
= 0; i
< ir
->type
->vector_elements
; i
++) {
2373 values
[i
].f
= ir
->value
.f
[i
];
2376 case GLSL_TYPE_UINT
:
2377 gl_type
= native_integers
? GL_UNSIGNED_INT
: GL_FLOAT
;
2378 for (i
= 0; i
< ir
->type
->vector_elements
; i
++) {
2379 if (native_integers
)
2380 values
[i
].u
= ir
->value
.u
[i
];
2382 values
[i
].f
= ir
->value
.u
[i
];
2386 gl_type
= native_integers
? GL_INT
: GL_FLOAT
;
2387 for (i
= 0; i
< ir
->type
->vector_elements
; i
++) {
2388 if (native_integers
)
2389 values
[i
].i
= ir
->value
.i
[i
];
2391 values
[i
].f
= ir
->value
.i
[i
];
2394 case GLSL_TYPE_BOOL
:
2395 gl_type
= native_integers
? GL_BOOL
: GL_FLOAT
;
2396 for (i
= 0; i
< ir
->type
->vector_elements
; i
++) {
2397 if (native_integers
)
2398 values
[i
].b
= ir
->value
.b
[i
];
2400 values
[i
].f
= ir
->value
.b
[i
];
2404 assert(!"Non-float/uint/int/bool constant");
2407 this->result
= st_src_reg(file
, -1, ir
->type
);
2408 this->result
.index
= add_constant(file
,
2410 ir
->type
->vector_elements
,
2412 &this->result
.swizzle
);
2416 glsl_to_tgsi_visitor::get_function_signature(ir_function_signature
*sig
)
2418 function_entry
*entry
;
2420 foreach_iter(exec_list_iterator
, iter
, this->function_signatures
) {
2421 entry
= (function_entry
*)iter
.get();
2423 if (entry
->sig
== sig
)
2427 entry
= ralloc(mem_ctx
, function_entry
);
2429 entry
->sig_id
= this->next_signature_id
++;
2430 entry
->bgn_inst
= NULL
;
2432 /* Allocate storage for all the parameters. */
2433 foreach_iter(exec_list_iterator
, iter
, sig
->parameters
) {
2434 ir_variable
*param
= (ir_variable
*)iter
.get();
2435 variable_storage
*storage
;
2437 storage
= find_variable_storage(param
);
2440 storage
= new(mem_ctx
) variable_storage(param
, PROGRAM_TEMPORARY
,
2442 this->variables
.push_tail(storage
);
2444 this->next_temp
+= type_size(param
->type
);
2447 if (!sig
->return_type
->is_void()) {
2448 entry
->return_reg
= get_temp(sig
->return_type
);
2450 entry
->return_reg
= undef_src
;
2453 this->function_signatures
.push_tail(entry
);
2458 glsl_to_tgsi_visitor::visit(ir_call
*ir
)
2460 glsl_to_tgsi_instruction
*call_inst
;
2461 ir_function_signature
*sig
= ir
->get_callee();
2462 function_entry
*entry
= get_function_signature(sig
);
2465 /* Process in parameters. */
2466 exec_list_iterator sig_iter
= sig
->parameters
.iterator();
2467 foreach_iter(exec_list_iterator
, iter
, *ir
) {
2468 ir_rvalue
*param_rval
= (ir_rvalue
*)iter
.get();
2469 ir_variable
*param
= (ir_variable
*)sig_iter
.get();
2471 if (param
->mode
== ir_var_in
||
2472 param
->mode
== ir_var_inout
) {
2473 variable_storage
*storage
= find_variable_storage(param
);
2476 param_rval
->accept(this);
2477 st_src_reg r
= this->result
;
2480 l
.file
= storage
->file
;
2481 l
.index
= storage
->index
;
2483 l
.writemask
= WRITEMASK_XYZW
;
2484 l
.cond_mask
= COND_TR
;
2486 for (i
= 0; i
< type_size(param
->type
); i
++) {
2487 emit(ir
, TGSI_OPCODE_MOV
, l
, r
);
2495 assert(!sig_iter
.has_next());
2497 /* Emit call instruction */
2498 call_inst
= emit(ir
, TGSI_OPCODE_CAL
);
2499 call_inst
->function
= entry
;
2501 /* Process out parameters. */
2502 sig_iter
= sig
->parameters
.iterator();
2503 foreach_iter(exec_list_iterator
, iter
, *ir
) {
2504 ir_rvalue
*param_rval
= (ir_rvalue
*)iter
.get();
2505 ir_variable
*param
= (ir_variable
*)sig_iter
.get();
2507 if (param
->mode
== ir_var_out
||
2508 param
->mode
== ir_var_inout
) {
2509 variable_storage
*storage
= find_variable_storage(param
);
2513 r
.file
= storage
->file
;
2514 r
.index
= storage
->index
;
2516 r
.swizzle
= SWIZZLE_NOOP
;
2519 param_rval
->accept(this);
2520 st_dst_reg l
= st_dst_reg(this->result
);
2522 for (i
= 0; i
< type_size(param
->type
); i
++) {
2523 emit(ir
, TGSI_OPCODE_MOV
, l
, r
);
2531 assert(!sig_iter
.has_next());
2533 /* Process return value. */
2534 this->result
= entry
->return_reg
;
2538 glsl_to_tgsi_visitor::visit(ir_texture
*ir
)
2540 st_src_reg result_src
, coord
, lod_info
, projector
, dx
, dy
, offset
;
2541 st_dst_reg result_dst
, coord_dst
;
2542 glsl_to_tgsi_instruction
*inst
= NULL
;
2543 unsigned opcode
= TGSI_OPCODE_NOP
;
2545 if (ir
->coordinate
) {
2546 ir
->coordinate
->accept(this);
2548 /* Put our coords in a temp. We'll need to modify them for shadow,
2549 * projection, or LOD, so the only case we'd use it as is is if
2550 * we're doing plain old texturing. The optimization passes on
2551 * glsl_to_tgsi_visitor should handle cleaning up our mess in that case.
2553 coord
= get_temp(glsl_type::vec4_type
);
2554 coord_dst
= st_dst_reg(coord
);
2555 emit(ir
, TGSI_OPCODE_MOV
, coord_dst
, this->result
);
2558 if (ir
->projector
) {
2559 ir
->projector
->accept(this);
2560 projector
= this->result
;
2563 /* Storage for our result. Ideally for an assignment we'd be using
2564 * the actual storage for the result here, instead.
2566 result_src
= get_temp(glsl_type::vec4_type
);
2567 result_dst
= st_dst_reg(result_src
);
2571 opcode
= TGSI_OPCODE_TEX
;
2574 opcode
= TGSI_OPCODE_TXB
;
2575 ir
->lod_info
.bias
->accept(this);
2576 lod_info
= this->result
;
2579 opcode
= TGSI_OPCODE_TXL
;
2580 ir
->lod_info
.lod
->accept(this);
2581 lod_info
= this->result
;
2584 opcode
= TGSI_OPCODE_TXD
;
2585 ir
->lod_info
.grad
.dPdx
->accept(this);
2587 ir
->lod_info
.grad
.dPdy
->accept(this);
2591 opcode
= TGSI_OPCODE_TXQ
;
2592 ir
->lod_info
.lod
->accept(this);
2593 lod_info
= this->result
;
2596 opcode
= TGSI_OPCODE_TXF
;
2597 ir
->lod_info
.lod
->accept(this);
2598 lod_info
= this->result
;
2600 ir
->offset
->accept(this);
2601 offset
= this->result
;
2606 const glsl_type
*sampler_type
= ir
->sampler
->type
;
2608 if (ir
->projector
) {
2609 if (opcode
== TGSI_OPCODE_TEX
) {
2610 /* Slot the projector in as the last component of the coord. */
2611 coord_dst
.writemask
= WRITEMASK_W
;
2612 emit(ir
, TGSI_OPCODE_MOV
, coord_dst
, projector
);
2613 coord_dst
.writemask
= WRITEMASK_XYZW
;
2614 opcode
= TGSI_OPCODE_TXP
;
2616 st_src_reg coord_w
= coord
;
2617 coord_w
.swizzle
= SWIZZLE_WWWW
;
2619 /* For the other TEX opcodes there's no projective version
2620 * since the last slot is taken up by LOD info. Do the
2621 * projective divide now.
2623 coord_dst
.writemask
= WRITEMASK_W
;
2624 emit(ir
, TGSI_OPCODE_RCP
, coord_dst
, projector
);
2626 /* In the case where we have to project the coordinates "by hand,"
2627 * the shadow comparator value must also be projected.
2629 st_src_reg tmp_src
= coord
;
2630 if (ir
->shadow_comparitor
) {
2631 /* Slot the shadow value in as the second to last component of the
2634 ir
->shadow_comparitor
->accept(this);
2636 tmp_src
= get_temp(glsl_type::vec4_type
);
2637 st_dst_reg tmp_dst
= st_dst_reg(tmp_src
);
2639 /* Projective division not allowed for array samplers. */
2640 assert(!sampler_type
->sampler_array
);
2642 tmp_dst
.writemask
= WRITEMASK_Z
;
2643 emit(ir
, TGSI_OPCODE_MOV
, tmp_dst
, this->result
);
2645 tmp_dst
.writemask
= WRITEMASK_XY
;
2646 emit(ir
, TGSI_OPCODE_MOV
, tmp_dst
, coord
);
2649 coord_dst
.writemask
= WRITEMASK_XYZ
;
2650 emit(ir
, TGSI_OPCODE_MUL
, coord_dst
, tmp_src
, coord_w
);
2652 coord_dst
.writemask
= WRITEMASK_XYZW
;
2653 coord
.swizzle
= SWIZZLE_XYZW
;
2657 /* If projection is done and the opcode is not TGSI_OPCODE_TXP, then the shadow
2658 * comparator was put in the correct place (and projected) by the code,
2659 * above, that handles by-hand projection.
2661 if (ir
->shadow_comparitor
&& (!ir
->projector
|| opcode
== TGSI_OPCODE_TXP
)) {
2662 /* Slot the shadow value in as the second to last component of the
2665 ir
->shadow_comparitor
->accept(this);
2667 /* XXX This will need to be updated for cubemap array samplers. */
2668 if (sampler_type
->sampler_dimensionality
== GLSL_SAMPLER_DIM_2D
&&
2669 sampler_type
->sampler_array
) {
2670 coord_dst
.writemask
= WRITEMASK_W
;
2672 coord_dst
.writemask
= WRITEMASK_Z
;
2675 emit(ir
, TGSI_OPCODE_MOV
, coord_dst
, this->result
);
2676 coord_dst
.writemask
= WRITEMASK_XYZW
;
2679 if (opcode
== TGSI_OPCODE_TXL
|| opcode
== TGSI_OPCODE_TXB
||
2680 opcode
== TGSI_OPCODE_TXF
) {
2681 /* TGSI stores LOD or LOD bias in the last channel of the coords. */
2682 coord_dst
.writemask
= WRITEMASK_W
;
2683 emit(ir
, TGSI_OPCODE_MOV
, coord_dst
, lod_info
);
2684 coord_dst
.writemask
= WRITEMASK_XYZW
;
2687 if (opcode
== TGSI_OPCODE_TXD
)
2688 inst
= emit(ir
, opcode
, result_dst
, coord
, dx
, dy
);
2689 else if (opcode
== TGSI_OPCODE_TXQ
)
2690 inst
= emit(ir
, opcode
, result_dst
, lod_info
);
2691 else if (opcode
== TGSI_OPCODE_TXF
) {
2692 inst
= emit(ir
, opcode
, result_dst
, coord
);
2694 inst
= emit(ir
, opcode
, result_dst
, coord
);
2696 if (ir
->shadow_comparitor
)
2697 inst
->tex_shadow
= GL_TRUE
;
2699 inst
->sampler
= _mesa_get_sampler_uniform_value(ir
->sampler
,
2700 this->shader_program
,
2704 inst
->tex_offset_num_offset
= 1;
2705 inst
->tex_offsets
[0].Index
= offset
.index
;
2706 inst
->tex_offsets
[0].File
= offset
.file
;
2707 inst
->tex_offsets
[0].SwizzleX
= GET_SWZ(offset
.swizzle
, 0);
2708 inst
->tex_offsets
[0].SwizzleY
= GET_SWZ(offset
.swizzle
, 1);
2709 inst
->tex_offsets
[0].SwizzleZ
= GET_SWZ(offset
.swizzle
, 2);
2712 switch (sampler_type
->sampler_dimensionality
) {
2713 case GLSL_SAMPLER_DIM_1D
:
2714 inst
->tex_target
= (sampler_type
->sampler_array
)
2715 ? TEXTURE_1D_ARRAY_INDEX
: TEXTURE_1D_INDEX
;
2717 case GLSL_SAMPLER_DIM_2D
:
2718 inst
->tex_target
= (sampler_type
->sampler_array
)
2719 ? TEXTURE_2D_ARRAY_INDEX
: TEXTURE_2D_INDEX
;
2721 case GLSL_SAMPLER_DIM_3D
:
2722 inst
->tex_target
= TEXTURE_3D_INDEX
;
2724 case GLSL_SAMPLER_DIM_CUBE
:
2725 inst
->tex_target
= TEXTURE_CUBE_INDEX
;
2727 case GLSL_SAMPLER_DIM_RECT
:
2728 inst
->tex_target
= TEXTURE_RECT_INDEX
;
2730 case GLSL_SAMPLER_DIM_BUF
:
2731 assert(!"FINISHME: Implement ARB_texture_buffer_object");
2733 case GLSL_SAMPLER_DIM_EXTERNAL
:
2734 inst
->tex_target
= TEXTURE_EXTERNAL_INDEX
;
2737 assert(!"Should not get here.");
2740 this->result
= result_src
;
2744 glsl_to_tgsi_visitor::visit(ir_return
*ir
)
2746 if (ir
->get_value()) {
2750 assert(current_function
);
2752 ir
->get_value()->accept(this);
2753 st_src_reg r
= this->result
;
2755 l
= st_dst_reg(current_function
->return_reg
);
2757 for (i
= 0; i
< type_size(current_function
->sig
->return_type
); i
++) {
2758 emit(ir
, TGSI_OPCODE_MOV
, l
, r
);
2764 emit(ir
, TGSI_OPCODE_RET
);
2768 glsl_to_tgsi_visitor::visit(ir_discard
*ir
)
2770 struct gl_fragment_program
*fp
= (struct gl_fragment_program
*)this->prog
;
2772 if (ir
->condition
) {
2773 ir
->condition
->accept(this);
2774 this->result
.negate
= ~this->result
.negate
;
2775 emit(ir
, TGSI_OPCODE_KIL
, undef_dst
, this->result
);
2777 emit(ir
, TGSI_OPCODE_KILP
);
2780 fp
->UsesKill
= GL_TRUE
;
2784 glsl_to_tgsi_visitor::visit(ir_if
*ir
)
2786 glsl_to_tgsi_instruction
*cond_inst
, *if_inst
;
2787 glsl_to_tgsi_instruction
*prev_inst
;
2789 prev_inst
= (glsl_to_tgsi_instruction
*)this->instructions
.get_tail();
2791 ir
->condition
->accept(this);
2792 assert(this->result
.file
!= PROGRAM_UNDEFINED
);
2794 if (this->options
->EmitCondCodes
) {
2795 cond_inst
= (glsl_to_tgsi_instruction
*)this->instructions
.get_tail();
2797 /* See if we actually generated any instruction for generating
2798 * the condition. If not, then cook up a move to a temp so we
2799 * have something to set cond_update on.
2801 if (cond_inst
== prev_inst
) {
2802 st_src_reg temp
= get_temp(glsl_type::bool_type
);
2803 cond_inst
= emit(ir
->condition
, TGSI_OPCODE_MOV
, st_dst_reg(temp
), result
);
2805 cond_inst
->cond_update
= GL_TRUE
;
2807 if_inst
= emit(ir
->condition
, TGSI_OPCODE_IF
);
2808 if_inst
->dst
.cond_mask
= COND_NE
;
2810 if_inst
= emit(ir
->condition
, TGSI_OPCODE_IF
, undef_dst
, this->result
);
2813 this->instructions
.push_tail(if_inst
);
2815 visit_exec_list(&ir
->then_instructions
, this);
2817 if (!ir
->else_instructions
.is_empty()) {
2818 emit(ir
->condition
, TGSI_OPCODE_ELSE
);
2819 visit_exec_list(&ir
->else_instructions
, this);
2822 if_inst
= emit(ir
->condition
, TGSI_OPCODE_ENDIF
);
2825 glsl_to_tgsi_visitor::glsl_to_tgsi_visitor()
2827 result
.file
= PROGRAM_UNDEFINED
;
2829 next_signature_id
= 1;
2831 current_function
= NULL
;
2832 num_address_regs
= 0;
2833 indirect_addr_temps
= false;
2834 indirect_addr_consts
= false;
2835 mem_ctx
= ralloc_context(NULL
);
2838 glsl_to_tgsi_visitor::~glsl_to_tgsi_visitor()
2840 ralloc_free(mem_ctx
);
2843 extern "C" void free_glsl_to_tgsi_visitor(glsl_to_tgsi_visitor
*v
)
2850 * Count resources used by the given gpu program (number of texture
2854 count_resources(glsl_to_tgsi_visitor
*v
, gl_program
*prog
)
2856 v
->samplers_used
= 0;
2858 foreach_iter(exec_list_iterator
, iter
, v
->instructions
) {
2859 glsl_to_tgsi_instruction
*inst
= (glsl_to_tgsi_instruction
*)iter
.get();
2861 if (is_tex_instruction(inst
->op
)) {
2862 v
->samplers_used
|= 1 << inst
->sampler
;
2864 prog
->SamplerTargets
[inst
->sampler
] =
2865 (gl_texture_index
)inst
->tex_target
;
2866 if (inst
->tex_shadow
) {
2867 prog
->ShadowSamplers
|= 1 << inst
->sampler
;
2872 prog
->SamplersUsed
= v
->samplers_used
;
2873 _mesa_update_shader_textures_used(prog
);
2878 * Check if the given vertex/fragment/shader program is within the
2879 * resource limits of the context (number of texture units, etc).
2880 * If any of those checks fail, record a linker error.
2882 * XXX more checks are needed...
2885 check_resources(const struct gl_context
*ctx
,
2886 struct gl_shader_program
*shader_program
,
2887 glsl_to_tgsi_visitor
*prog
,
2888 struct gl_program
*proginfo
)
2890 switch (proginfo
->Target
) {
2891 case GL_VERTEX_PROGRAM_ARB
:
2892 if (_mesa_bitcount(prog
->samplers_used
) >
2893 ctx
->Const
.MaxVertexTextureImageUnits
) {
2894 fail_link(shader_program
, "Too many vertex shader texture samplers");
2896 if (proginfo
->Parameters
->NumParameters
> MAX_UNIFORMS
) {
2897 fail_link(shader_program
, "Too many vertex shader constants");
2900 case MESA_GEOMETRY_PROGRAM
:
2901 if (_mesa_bitcount(prog
->samplers_used
) >
2902 ctx
->Const
.MaxGeometryTextureImageUnits
) {
2903 fail_link(shader_program
, "Too many geometry shader texture samplers");
2905 if (proginfo
->Parameters
->NumParameters
>
2906 MAX_GEOMETRY_UNIFORM_COMPONENTS
/ 4) {
2907 fail_link(shader_program
, "Too many geometry shader constants");
2910 case GL_FRAGMENT_PROGRAM_ARB
:
2911 if (_mesa_bitcount(prog
->samplers_used
) >
2912 ctx
->Const
.MaxTextureImageUnits
) {
2913 fail_link(shader_program
, "Too many fragment shader texture samplers");
2915 if (proginfo
->Parameters
->NumParameters
> MAX_UNIFORMS
) {
2916 fail_link(shader_program
, "Too many fragment shader constants");
2920 _mesa_problem(ctx
, "unexpected program type in check_resources()");
2926 set_uniform_initializer(struct gl_context
*ctx
, void *mem_ctx
,
2927 struct gl_shader_program
*shader_program
,
2928 const char *name
, const glsl_type
*type
,
2931 if (type
->is_record()) {
2932 ir_constant
*field_constant
;
2934 field_constant
= (ir_constant
*)val
->components
.get_head();
2936 for (unsigned int i
= 0; i
< type
->length
; i
++) {
2937 const glsl_type
*field_type
= type
->fields
.structure
[i
].type
;
2938 const char *field_name
= ralloc_asprintf(mem_ctx
, "%s.%s", name
,
2939 type
->fields
.structure
[i
].name
);
2940 set_uniform_initializer(ctx
, mem_ctx
, shader_program
, field_name
,
2941 field_type
, field_constant
);
2942 field_constant
= (ir_constant
*)field_constant
->next
;
2947 int loc
= _mesa_get_uniform_location(ctx
, shader_program
, name
);
2950 fail_link(shader_program
,
2951 "Couldn't find uniform for initializer %s\n", name
);
2955 for (unsigned int i
= 0; i
< (type
->is_array() ? type
->length
: 1); i
++) {
2956 ir_constant
*element
;
2957 const glsl_type
*element_type
;
2958 if (type
->is_array()) {
2959 element
= val
->array_elements
[i
];
2960 element_type
= type
->fields
.array
;
2963 element_type
= type
;
2968 if (element_type
->base_type
== GLSL_TYPE_BOOL
) {
2969 int *conv
= ralloc_array(mem_ctx
, int, element_type
->components());
2970 for (unsigned int j
= 0; j
< element_type
->components(); j
++) {
2971 conv
[j
] = element
->value
.b
[j
];
2973 values
= (void *)conv
;
2974 element_type
= glsl_type::get_instance(GLSL_TYPE_INT
,
2975 element_type
->vector_elements
,
2978 values
= &element
->value
;
2981 if (element_type
->is_matrix()) {
2982 _mesa_uniform_matrix(ctx
, shader_program
,
2983 element_type
->matrix_columns
,
2984 element_type
->vector_elements
,
2985 loc
, 1, GL_FALSE
, (GLfloat
*)values
);
2987 _mesa_uniform(ctx
, shader_program
, loc
, element_type
->matrix_columns
,
2988 values
, element_type
->gl_type
);
2996 * Scan/rewrite program to remove reads of custom (output) registers.
2997 * The passed type has to be either PROGRAM_OUTPUT or PROGRAM_VARYING
2998 * (for vertex shaders).
2999 * In GLSL shaders, varying vars can be read and written.
3000 * On some hardware, trying to read an output register causes trouble.
3001 * So, rewrite the program to use a temporary register in this case.
3003 * Based on _mesa_remove_output_reads from programopt.c.
3006 glsl_to_tgsi_visitor::remove_output_reads(gl_register_file type
)
3009 GLint outputMap
[VERT_RESULT_MAX
];
3010 GLint outputTypes
[VERT_RESULT_MAX
];
3011 GLuint numVaryingReads
= 0;
3012 GLboolean
*usedTemps
;
3013 GLuint firstTemp
= 0;
3015 usedTemps
= new GLboolean
[MAX_TEMPS
];
3019 _mesa_find_used_registers(prog
, PROGRAM_TEMPORARY
,
3020 usedTemps
, MAX_TEMPS
);
3022 assert(type
== PROGRAM_VARYING
|| type
== PROGRAM_OUTPUT
);
3023 assert(prog
->Target
== GL_VERTEX_PROGRAM_ARB
|| type
!= PROGRAM_VARYING
);
3025 for (i
= 0; i
< VERT_RESULT_MAX
; i
++)
3028 /* look for instructions which read from varying vars */
3029 foreach_iter(exec_list_iterator
, iter
, this->instructions
) {
3030 glsl_to_tgsi_instruction
*inst
= (glsl_to_tgsi_instruction
*)iter
.get();
3031 const GLuint numSrc
= num_inst_src_regs(inst
->op
);
3033 for (j
= 0; j
< numSrc
; j
++) {
3034 if (inst
->src
[j
].file
== type
) {
3035 /* replace the read with a temp reg */
3036 const GLuint var
= inst
->src
[j
].index
;
3037 if (outputMap
[var
] == -1) {
3039 outputMap
[var
] = _mesa_find_free_register(usedTemps
,
3042 outputTypes
[var
] = inst
->src
[j
].type
;
3043 firstTemp
= outputMap
[var
] + 1;
3045 inst
->src
[j
].file
= PROGRAM_TEMPORARY
;
3046 inst
->src
[j
].index
= outputMap
[var
];
3051 delete [] usedTemps
;
3053 if (numVaryingReads
== 0)
3054 return; /* nothing to be done */
3056 /* look for instructions which write to the varying vars identified above */
3057 foreach_iter(exec_list_iterator
, iter
, this->instructions
) {
3058 glsl_to_tgsi_instruction
*inst
= (glsl_to_tgsi_instruction
*)iter
.get();
3059 if (inst
->dst
.file
== type
&& outputMap
[inst
->dst
.index
] >= 0) {
3060 /* change inst to write to the temp reg, instead of the varying */
3061 inst
->dst
.file
= PROGRAM_TEMPORARY
;
3062 inst
->dst
.index
= outputMap
[inst
->dst
.index
];
3066 /* insert new MOV instructions at the end */
3067 for (i
= 0; i
< VERT_RESULT_MAX
; i
++) {
3068 if (outputMap
[i
] >= 0) {
3069 /* MOV VAR[i], TEMP[tmp]; */
3070 st_src_reg src
= st_src_reg(PROGRAM_TEMPORARY
, outputMap
[i
], outputTypes
[i
]);
3071 st_dst_reg dst
= st_dst_reg(type
, WRITEMASK_XYZW
, outputTypes
[i
]);
3073 this->emit(NULL
, TGSI_OPCODE_MOV
, dst
, src
);
3079 * Returns the mask of channels (bitmask of WRITEMASK_X,Y,Z,W) which
3080 * are read from the given src in this instruction
3083 get_src_arg_mask(st_dst_reg dst
, st_src_reg src
)
3085 int read_mask
= 0, comp
;
3087 /* Now, given the src swizzle and the written channels, find which
3088 * components are actually read
3090 for (comp
= 0; comp
< 4; ++comp
) {
3091 const unsigned coord
= GET_SWZ(src
.swizzle
, comp
);
3093 if (dst
.writemask
& (1 << comp
) && coord
<= SWIZZLE_W
)
3094 read_mask
|= 1 << coord
;
3101 * This pass replaces CMP T0, T1 T2 T0 with MOV T0, T2 when the CMP
3102 * instruction is the first instruction to write to register T0. There are
3103 * several lowering passes done in GLSL IR (e.g. branches and
3104 * relative addressing) that create a large number of conditional assignments
3105 * that ir_to_mesa converts to CMP instructions like the one mentioned above.
3107 * Here is why this conversion is safe:
3108 * CMP T0, T1 T2 T0 can be expanded to:
3114 * If (T1 < 0.0) evaluates to true then our replacement MOV T0, T2 is the same
3115 * as the original program. If (T1 < 0.0) evaluates to false, executing
3116 * MOV T0, T0 will store a garbage value in T0 since T0 is uninitialized.
3117 * Therefore, it doesn't matter that we are replacing MOV T0, T0 with MOV T0, T2
3118 * because any instruction that was going to read from T0 after this was going
3119 * to read a garbage value anyway.
3122 glsl_to_tgsi_visitor::simplify_cmp(void)
3124 unsigned *tempWrites
;
3125 unsigned outputWrites
[MAX_PROGRAM_OUTPUTS
];
3127 tempWrites
= new unsigned[MAX_TEMPS
];
3131 memset(tempWrites
, 0, sizeof(tempWrites
));
3132 memset(outputWrites
, 0, sizeof(outputWrites
));
3134 foreach_iter(exec_list_iterator
, iter
, this->instructions
) {
3135 glsl_to_tgsi_instruction
*inst
= (glsl_to_tgsi_instruction
*)iter
.get();
3136 unsigned prevWriteMask
= 0;
3138 /* Give up if we encounter relative addressing or flow control. */
3139 if (inst
->dst
.reladdr
||
3140 tgsi_get_opcode_info(inst
->op
)->is_branch
||
3141 inst
->op
== TGSI_OPCODE_BGNSUB
||
3142 inst
->op
== TGSI_OPCODE_CONT
||
3143 inst
->op
== TGSI_OPCODE_END
||
3144 inst
->op
== TGSI_OPCODE_ENDSUB
||
3145 inst
->op
== TGSI_OPCODE_RET
) {
3149 if (inst
->dst
.file
== PROGRAM_OUTPUT
) {
3150 assert(inst
->dst
.index
< MAX_PROGRAM_OUTPUTS
);
3151 prevWriteMask
= outputWrites
[inst
->dst
.index
];
3152 outputWrites
[inst
->dst
.index
] |= inst
->dst
.writemask
;
3153 } else if (inst
->dst
.file
== PROGRAM_TEMPORARY
) {
3154 assert(inst
->dst
.index
< MAX_TEMPS
);
3155 prevWriteMask
= tempWrites
[inst
->dst
.index
];
3156 tempWrites
[inst
->dst
.index
] |= inst
->dst
.writemask
;
3159 /* For a CMP to be considered a conditional write, the destination
3160 * register and source register two must be the same. */
3161 if (inst
->op
== TGSI_OPCODE_CMP
3162 && !(inst
->dst
.writemask
& prevWriteMask
)
3163 && inst
->src
[2].file
== inst
->dst
.file
3164 && inst
->src
[2].index
== inst
->dst
.index
3165 && inst
->dst
.writemask
== get_src_arg_mask(inst
->dst
, inst
->src
[2])) {
3167 inst
->op
= TGSI_OPCODE_MOV
;
3168 inst
->src
[0] = inst
->src
[1];
3172 delete [] tempWrites
;
3175 /* Replaces all references to a temporary register index with another index. */
3177 glsl_to_tgsi_visitor::rename_temp_register(int index
, int new_index
)
3179 foreach_iter(exec_list_iterator
, iter
, this->instructions
) {
3180 glsl_to_tgsi_instruction
*inst
= (glsl_to_tgsi_instruction
*)iter
.get();
3183 for (j
=0; j
< num_inst_src_regs(inst
->op
); j
++) {
3184 if (inst
->src
[j
].file
== PROGRAM_TEMPORARY
&&
3185 inst
->src
[j
].index
== index
) {
3186 inst
->src
[j
].index
= new_index
;
3190 if (inst
->dst
.file
== PROGRAM_TEMPORARY
&& inst
->dst
.index
== index
) {
3191 inst
->dst
.index
= new_index
;
3197 glsl_to_tgsi_visitor::get_first_temp_read(int index
)
3199 int depth
= 0; /* loop depth */
3200 int loop_start
= -1; /* index of the first active BGNLOOP (if any) */
3203 foreach_iter(exec_list_iterator
, iter
, this->instructions
) {
3204 glsl_to_tgsi_instruction
*inst
= (glsl_to_tgsi_instruction
*)iter
.get();
3206 for (j
=0; j
< num_inst_src_regs(inst
->op
); j
++) {
3207 if (inst
->src
[j
].file
== PROGRAM_TEMPORARY
&&
3208 inst
->src
[j
].index
== index
) {
3209 return (depth
== 0) ? i
: loop_start
;
3213 if (inst
->op
== TGSI_OPCODE_BGNLOOP
) {
3216 } else if (inst
->op
== TGSI_OPCODE_ENDLOOP
) {
3229 glsl_to_tgsi_visitor::get_first_temp_write(int index
)
3231 int depth
= 0; /* loop depth */
3232 int loop_start
= -1; /* index of the first active BGNLOOP (if any) */
3235 foreach_iter(exec_list_iterator
, iter
, this->instructions
) {
3236 glsl_to_tgsi_instruction
*inst
= (glsl_to_tgsi_instruction
*)iter
.get();
3238 if (inst
->dst
.file
== PROGRAM_TEMPORARY
&& inst
->dst
.index
== index
) {
3239 return (depth
== 0) ? i
: loop_start
;
3242 if (inst
->op
== TGSI_OPCODE_BGNLOOP
) {
3245 } else if (inst
->op
== TGSI_OPCODE_ENDLOOP
) {
3258 glsl_to_tgsi_visitor::get_last_temp_read(int index
)
3260 int depth
= 0; /* loop depth */
3261 int last
= -1; /* index of last instruction that reads the temporary */
3264 foreach_iter(exec_list_iterator
, iter
, this->instructions
) {
3265 glsl_to_tgsi_instruction
*inst
= (glsl_to_tgsi_instruction
*)iter
.get();
3267 for (j
=0; j
< num_inst_src_regs(inst
->op
); j
++) {
3268 if (inst
->src
[j
].file
== PROGRAM_TEMPORARY
&&
3269 inst
->src
[j
].index
== index
) {
3270 last
= (depth
== 0) ? i
: -2;
3274 if (inst
->op
== TGSI_OPCODE_BGNLOOP
)
3276 else if (inst
->op
== TGSI_OPCODE_ENDLOOP
)
3277 if (--depth
== 0 && last
== -2)
3289 glsl_to_tgsi_visitor::get_last_temp_write(int index
)
3291 int depth
= 0; /* loop depth */
3292 int last
= -1; /* index of last instruction that writes to the temporary */
3295 foreach_iter(exec_list_iterator
, iter
, this->instructions
) {
3296 glsl_to_tgsi_instruction
*inst
= (glsl_to_tgsi_instruction
*)iter
.get();
3298 if (inst
->dst
.file
== PROGRAM_TEMPORARY
&& inst
->dst
.index
== index
)
3299 last
= (depth
== 0) ? i
: -2;
3301 if (inst
->op
== TGSI_OPCODE_BGNLOOP
)
3303 else if (inst
->op
== TGSI_OPCODE_ENDLOOP
)
3304 if (--depth
== 0 && last
== -2)
3316 * On a basic block basis, tracks available PROGRAM_TEMPORARY register
3317 * channels for copy propagation and updates following instructions to
3318 * use the original versions.
3320 * The glsl_to_tgsi_visitor lazily produces code assuming that this pass
3321 * will occur. As an example, a TXP production before this pass:
3323 * 0: MOV TEMP[1], INPUT[4].xyyy;
3324 * 1: MOV TEMP[1].w, INPUT[4].wwww;
3325 * 2: TXP TEMP[2], TEMP[1], texture[0], 2D;
3329 * 0: MOV TEMP[1], INPUT[4].xyyy;
3330 * 1: MOV TEMP[1].w, INPUT[4].wwww;
3331 * 2: TXP TEMP[2], INPUT[4].xyyw, texture[0], 2D;
3333 * which allows for dead code elimination on TEMP[1]'s writes.
3336 glsl_to_tgsi_visitor::copy_propagate(void)
3338 glsl_to_tgsi_instruction
**acp
= rzalloc_array(mem_ctx
,
3339 glsl_to_tgsi_instruction
*,
3340 this->next_temp
* 4);
3341 int *acp_level
= rzalloc_array(mem_ctx
, int, this->next_temp
* 4);
3344 foreach_iter(exec_list_iterator
, iter
, this->instructions
) {
3345 glsl_to_tgsi_instruction
*inst
= (glsl_to_tgsi_instruction
*)iter
.get();
3347 assert(inst
->dst
.file
!= PROGRAM_TEMPORARY
3348 || inst
->dst
.index
< this->next_temp
);
3350 /* First, do any copy propagation possible into the src regs. */
3351 for (int r
= 0; r
< 3; r
++) {
3352 glsl_to_tgsi_instruction
*first
= NULL
;
3354 int acp_base
= inst
->src
[r
].index
* 4;
3356 if (inst
->src
[r
].file
!= PROGRAM_TEMPORARY
||
3357 inst
->src
[r
].reladdr
)
3360 /* See if we can find entries in the ACP consisting of MOVs
3361 * from the same src register for all the swizzled channels
3362 * of this src register reference.
3364 for (int i
= 0; i
< 4; i
++) {
3365 int src_chan
= GET_SWZ(inst
->src
[r
].swizzle
, i
);
3366 glsl_to_tgsi_instruction
*copy_chan
= acp
[acp_base
+ src_chan
];
3373 assert(acp_level
[acp_base
+ src_chan
] <= level
);
3378 if (first
->src
[0].file
!= copy_chan
->src
[0].file
||
3379 first
->src
[0].index
!= copy_chan
->src
[0].index
) {
3387 /* We've now validated that we can copy-propagate to
3388 * replace this src register reference. Do it.
3390 inst
->src
[r
].file
= first
->src
[0].file
;
3391 inst
->src
[r
].index
= first
->src
[0].index
;
3394 for (int i
= 0; i
< 4; i
++) {
3395 int src_chan
= GET_SWZ(inst
->src
[r
].swizzle
, i
);
3396 glsl_to_tgsi_instruction
*copy_inst
= acp
[acp_base
+ src_chan
];
3397 swizzle
|= (GET_SWZ(copy_inst
->src
[0].swizzle
, src_chan
) <<
3400 inst
->src
[r
].swizzle
= swizzle
;
3405 case TGSI_OPCODE_BGNLOOP
:
3406 case TGSI_OPCODE_ENDLOOP
:
3407 /* End of a basic block, clear the ACP entirely. */
3408 memset(acp
, 0, sizeof(*acp
) * this->next_temp
* 4);
3411 case TGSI_OPCODE_IF
:
3415 case TGSI_OPCODE_ENDIF
:
3416 case TGSI_OPCODE_ELSE
:
3417 /* Clear all channels written inside the block from the ACP, but
3418 * leaving those that were not touched.
3420 for (int r
= 0; r
< this->next_temp
; r
++) {
3421 for (int c
= 0; c
< 4; c
++) {
3422 if (!acp
[4 * r
+ c
])
3425 if (acp_level
[4 * r
+ c
] >= level
)
3426 acp
[4 * r
+ c
] = NULL
;
3429 if (inst
->op
== TGSI_OPCODE_ENDIF
)
3434 /* Continuing the block, clear any written channels from
3437 if (inst
->dst
.file
== PROGRAM_TEMPORARY
&& inst
->dst
.reladdr
) {
3438 /* Any temporary might be written, so no copy propagation
3439 * across this instruction.
3441 memset(acp
, 0, sizeof(*acp
) * this->next_temp
* 4);
3442 } else if (inst
->dst
.file
== PROGRAM_OUTPUT
&&
3443 inst
->dst
.reladdr
) {
3444 /* Any output might be written, so no copy propagation
3445 * from outputs across this instruction.
3447 for (int r
= 0; r
< this->next_temp
; r
++) {
3448 for (int c
= 0; c
< 4; c
++) {
3449 if (!acp
[4 * r
+ c
])
3452 if (acp
[4 * r
+ c
]->src
[0].file
== PROGRAM_OUTPUT
)
3453 acp
[4 * r
+ c
] = NULL
;
3456 } else if (inst
->dst
.file
== PROGRAM_TEMPORARY
||
3457 inst
->dst
.file
== PROGRAM_OUTPUT
) {
3458 /* Clear where it's used as dst. */
3459 if (inst
->dst
.file
== PROGRAM_TEMPORARY
) {
3460 for (int c
= 0; c
< 4; c
++) {
3461 if (inst
->dst
.writemask
& (1 << c
)) {
3462 acp
[4 * inst
->dst
.index
+ c
] = NULL
;
3467 /* Clear where it's used as src. */
3468 for (int r
= 0; r
< this->next_temp
; r
++) {
3469 for (int c
= 0; c
< 4; c
++) {
3470 if (!acp
[4 * r
+ c
])
3473 int src_chan
= GET_SWZ(acp
[4 * r
+ c
]->src
[0].swizzle
, c
);
3475 if (acp
[4 * r
+ c
]->src
[0].file
== inst
->dst
.file
&&
3476 acp
[4 * r
+ c
]->src
[0].index
== inst
->dst
.index
&&
3477 inst
->dst
.writemask
& (1 << src_chan
))
3479 acp
[4 * r
+ c
] = NULL
;
3487 /* If this is a copy, add it to the ACP. */
3488 if (inst
->op
== TGSI_OPCODE_MOV
&&
3489 inst
->dst
.file
== PROGRAM_TEMPORARY
&&
3490 !inst
->dst
.reladdr
&&
3492 !inst
->src
[0].reladdr
&&
3493 !inst
->src
[0].negate
) {
3494 for (int i
= 0; i
< 4; i
++) {
3495 if (inst
->dst
.writemask
& (1 << i
)) {
3496 acp
[4 * inst
->dst
.index
+ i
] = inst
;
3497 acp_level
[4 * inst
->dst
.index
+ i
] = level
;
3503 ralloc_free(acp_level
);
3508 * Tracks available PROGRAM_TEMPORARY registers for dead code elimination.
3510 * The glsl_to_tgsi_visitor lazily produces code assuming that this pass
3511 * will occur. As an example, a TXP production after copy propagation but
3514 * 0: MOV TEMP[1], INPUT[4].xyyy;
3515 * 1: MOV TEMP[1].w, INPUT[4].wwww;
3516 * 2: TXP TEMP[2], INPUT[4].xyyw, texture[0], 2D;
3518 * and after this pass:
3520 * 0: TXP TEMP[2], INPUT[4].xyyw, texture[0], 2D;
3522 * FIXME: assumes that all functions are inlined (no support for BGNSUB/ENDSUB)
3523 * FIXME: doesn't eliminate all dead code inside of loops; it steps around them
3526 glsl_to_tgsi_visitor::eliminate_dead_code(void)
3530 for (i
=0; i
< this->next_temp
; i
++) {
3531 int last_read
= get_last_temp_read(i
);
3534 foreach_iter(exec_list_iterator
, iter
, this->instructions
) {
3535 glsl_to_tgsi_instruction
*inst
= (glsl_to_tgsi_instruction
*)iter
.get();
3537 if (inst
->dst
.file
== PROGRAM_TEMPORARY
&& inst
->dst
.index
== i
&&
3550 * On a basic block basis, tracks available PROGRAM_TEMPORARY registers for dead
3551 * code elimination. This is less primitive than eliminate_dead_code(), as it
3552 * is per-channel and can detect consecutive writes without a read between them
3553 * as dead code. However, there is some dead code that can be eliminated by
3554 * eliminate_dead_code() but not this function - for example, this function
3555 * cannot eliminate an instruction writing to a register that is never read and
3556 * is the only instruction writing to that register.
3558 * The glsl_to_tgsi_visitor lazily produces code assuming that this pass
3562 glsl_to_tgsi_visitor::eliminate_dead_code_advanced(void)
3564 glsl_to_tgsi_instruction
**writes
= rzalloc_array(mem_ctx
,
3565 glsl_to_tgsi_instruction
*,
3566 this->next_temp
* 4);
3567 int *write_level
= rzalloc_array(mem_ctx
, int, this->next_temp
* 4);
3571 foreach_iter(exec_list_iterator
, iter
, this->instructions
) {
3572 glsl_to_tgsi_instruction
*inst
= (glsl_to_tgsi_instruction
*)iter
.get();
3574 assert(inst
->dst
.file
!= PROGRAM_TEMPORARY
3575 || inst
->dst
.index
< this->next_temp
);
3578 case TGSI_OPCODE_BGNLOOP
:
3579 case TGSI_OPCODE_ENDLOOP
:
3580 /* End of a basic block, clear the write array entirely.
3581 * FIXME: This keeps us from killing dead code when the writes are
3582 * on either side of a loop, even when the register isn't touched
3585 memset(writes
, 0, sizeof(*writes
) * this->next_temp
* 4);
3588 case TGSI_OPCODE_ENDIF
:
3592 case TGSI_OPCODE_ELSE
:
3593 /* Clear all channels written inside the preceding if block from the
3594 * write array, but leave those that were not touched.
3596 * FIXME: This destroys opportunities to remove dead code inside of
3597 * IF blocks that are followed by an ELSE block.
3599 for (int r
= 0; r
< this->next_temp
; r
++) {
3600 for (int c
= 0; c
< 4; c
++) {
3601 if (!writes
[4 * r
+ c
])
3604 if (write_level
[4 * r
+ c
] >= level
)
3605 writes
[4 * r
+ c
] = NULL
;
3610 case TGSI_OPCODE_IF
:
3612 /* fallthrough to default case to mark the condition as read */
3615 /* Continuing the block, clear any channels from the write array that
3616 * are read by this instruction.
3618 for (unsigned i
= 0; i
< Elements(inst
->src
); i
++) {
3619 if (inst
->src
[i
].file
== PROGRAM_TEMPORARY
&& inst
->src
[i
].reladdr
){
3620 /* Any temporary might be read, so no dead code elimination
3621 * across this instruction.
3623 memset(writes
, 0, sizeof(*writes
) * this->next_temp
* 4);
3624 } else if (inst
->src
[i
].file
== PROGRAM_TEMPORARY
) {
3625 /* Clear where it's used as src. */
3626 int src_chans
= 1 << GET_SWZ(inst
->src
[i
].swizzle
, 0);
3627 src_chans
|= 1 << GET_SWZ(inst
->src
[i
].swizzle
, 1);
3628 src_chans
|= 1 << GET_SWZ(inst
->src
[i
].swizzle
, 2);
3629 src_chans
|= 1 << GET_SWZ(inst
->src
[i
].swizzle
, 3);
3631 for (int c
= 0; c
< 4; c
++) {
3632 if (src_chans
& (1 << c
)) {
3633 writes
[4 * inst
->src
[i
].index
+ c
] = NULL
;
3641 /* If this instruction writes to a temporary, add it to the write array.
3642 * If there is already an instruction in the write array for one or more
3643 * of the channels, flag that channel write as dead.
3645 if (inst
->dst
.file
== PROGRAM_TEMPORARY
&&
3646 !inst
->dst
.reladdr
&&
3648 for (int c
= 0; c
< 4; c
++) {
3649 if (inst
->dst
.writemask
& (1 << c
)) {
3650 if (writes
[4 * inst
->dst
.index
+ c
]) {
3651 if (write_level
[4 * inst
->dst
.index
+ c
] < level
)
3654 writes
[4 * inst
->dst
.index
+ c
]->dead_mask
|= (1 << c
);
3656 writes
[4 * inst
->dst
.index
+ c
] = inst
;
3657 write_level
[4 * inst
->dst
.index
+ c
] = level
;
3663 /* Anything still in the write array at this point is dead code. */
3664 for (int r
= 0; r
< this->next_temp
; r
++) {
3665 for (int c
= 0; c
< 4; c
++) {
3666 glsl_to_tgsi_instruction
*inst
= writes
[4 * r
+ c
];
3668 inst
->dead_mask
|= (1 << c
);
3672 /* Now actually remove the instructions that are completely dead and update
3673 * the writemask of other instructions with dead channels.
3675 foreach_iter(exec_list_iterator
, iter
, this->instructions
) {
3676 glsl_to_tgsi_instruction
*inst
= (glsl_to_tgsi_instruction
*)iter
.get();
3678 if (!inst
->dead_mask
|| !inst
->dst
.writemask
)
3680 else if ((inst
->dst
.writemask
& ~inst
->dead_mask
) == 0) {
3685 inst
->dst
.writemask
&= ~(inst
->dead_mask
);
3688 ralloc_free(write_level
);
3689 ralloc_free(writes
);
3694 /* Merges temporary registers together where possible to reduce the number of
3695 * registers needed to run a program.
3697 * Produces optimal code only after copy propagation and dead code elimination
3700 glsl_to_tgsi_visitor::merge_registers(void)
3702 int *last_reads
= rzalloc_array(mem_ctx
, int, this->next_temp
);
3703 int *first_writes
= rzalloc_array(mem_ctx
, int, this->next_temp
);
3706 /* Read the indices of the last read and first write to each temp register
3707 * into an array so that we don't have to traverse the instruction list as
3709 for (i
=0; i
< this->next_temp
; i
++) {
3710 last_reads
[i
] = get_last_temp_read(i
);
3711 first_writes
[i
] = get_first_temp_write(i
);
3714 /* Start looking for registers with non-overlapping usages that can be
3715 * merged together. */
3716 for (i
=0; i
< this->next_temp
; i
++) {
3717 /* Don't touch unused registers. */
3718 if (last_reads
[i
] < 0 || first_writes
[i
] < 0) continue;
3720 for (j
=0; j
< this->next_temp
; j
++) {
3721 /* Don't touch unused registers. */
3722 if (last_reads
[j
] < 0 || first_writes
[j
] < 0) continue;
3724 /* We can merge the two registers if the first write to j is after or
3725 * in the same instruction as the last read from i. Note that the
3726 * register at index i will always be used earlier or at the same time
3727 * as the register at index j. */
3728 if (first_writes
[i
] <= first_writes
[j
] &&
3729 last_reads
[i
] <= first_writes
[j
])
3731 rename_temp_register(j
, i
); /* Replace all references to j with i.*/
3733 /* Update the first_writes and last_reads arrays with the new
3734 * values for the merged register index, and mark the newly unused
3735 * register index as such. */
3736 last_reads
[i
] = last_reads
[j
];
3737 first_writes
[j
] = -1;
3743 ralloc_free(last_reads
);
3744 ralloc_free(first_writes
);
3747 /* Reassign indices to temporary registers by reusing unused indices created
3748 * by optimization passes. */
3750 glsl_to_tgsi_visitor::renumber_registers(void)
3755 for (i
=0; i
< this->next_temp
; i
++) {
3756 if (get_first_temp_read(i
) < 0) continue;
3758 rename_temp_register(i
, new_index
);
3762 this->next_temp
= new_index
;
3766 * Returns a fragment program which implements the current pixel transfer ops.
3767 * Based on get_pixel_transfer_program in st_atom_pixeltransfer.c.
3770 get_pixel_transfer_visitor(struct st_fragment_program
*fp
,
3771 glsl_to_tgsi_visitor
*original
,
3772 int scale_and_bias
, int pixel_maps
)
3774 glsl_to_tgsi_visitor
*v
= new glsl_to_tgsi_visitor();
3775 struct st_context
*st
= st_context(original
->ctx
);
3776 struct gl_program
*prog
= &fp
->Base
.Base
;
3777 struct gl_program_parameter_list
*params
= _mesa_new_parameter_list();
3778 st_src_reg coord
, src0
;
3780 glsl_to_tgsi_instruction
*inst
;
3782 /* Copy attributes of the glsl_to_tgsi_visitor in the original shader. */
3783 v
->ctx
= original
->ctx
;
3785 v
->glsl_version
= original
->glsl_version
;
3786 v
->native_integers
= original
->native_integers
;
3787 v
->options
= original
->options
;
3788 v
->next_temp
= original
->next_temp
;
3789 v
->num_address_regs
= original
->num_address_regs
;
3790 v
->samplers_used
= prog
->SamplersUsed
= original
->samplers_used
;
3791 v
->indirect_addr_temps
= original
->indirect_addr_temps
;
3792 v
->indirect_addr_consts
= original
->indirect_addr_consts
;
3793 memcpy(&v
->immediates
, &original
->immediates
, sizeof(v
->immediates
));
3796 * Get initial pixel color from the texture.
3797 * TEX colorTemp, fragment.texcoord[0], texture[0], 2D;
3799 coord
= st_src_reg(PROGRAM_INPUT
, FRAG_ATTRIB_TEX0
, glsl_type::vec2_type
);
3800 src0
= v
->get_temp(glsl_type::vec4_type
);
3801 dst0
= st_dst_reg(src0
);
3802 inst
= v
->emit(NULL
, TGSI_OPCODE_TEX
, dst0
, coord
);
3804 inst
->tex_target
= TEXTURE_2D_INDEX
;
3806 prog
->InputsRead
|= (1 << FRAG_ATTRIB_TEX0
);
3807 prog
->SamplersUsed
|= (1 << 0); /* mark sampler 0 as used */
3808 v
->samplers_used
|= (1 << 0);
3810 if (scale_and_bias
) {
3811 static const gl_state_index scale_state
[STATE_LENGTH
] =
3812 { STATE_INTERNAL
, STATE_PT_SCALE
,
3813 (gl_state_index
) 0, (gl_state_index
) 0, (gl_state_index
) 0 };
3814 static const gl_state_index bias_state
[STATE_LENGTH
] =
3815 { STATE_INTERNAL
, STATE_PT_BIAS
,
3816 (gl_state_index
) 0, (gl_state_index
) 0, (gl_state_index
) 0 };
3817 GLint scale_p
, bias_p
;
3818 st_src_reg scale
, bias
;
3820 scale_p
= _mesa_add_state_reference(params
, scale_state
);
3821 bias_p
= _mesa_add_state_reference(params
, bias_state
);
3823 /* MAD colorTemp, colorTemp, scale, bias; */
3824 scale
= st_src_reg(PROGRAM_STATE_VAR
, scale_p
, GLSL_TYPE_FLOAT
);
3825 bias
= st_src_reg(PROGRAM_STATE_VAR
, bias_p
, GLSL_TYPE_FLOAT
);
3826 inst
= v
->emit(NULL
, TGSI_OPCODE_MAD
, dst0
, src0
, scale
, bias
);
3830 st_src_reg temp
= v
->get_temp(glsl_type::vec4_type
);
3831 st_dst_reg temp_dst
= st_dst_reg(temp
);
3833 assert(st
->pixel_xfer
.pixelmap_texture
);
3835 /* With a little effort, we can do four pixel map look-ups with
3836 * two TEX instructions:
3839 /* TEX temp.rg, colorTemp.rgba, texture[1], 2D; */
3840 temp_dst
.writemask
= WRITEMASK_XY
; /* write R,G */
3841 inst
= v
->emit(NULL
, TGSI_OPCODE_TEX
, temp_dst
, src0
);
3843 inst
->tex_target
= TEXTURE_2D_INDEX
;
3845 /* TEX temp.ba, colorTemp.baba, texture[1], 2D; */
3846 src0
.swizzle
= MAKE_SWIZZLE4(SWIZZLE_Z
, SWIZZLE_W
, SWIZZLE_Z
, SWIZZLE_W
);
3847 temp_dst
.writemask
= WRITEMASK_ZW
; /* write B,A */
3848 inst
= v
->emit(NULL
, TGSI_OPCODE_TEX
, temp_dst
, src0
);
3850 inst
->tex_target
= TEXTURE_2D_INDEX
;
3852 prog
->SamplersUsed
|= (1 << 1); /* mark sampler 1 as used */
3853 v
->samplers_used
|= (1 << 1);
3855 /* MOV colorTemp, temp; */
3856 inst
= v
->emit(NULL
, TGSI_OPCODE_MOV
, dst0
, temp
);
3859 /* Now copy the instructions from the original glsl_to_tgsi_visitor into the
3861 foreach_iter(exec_list_iterator
, iter
, original
->instructions
) {
3862 glsl_to_tgsi_instruction
*inst
= (glsl_to_tgsi_instruction
*)iter
.get();
3863 st_src_reg src_regs
[3];
3865 if (inst
->dst
.file
== PROGRAM_OUTPUT
)
3866 prog
->OutputsWritten
|= BITFIELD64_BIT(inst
->dst
.index
);
3868 for (int i
=0; i
<3; i
++) {
3869 src_regs
[i
] = inst
->src
[i
];
3870 if (src_regs
[i
].file
== PROGRAM_INPUT
&&
3871 src_regs
[i
].index
== FRAG_ATTRIB_COL0
)
3873 src_regs
[i
].file
= PROGRAM_TEMPORARY
;
3874 src_regs
[i
].index
= src0
.index
;
3876 else if (src_regs
[i
].file
== PROGRAM_INPUT
)
3877 prog
->InputsRead
|= (1 << src_regs
[i
].index
);
3880 v
->emit(NULL
, inst
->op
, inst
->dst
, src_regs
[0], src_regs
[1], src_regs
[2]);
3883 /* Make modifications to fragment program info. */
3884 prog
->Parameters
= _mesa_combine_parameter_lists(params
,
3885 original
->prog
->Parameters
);
3886 _mesa_free_parameter_list(params
);
3887 count_resources(v
, prog
);
3888 fp
->glsl_to_tgsi
= v
;
3892 * Make fragment program for glBitmap:
3893 * Sample the texture and kill the fragment if the bit is 0.
3894 * This program will be combined with the user's fragment program.
3896 * Based on make_bitmap_fragment_program in st_cb_bitmap.c.
3899 get_bitmap_visitor(struct st_fragment_program
*fp
,
3900 glsl_to_tgsi_visitor
*original
, int samplerIndex
)
3902 glsl_to_tgsi_visitor
*v
= new glsl_to_tgsi_visitor();
3903 struct st_context
*st
= st_context(original
->ctx
);
3904 struct gl_program
*prog
= &fp
->Base
.Base
;
3905 st_src_reg coord
, src0
;
3907 glsl_to_tgsi_instruction
*inst
;
3909 /* Copy attributes of the glsl_to_tgsi_visitor in the original shader. */
3910 v
->ctx
= original
->ctx
;
3912 v
->glsl_version
= original
->glsl_version
;
3913 v
->native_integers
= original
->native_integers
;
3914 v
->options
= original
->options
;
3915 v
->next_temp
= original
->next_temp
;
3916 v
->num_address_regs
= original
->num_address_regs
;
3917 v
->samplers_used
= prog
->SamplersUsed
= original
->samplers_used
;
3918 v
->indirect_addr_temps
= original
->indirect_addr_temps
;
3919 v
->indirect_addr_consts
= original
->indirect_addr_consts
;
3920 memcpy(&v
->immediates
, &original
->immediates
, sizeof(v
->immediates
));
3922 /* TEX tmp0, fragment.texcoord[0], texture[0], 2D; */
3923 coord
= st_src_reg(PROGRAM_INPUT
, FRAG_ATTRIB_TEX0
, glsl_type::vec2_type
);
3924 src0
= v
->get_temp(glsl_type::vec4_type
);
3925 dst0
= st_dst_reg(src0
);
3926 inst
= v
->emit(NULL
, TGSI_OPCODE_TEX
, dst0
, coord
);
3927 inst
->sampler
= samplerIndex
;
3928 inst
->tex_target
= TEXTURE_2D_INDEX
;
3930 prog
->InputsRead
|= (1 << FRAG_ATTRIB_TEX0
);
3931 prog
->SamplersUsed
|= (1 << samplerIndex
); /* mark sampler as used */
3932 v
->samplers_used
|= (1 << samplerIndex
);
3934 /* KIL if -tmp0 < 0 # texel=0 -> keep / texel=0 -> discard */
3935 src0
.negate
= NEGATE_XYZW
;
3936 if (st
->bitmap
.tex_format
== PIPE_FORMAT_L8_UNORM
)
3937 src0
.swizzle
= SWIZZLE_XXXX
;
3938 inst
= v
->emit(NULL
, TGSI_OPCODE_KIL
, undef_dst
, src0
);
3940 /* Now copy the instructions from the original glsl_to_tgsi_visitor into the
3942 foreach_iter(exec_list_iterator
, iter
, original
->instructions
) {
3943 glsl_to_tgsi_instruction
*inst
= (glsl_to_tgsi_instruction
*)iter
.get();
3944 st_src_reg src_regs
[3];
3946 if (inst
->dst
.file
== PROGRAM_OUTPUT
)
3947 prog
->OutputsWritten
|= BITFIELD64_BIT(inst
->dst
.index
);
3949 for (int i
=0; i
<3; i
++) {
3950 src_regs
[i
] = inst
->src
[i
];
3951 if (src_regs
[i
].file
== PROGRAM_INPUT
)
3952 prog
->InputsRead
|= (1 << src_regs
[i
].index
);
3955 v
->emit(NULL
, inst
->op
, inst
->dst
, src_regs
[0], src_regs
[1], src_regs
[2]);
3958 /* Make modifications to fragment program info. */
3959 prog
->Parameters
= _mesa_clone_parameter_list(original
->prog
->Parameters
);
3960 count_resources(v
, prog
);
3961 fp
->glsl_to_tgsi
= v
;
3964 /* ------------------------- TGSI conversion stuff -------------------------- */
3966 unsigned branch_target
;
3971 * Intermediate state used during shader translation.
3973 struct st_translate
{
3974 struct ureg_program
*ureg
;
3976 struct ureg_dst temps
[MAX_TEMPS
];
3977 struct ureg_src
*constants
;
3978 struct ureg_src
*immediates
;
3979 struct ureg_dst outputs
[PIPE_MAX_SHADER_OUTPUTS
];
3980 struct ureg_src inputs
[PIPE_MAX_SHADER_INPUTS
];
3981 struct ureg_dst address
[1];
3982 struct ureg_src samplers
[PIPE_MAX_SAMPLERS
];
3983 struct ureg_src systemValues
[SYSTEM_VALUE_MAX
];
3985 /* Extra info for handling point size clamping in vertex shader */
3986 struct ureg_dst pointSizeResult
; /**< Actual point size output register */
3987 struct ureg_src pointSizeConst
; /**< Point size range constant register */
3988 GLint pointSizeOutIndex
; /**< Temp point size output register */
3989 GLboolean prevInstWrotePointSize
;
3991 const GLuint
*inputMapping
;
3992 const GLuint
*outputMapping
;
3994 /* For every instruction that contains a label (eg CALL), keep
3995 * details so that we can go back afterwards and emit the correct
3996 * tgsi instruction number for each label.
3998 struct label
*labels
;
3999 unsigned labels_size
;
4000 unsigned labels_count
;
4002 /* Keep a record of the tgsi instruction number that each mesa
4003 * instruction starts at, will be used to fix up labels after
4008 unsigned insn_count
;
4010 unsigned procType
; /**< TGSI_PROCESSOR_VERTEX/FRAGMENT */
4015 /** Map Mesa's SYSTEM_VALUE_x to TGSI_SEMANTIC_x */
4016 static unsigned mesa_sysval_to_semantic
[SYSTEM_VALUE_MAX
] = {
4018 TGSI_SEMANTIC_INSTANCEID
4022 * Make note of a branch to a label in the TGSI code.
4023 * After we've emitted all instructions, we'll go over the list
4024 * of labels built here and patch the TGSI code with the actual
4025 * location of each label.
4027 static unsigned *get_label(struct st_translate
*t
, unsigned branch_target
)
4031 if (t
->labels_count
+ 1 >= t
->labels_size
) {
4032 t
->labels_size
= 1 << (util_logbase2(t
->labels_size
) + 1);
4033 t
->labels
= (struct label
*)realloc(t
->labels
,
4034 t
->labels_size
* sizeof(struct label
));
4035 if (t
->labels
== NULL
) {
4036 static unsigned dummy
;
4042 i
= t
->labels_count
++;
4043 t
->labels
[i
].branch_target
= branch_target
;
4044 return &t
->labels
[i
].token
;
4048 * Called prior to emitting the TGSI code for each instruction.
4049 * Allocate additional space for instructions if needed.
4050 * Update the insn[] array so the next glsl_to_tgsi_instruction points to
4051 * the next TGSI instruction.
4053 static void set_insn_start(struct st_translate
*t
, unsigned start
)
4055 if (t
->insn_count
+ 1 >= t
->insn_size
) {
4056 t
->insn_size
= 1 << (util_logbase2(t
->insn_size
) + 1);
4057 t
->insn
= (unsigned *)realloc(t
->insn
, t
->insn_size
* sizeof(t
->insn
[0]));
4058 if (t
->insn
== NULL
) {
4064 t
->insn
[t
->insn_count
++] = start
;
4068 * Map a glsl_to_tgsi constant/immediate to a TGSI immediate.
4070 static struct ureg_src
4071 emit_immediate(struct st_translate
*t
,
4072 gl_constant_value values
[4],
4075 struct ureg_program
*ureg
= t
->ureg
;
4080 return ureg_DECL_immediate(ureg
, &values
[0].f
, size
);
4082 return ureg_DECL_immediate_int(ureg
, &values
[0].i
, size
);
4083 case GL_UNSIGNED_INT
:
4085 return ureg_DECL_immediate_uint(ureg
, &values
[0].u
, size
);
4087 assert(!"should not get here - type must be float, int, uint, or bool");
4088 return ureg_src_undef();
4093 * Map a glsl_to_tgsi dst register to a TGSI ureg_dst register.
4095 static struct ureg_dst
4096 dst_register(struct st_translate
*t
,
4097 gl_register_file file
,
4101 case PROGRAM_UNDEFINED
:
4102 return ureg_dst_undef();
4104 case PROGRAM_TEMPORARY
:
4105 if (ureg_dst_is_undef(t
->temps
[index
]))
4106 t
->temps
[index
] = ureg_DECL_temporary(t
->ureg
);
4108 return t
->temps
[index
];
4110 case PROGRAM_OUTPUT
:
4111 if (t
->procType
== TGSI_PROCESSOR_VERTEX
&& index
== VERT_RESULT_PSIZ
)
4112 t
->prevInstWrotePointSize
= GL_TRUE
;
4114 if (t
->procType
== TGSI_PROCESSOR_VERTEX
)
4115 assert(index
< VERT_RESULT_MAX
);
4116 else if (t
->procType
== TGSI_PROCESSOR_FRAGMENT
)
4117 assert(index
< FRAG_RESULT_MAX
);
4119 assert(index
< GEOM_RESULT_MAX
);
4121 assert(t
->outputMapping
[index
] < Elements(t
->outputs
));
4123 return t
->outputs
[t
->outputMapping
[index
]];
4125 case PROGRAM_ADDRESS
:
4126 return t
->address
[index
];
4129 assert(!"unknown dst register file");
4130 return ureg_dst_undef();
4135 * Map a glsl_to_tgsi src register to a TGSI ureg_src register.
4137 static struct ureg_src
4138 src_register(struct st_translate
*t
,
4139 gl_register_file file
,
4143 case PROGRAM_UNDEFINED
:
4144 return ureg_src_undef();
4146 case PROGRAM_TEMPORARY
:
4148 assert(index
< Elements(t
->temps
));
4149 if (ureg_dst_is_undef(t
->temps
[index
]))
4150 t
->temps
[index
] = ureg_DECL_temporary(t
->ureg
);
4151 return ureg_src(t
->temps
[index
]);
4153 case PROGRAM_NAMED_PARAM
:
4154 case PROGRAM_ENV_PARAM
:
4155 case PROGRAM_LOCAL_PARAM
:
4156 case PROGRAM_UNIFORM
:
4158 return t
->constants
[index
];
4159 case PROGRAM_STATE_VAR
:
4160 case PROGRAM_CONSTANT
: /* ie, immediate */
4162 return ureg_DECL_constant(t
->ureg
, 0);
4164 return t
->constants
[index
];
4166 case PROGRAM_IMMEDIATE
:
4167 return t
->immediates
[index
];
4170 assert(t
->inputMapping
[index
] < Elements(t
->inputs
));
4171 return t
->inputs
[t
->inputMapping
[index
]];
4173 case PROGRAM_OUTPUT
:
4174 assert(t
->outputMapping
[index
] < Elements(t
->outputs
));
4175 return ureg_src(t
->outputs
[t
->outputMapping
[index
]]); /* not needed? */
4177 case PROGRAM_ADDRESS
:
4178 return ureg_src(t
->address
[index
]);
4180 case PROGRAM_SYSTEM_VALUE
:
4181 assert(index
< Elements(t
->systemValues
));
4182 return t
->systemValues
[index
];
4185 assert(!"unknown src register file");
4186 return ureg_src_undef();
4191 * Create a TGSI ureg_dst register from an st_dst_reg.
4193 static struct ureg_dst
4194 translate_dst(struct st_translate
*t
,
4195 const st_dst_reg
*dst_reg
,
4198 struct ureg_dst dst
= dst_register(t
,
4202 dst
= ureg_writemask(dst
, dst_reg
->writemask
);
4205 dst
= ureg_saturate(dst
);
4207 if (dst_reg
->reladdr
!= NULL
)
4208 dst
= ureg_dst_indirect(dst
, ureg_src(t
->address
[0]));
4214 * Create a TGSI ureg_src register from an st_src_reg.
4216 static struct ureg_src
4217 translate_src(struct st_translate
*t
, const st_src_reg
*src_reg
)
4219 struct ureg_src src
= src_register(t
, src_reg
->file
, src_reg
->index
);
4221 src
= ureg_swizzle(src
,
4222 GET_SWZ(src_reg
->swizzle
, 0) & 0x3,
4223 GET_SWZ(src_reg
->swizzle
, 1) & 0x3,
4224 GET_SWZ(src_reg
->swizzle
, 2) & 0x3,
4225 GET_SWZ(src_reg
->swizzle
, 3) & 0x3);
4227 if ((src_reg
->negate
& 0xf) == NEGATE_XYZW
)
4228 src
= ureg_negate(src
);
4230 if (src_reg
->reladdr
!= NULL
) {
4231 /* Normally ureg_src_indirect() would be used here, but a stupid compiler
4232 * bug in g++ makes ureg_src_indirect (an inline C function) erroneously
4233 * set the bit for src.Negate. So we have to do the operation manually
4234 * here to work around the compiler's problems. */
4235 /*src = ureg_src_indirect(src, ureg_src(t->address[0]));*/
4236 struct ureg_src addr
= ureg_src(t
->address
[0]);
4238 src
.IndirectFile
= addr
.File
;
4239 src
.IndirectIndex
= addr
.Index
;
4240 src
.IndirectSwizzle
= addr
.SwizzleX
;
4242 if (src_reg
->file
!= PROGRAM_INPUT
&&
4243 src_reg
->file
!= PROGRAM_OUTPUT
) {
4244 /* If src_reg->index was negative, it was set to zero in
4245 * src_register(). Reassign it now. But don't do this
4246 * for input/output regs since they get remapped while
4247 * const buffers don't.
4249 src
.Index
= src_reg
->index
;
4256 static struct tgsi_texture_offset
4257 translate_tex_offset(struct st_translate
*t
,
4258 const struct tgsi_texture_offset
*in_offset
)
4260 struct tgsi_texture_offset offset
;
4262 assert(in_offset
->File
== PROGRAM_IMMEDIATE
);
4264 offset
.File
= TGSI_FILE_IMMEDIATE
;
4265 offset
.Index
= in_offset
->Index
;
4266 offset
.SwizzleX
= in_offset
->SwizzleX
;
4267 offset
.SwizzleY
= in_offset
->SwizzleY
;
4268 offset
.SwizzleZ
= in_offset
->SwizzleZ
;
4274 compile_tgsi_instruction(struct st_translate
*t
,
4275 const glsl_to_tgsi_instruction
*inst
)
4277 struct ureg_program
*ureg
= t
->ureg
;
4279 struct ureg_dst dst
[1];
4280 struct ureg_src src
[4];
4281 struct tgsi_texture_offset texoffsets
[MAX_GLSL_TEXTURE_OFFSET
];
4286 num_dst
= num_inst_dst_regs(inst
->op
);
4287 num_src
= num_inst_src_regs(inst
->op
);
4290 dst
[0] = translate_dst(t
,
4294 for (i
= 0; i
< num_src
; i
++)
4295 src
[i
] = translate_src(t
, &inst
->src
[i
]);
4298 case TGSI_OPCODE_BGNLOOP
:
4299 case TGSI_OPCODE_CAL
:
4300 case TGSI_OPCODE_ELSE
:
4301 case TGSI_OPCODE_ENDLOOP
:
4302 case TGSI_OPCODE_IF
:
4303 assert(num_dst
== 0);
4304 ureg_label_insn(ureg
,
4308 inst
->op
== TGSI_OPCODE_CAL
? inst
->function
->sig_id
: 0));
4311 case TGSI_OPCODE_TEX
:
4312 case TGSI_OPCODE_TXB
:
4313 case TGSI_OPCODE_TXD
:
4314 case TGSI_OPCODE_TXL
:
4315 case TGSI_OPCODE_TXP
:
4316 case TGSI_OPCODE_TXQ
:
4317 case TGSI_OPCODE_TXF
:
4318 src
[num_src
++] = t
->samplers
[inst
->sampler
];
4319 for (i
= 0; i
< inst
->tex_offset_num_offset
; i
++) {
4320 texoffsets
[i
] = translate_tex_offset(t
, &inst
->tex_offsets
[i
]);
4325 translate_texture_target(inst
->tex_target
, inst
->tex_shadow
),
4326 texoffsets
, inst
->tex_offset_num_offset
,
4330 case TGSI_OPCODE_SCS
:
4331 dst
[0] = ureg_writemask(dst
[0], TGSI_WRITEMASK_XY
);
4332 ureg_insn(ureg
, inst
->op
, dst
, num_dst
, src
, num_src
);
4345 * Emit the TGSI instructions for inverting and adjusting WPOS.
4346 * This code is unavoidable because it also depends on whether
4347 * a FBO is bound (STATE_FB_WPOS_Y_TRANSFORM).
4350 emit_wpos_adjustment( struct st_translate
*t
,
4351 const struct gl_program
*program
,
4353 GLfloat adjX
, GLfloat adjY
[2])
4355 struct ureg_program
*ureg
= t
->ureg
;
4357 /* Fragment program uses fragment position input.
4358 * Need to replace instances of INPUT[WPOS] with temp T
4359 * where T = INPUT[WPOS] by y is inverted.
4361 static const gl_state_index wposTransformState
[STATE_LENGTH
]
4362 = { STATE_INTERNAL
, STATE_FB_WPOS_Y_TRANSFORM
,
4363 (gl_state_index
)0, (gl_state_index
)0, (gl_state_index
)0 };
4365 /* XXX: note we are modifying the incoming shader here! Need to
4366 * do this before emitting the constant decls below, or this
4369 unsigned wposTransConst
= _mesa_add_state_reference(program
->Parameters
,
4370 wposTransformState
);
4372 struct ureg_src wpostrans
= ureg_DECL_constant( ureg
, wposTransConst
);
4373 struct ureg_dst wpos_temp
= ureg_DECL_temporary( ureg
);
4374 struct ureg_src wpos_input
= t
->inputs
[t
->inputMapping
[FRAG_ATTRIB_WPOS
]];
4376 /* First, apply the coordinate shift: */
4377 if (adjX
|| adjY
[0] || adjY
[1]) {
4378 if (adjY
[0] != adjY
[1]) {
4379 /* Adjust the y coordinate by adjY[1] or adjY[0] respectively
4380 * depending on whether inversion is actually going to be applied
4381 * or not, which is determined by testing against the inversion
4382 * state variable used below, which will be either +1 or -1.
4384 struct ureg_dst adj_temp
= ureg_DECL_temporary(ureg
);
4386 ureg_CMP(ureg
, adj_temp
,
4387 ureg_scalar(wpostrans
, invert
? 2 : 0),
4388 ureg_imm4f(ureg
, adjX
, adjY
[0], 0.0f
, 0.0f
),
4389 ureg_imm4f(ureg
, adjX
, adjY
[1], 0.0f
, 0.0f
));
4390 ureg_ADD(ureg
, wpos_temp
, wpos_input
, ureg_src(adj_temp
));
4392 ureg_ADD(ureg
, wpos_temp
, wpos_input
,
4393 ureg_imm4f(ureg
, adjX
, adjY
[0], 0.0f
, 0.0f
));
4395 wpos_input
= ureg_src(wpos_temp
);
4397 /* MOV wpos_temp, input[wpos]
4399 ureg_MOV( ureg
, wpos_temp
, wpos_input
);
4402 /* Now the conditional y flip: STATE_FB_WPOS_Y_TRANSFORM.xy/zw will be
4403 * inversion/identity, or the other way around if we're drawing to an FBO.
4406 /* MAD wpos_temp.y, wpos_input, wpostrans.xxxx, wpostrans.yyyy
4409 ureg_writemask(wpos_temp
, TGSI_WRITEMASK_Y
),
4411 ureg_scalar(wpostrans
, 0),
4412 ureg_scalar(wpostrans
, 1));
4414 /* MAD wpos_temp.y, wpos_input, wpostrans.zzzz, wpostrans.wwww
4417 ureg_writemask(wpos_temp
, TGSI_WRITEMASK_Y
),
4419 ureg_scalar(wpostrans
, 2),
4420 ureg_scalar(wpostrans
, 3));
4423 /* Use wpos_temp as position input from here on:
4425 t
->inputs
[t
->inputMapping
[FRAG_ATTRIB_WPOS
]] = ureg_src(wpos_temp
);
4430 * Emit fragment position/ooordinate code.
4433 emit_wpos(struct st_context
*st
,
4434 struct st_translate
*t
,
4435 const struct gl_program
*program
,
4436 struct ureg_program
*ureg
)
4438 const struct gl_fragment_program
*fp
=
4439 (const struct gl_fragment_program
*) program
;
4440 struct pipe_screen
*pscreen
= st
->pipe
->screen
;
4441 GLfloat adjX
= 0.0f
;
4442 GLfloat adjY
[2] = { 0.0f
, 0.0f
};
4443 boolean invert
= FALSE
;
4445 /* Query the pixel center conventions supported by the pipe driver and set
4446 * adjX, adjY to help out if it cannot handle the requested one internally.
4448 * The bias of the y-coordinate depends on whether y-inversion takes place
4449 * (adjY[1]) or not (adjY[0]), which is in turn dependent on whether we are
4450 * drawing to an FBO (causes additional inversion), and whether the the pipe
4451 * driver origin and the requested origin differ (the latter condition is
4452 * stored in the 'invert' variable).
4454 * For height = 100 (i = integer, h = half-integer, l = lower, u = upper):
4456 * center shift only:
4461 * l,i -> u,i: ( 0.0 + 1.0) * -1 + 100 = 99
4462 * l,h -> u,h: ( 0.5 + 0.0) * -1 + 100 = 99.5
4463 * u,i -> l,i: (99.0 + 1.0) * -1 + 100 = 0
4464 * u,h -> l,h: (99.5 + 0.0) * -1 + 100 = 0.5
4466 * inversion and center shift:
4467 * l,i -> u,h: ( 0.0 + 0.5) * -1 + 100 = 99.5
4468 * l,h -> u,i: ( 0.5 + 0.5) * -1 + 100 = 99
4469 * u,i -> l,h: (99.0 + 0.5) * -1 + 100 = 0.5
4470 * u,h -> l,i: (99.5 + 0.5) * -1 + 100 = 0
4472 if (fp
->OriginUpperLeft
) {
4473 /* Fragment shader wants origin in upper-left */
4474 if (pscreen
->get_param(pscreen
, PIPE_CAP_TGSI_FS_COORD_ORIGIN_UPPER_LEFT
)) {
4475 /* the driver supports upper-left origin */
4477 else if (pscreen
->get_param(pscreen
, PIPE_CAP_TGSI_FS_COORD_ORIGIN_LOWER_LEFT
)) {
4478 /* the driver supports lower-left origin, need to invert Y */
4479 ureg_property_fs_coord_origin(ureg
, TGSI_FS_COORD_ORIGIN_LOWER_LEFT
);
4486 /* Fragment shader wants origin in lower-left */
4487 if (pscreen
->get_param(pscreen
, PIPE_CAP_TGSI_FS_COORD_ORIGIN_LOWER_LEFT
))
4488 /* the driver supports lower-left origin */
4489 ureg_property_fs_coord_origin(ureg
, TGSI_FS_COORD_ORIGIN_LOWER_LEFT
);
4490 else if (pscreen
->get_param(pscreen
, PIPE_CAP_TGSI_FS_COORD_ORIGIN_UPPER_LEFT
))
4491 /* the driver supports upper-left origin, need to invert Y */
4497 if (fp
->PixelCenterInteger
) {
4498 /* Fragment shader wants pixel center integer */
4499 if (pscreen
->get_param(pscreen
, PIPE_CAP_TGSI_FS_COORD_PIXEL_CENTER_INTEGER
)) {
4500 /* the driver supports pixel center integer */
4502 ureg_property_fs_coord_pixel_center(ureg
, TGSI_FS_COORD_PIXEL_CENTER_INTEGER
);
4504 else if (pscreen
->get_param(pscreen
, PIPE_CAP_TGSI_FS_COORD_PIXEL_CENTER_HALF_INTEGER
)) {
4505 /* the driver supports pixel center half integer, need to bias X,Y */
4514 /* Fragment shader wants pixel center half integer */
4515 if (pscreen
->get_param(pscreen
, PIPE_CAP_TGSI_FS_COORD_PIXEL_CENTER_HALF_INTEGER
)) {
4516 /* the driver supports pixel center half integer */
4518 else if (pscreen
->get_param(pscreen
, PIPE_CAP_TGSI_FS_COORD_PIXEL_CENTER_INTEGER
)) {
4519 /* the driver supports pixel center integer, need to bias X,Y */
4520 adjX
= adjY
[0] = adjY
[1] = 0.5f
;
4521 ureg_property_fs_coord_pixel_center(ureg
, TGSI_FS_COORD_PIXEL_CENTER_INTEGER
);
4527 /* we invert after adjustment so that we avoid the MOV to temporary,
4528 * and reuse the adjustment ADD instead */
4529 emit_wpos_adjustment(t
, program
, invert
, adjX
, adjY
);
4533 * OpenGL's fragment gl_FrontFace input is 1 for front-facing, 0 for back.
4534 * TGSI uses +1 for front, -1 for back.
4535 * This function converts the TGSI value to the GL value. Simply clamping/
4536 * saturating the value to [0,1] does the job.
4539 emit_face_var(struct st_translate
*t
)
4541 struct ureg_program
*ureg
= t
->ureg
;
4542 struct ureg_dst face_temp
= ureg_DECL_temporary(ureg
);
4543 struct ureg_src face_input
= t
->inputs
[t
->inputMapping
[FRAG_ATTRIB_FACE
]];
4545 /* MOV_SAT face_temp, input[face] */
4546 face_temp
= ureg_saturate(face_temp
);
4547 ureg_MOV(ureg
, face_temp
, face_input
);
4549 /* Use face_temp as face input from here on: */
4550 t
->inputs
[t
->inputMapping
[FRAG_ATTRIB_FACE
]] = ureg_src(face_temp
);
4554 emit_edgeflags(struct st_translate
*t
)
4556 struct ureg_program
*ureg
= t
->ureg
;
4557 struct ureg_dst edge_dst
= t
->outputs
[t
->outputMapping
[VERT_RESULT_EDGE
]];
4558 struct ureg_src edge_src
= t
->inputs
[t
->inputMapping
[VERT_ATTRIB_EDGEFLAG
]];
4560 ureg_MOV(ureg
, edge_dst
, edge_src
);
4564 * Translate intermediate IR (glsl_to_tgsi_instruction) to TGSI format.
4565 * \param program the program to translate
4566 * \param numInputs number of input registers used
4567 * \param inputMapping maps Mesa fragment program inputs to TGSI generic
4569 * \param inputSemanticName the TGSI_SEMANTIC flag for each input
4570 * \param inputSemanticIndex the semantic index (ex: which texcoord) for
4572 * \param interpMode the TGSI_INTERPOLATE_LINEAR/PERSP mode for each input
4573 * \param numOutputs number of output registers used
4574 * \param outputMapping maps Mesa fragment program outputs to TGSI
4576 * \param outputSemanticName the TGSI_SEMANTIC flag for each output
4577 * \param outputSemanticIndex the semantic index (ex: which texcoord) for
4580 * \return PIPE_OK or PIPE_ERROR_OUT_OF_MEMORY
4582 extern "C" enum pipe_error
4583 st_translate_program(
4584 struct gl_context
*ctx
,
4586 struct ureg_program
*ureg
,
4587 glsl_to_tgsi_visitor
*program
,
4588 const struct gl_program
*proginfo
,
4590 const GLuint inputMapping
[],
4591 const ubyte inputSemanticName
[],
4592 const ubyte inputSemanticIndex
[],
4593 const GLuint interpMode
[],
4595 const GLuint outputMapping
[],
4596 const ubyte outputSemanticName
[],
4597 const ubyte outputSemanticIndex
[],
4598 boolean passthrough_edgeflags
)
4600 struct st_translate
*t
;
4602 enum pipe_error ret
= PIPE_OK
;
4604 assert(numInputs
<= Elements(t
->inputs
));
4605 assert(numOutputs
<= Elements(t
->outputs
));
4607 t
= CALLOC_STRUCT(st_translate
);
4609 ret
= PIPE_ERROR_OUT_OF_MEMORY
;
4613 memset(t
, 0, sizeof *t
);
4615 t
->procType
= procType
;
4616 t
->inputMapping
= inputMapping
;
4617 t
->outputMapping
= outputMapping
;
4619 t
->pointSizeOutIndex
= -1;
4620 t
->prevInstWrotePointSize
= GL_FALSE
;
4623 * Declare input attributes.
4625 if (procType
== TGSI_PROCESSOR_FRAGMENT
) {
4626 for (i
= 0; i
< numInputs
; i
++) {
4627 t
->inputs
[i
] = ureg_DECL_fs_input(ureg
,
4628 inputSemanticName
[i
],
4629 inputSemanticIndex
[i
],
4633 if (proginfo
->InputsRead
& FRAG_BIT_WPOS
) {
4634 /* Must do this after setting up t->inputs, and before
4635 * emitting constant references, below:
4637 emit_wpos(st_context(ctx
), t
, proginfo
, ureg
);
4640 if (proginfo
->InputsRead
& FRAG_BIT_FACE
)
4644 * Declare output attributes.
4646 for (i
= 0; i
< numOutputs
; i
++) {
4647 switch (outputSemanticName
[i
]) {
4648 case TGSI_SEMANTIC_POSITION
:
4649 t
->outputs
[i
] = ureg_DECL_output(ureg
,
4650 TGSI_SEMANTIC_POSITION
, /* Z/Depth */
4651 outputSemanticIndex
[i
]);
4652 t
->outputs
[i
] = ureg_writemask(t
->outputs
[i
], TGSI_WRITEMASK_Z
);
4654 case TGSI_SEMANTIC_STENCIL
:
4655 t
->outputs
[i
] = ureg_DECL_output(ureg
,
4656 TGSI_SEMANTIC_STENCIL
, /* Stencil */
4657 outputSemanticIndex
[i
]);
4658 t
->outputs
[i
] = ureg_writemask(t
->outputs
[i
], TGSI_WRITEMASK_Y
);
4660 case TGSI_SEMANTIC_COLOR
:
4661 t
->outputs
[i
] = ureg_DECL_output(ureg
,
4662 TGSI_SEMANTIC_COLOR
,
4663 outputSemanticIndex
[i
]);
4666 assert(!"fragment shader outputs must be POSITION/STENCIL/COLOR");
4667 ret
= PIPE_ERROR_BAD_INPUT
;
4672 else if (procType
== TGSI_PROCESSOR_GEOMETRY
) {
4673 for (i
= 0; i
< numInputs
; i
++) {
4674 t
->inputs
[i
] = ureg_DECL_gs_input(ureg
,
4676 inputSemanticName
[i
],
4677 inputSemanticIndex
[i
]);
4680 for (i
= 0; i
< numOutputs
; i
++) {
4681 t
->outputs
[i
] = ureg_DECL_output(ureg
,
4682 outputSemanticName
[i
],
4683 outputSemanticIndex
[i
]);
4687 assert(procType
== TGSI_PROCESSOR_VERTEX
);
4689 for (i
= 0; i
< numInputs
; i
++) {
4690 t
->inputs
[i
] = ureg_DECL_vs_input(ureg
, i
);
4693 for (i
= 0; i
< numOutputs
; i
++) {
4694 t
->outputs
[i
] = ureg_DECL_output(ureg
,
4695 outputSemanticName
[i
],
4696 outputSemanticIndex
[i
]);
4697 if ((outputSemanticName
[i
] == TGSI_SEMANTIC_PSIZE
) && proginfo
->Id
) {
4698 /* Writing to the point size result register requires special
4699 * handling to implement clamping.
4701 static const gl_state_index pointSizeClampState
[STATE_LENGTH
]
4702 = { STATE_INTERNAL
, STATE_POINT_SIZE_IMPL_CLAMP
, (gl_state_index
)0, (gl_state_index
)0, (gl_state_index
)0 };
4703 /* XXX: note we are modifying the incoming shader here! Need to
4704 * do this before emitting the constant decls below, or this
4707 unsigned pointSizeClampConst
=
4708 _mesa_add_state_reference(proginfo
->Parameters
,
4709 pointSizeClampState
);
4710 struct ureg_dst psizregtemp
= ureg_DECL_temporary(ureg
);
4711 t
->pointSizeConst
= ureg_DECL_constant(ureg
, pointSizeClampConst
);
4712 t
->pointSizeResult
= t
->outputs
[i
];
4713 t
->pointSizeOutIndex
= i
;
4714 t
->outputs
[i
] = psizregtemp
;
4717 if (passthrough_edgeflags
)
4721 /* Declare address register.
4723 if (program
->num_address_regs
> 0) {
4724 assert(program
->num_address_regs
== 1);
4725 t
->address
[0] = ureg_DECL_address(ureg
);
4728 /* Declare misc input registers
4731 GLbitfield sysInputs
= proginfo
->SystemValuesRead
;
4732 unsigned numSys
= 0;
4733 for (i
= 0; sysInputs
; i
++) {
4734 if (sysInputs
& (1 << i
)) {
4735 unsigned semName
= mesa_sysval_to_semantic
[i
];
4736 t
->systemValues
[i
] = ureg_DECL_system_value(ureg
, numSys
, semName
, 0);
4738 sysInputs
&= ~(1 << i
);
4743 if (program
->indirect_addr_temps
) {
4744 /* If temps are accessed with indirect addressing, declare temporaries
4745 * in sequential order. Else, we declare them on demand elsewhere.
4746 * (Note: the number of temporaries is equal to program->next_temp)
4748 for (i
= 0; i
< (unsigned)program
->next_temp
; i
++) {
4749 /* XXX use TGSI_FILE_TEMPORARY_ARRAY when it's supported by ureg */
4750 t
->temps
[i
] = ureg_DECL_temporary(t
->ureg
);
4754 /* Emit constants and uniforms. TGSI uses a single index space for these,
4755 * so we put all the translated regs in t->constants.
4757 if (proginfo
->Parameters
) {
4758 t
->constants
= (struct ureg_src
*)CALLOC(proginfo
->Parameters
->NumParameters
* sizeof(t
->constants
[0]));
4759 if (t
->constants
== NULL
) {
4760 ret
= PIPE_ERROR_OUT_OF_MEMORY
;
4764 for (i
= 0; i
< proginfo
->Parameters
->NumParameters
; i
++) {
4765 switch (proginfo
->Parameters
->Parameters
[i
].Type
) {
4766 case PROGRAM_ENV_PARAM
:
4767 case PROGRAM_LOCAL_PARAM
:
4768 case PROGRAM_STATE_VAR
:
4769 case PROGRAM_NAMED_PARAM
:
4770 case PROGRAM_UNIFORM
:
4771 t
->constants
[i
] = ureg_DECL_constant(ureg
, i
);
4774 /* Emit immediates for PROGRAM_CONSTANT only when there's no indirect
4775 * addressing of the const buffer.
4776 * FIXME: Be smarter and recognize param arrays:
4777 * indirect addressing is only valid within the referenced
4780 case PROGRAM_CONSTANT
:
4781 if (program
->indirect_addr_consts
)
4782 t
->constants
[i
] = ureg_DECL_constant(ureg
, i
);
4784 t
->constants
[i
] = emit_immediate(t
,
4785 proginfo
->Parameters
->ParameterValues
[i
],
4786 proginfo
->Parameters
->Parameters
[i
].DataType
,
4795 /* Emit immediate values.
4797 t
->immediates
= (struct ureg_src
*)CALLOC(program
->num_immediates
* sizeof(struct ureg_src
));
4798 if (t
->immediates
== NULL
) {
4799 ret
= PIPE_ERROR_OUT_OF_MEMORY
;
4803 foreach_iter(exec_list_iterator
, iter
, program
->immediates
) {
4804 immediate_storage
*imm
= (immediate_storage
*)iter
.get();
4805 t
->immediates
[i
++] = emit_immediate(t
, imm
->values
, imm
->type
, imm
->size
);
4808 /* texture samplers */
4809 for (i
= 0; i
< ctx
->Const
.MaxTextureImageUnits
; i
++) {
4810 if (program
->samplers_used
& (1 << i
)) {
4811 t
->samplers
[i
] = ureg_DECL_sampler(ureg
, i
);
4815 /* Emit each instruction in turn:
4817 foreach_iter(exec_list_iterator
, iter
, program
->instructions
) {
4818 set_insn_start(t
, ureg_get_instruction_number(ureg
));
4819 compile_tgsi_instruction(t
, (glsl_to_tgsi_instruction
*)iter
.get());
4821 if (t
->prevInstWrotePointSize
&& proginfo
->Id
) {
4822 /* The previous instruction wrote to the (fake) vertex point size
4823 * result register. Now we need to clamp that value to the min/max
4824 * point size range, putting the result into the real point size
4826 * Note that we can't do this easily at the end of program due to
4827 * possible early return.
4829 set_insn_start(t
, ureg_get_instruction_number(ureg
));
4831 ureg_writemask(t
->outputs
[t
->pointSizeOutIndex
], WRITEMASK_X
),
4832 ureg_src(t
->outputs
[t
->pointSizeOutIndex
]),
4833 ureg_swizzle(t
->pointSizeConst
, 1,1,1,1));
4834 ureg_MIN(t
->ureg
, ureg_writemask(t
->pointSizeResult
, WRITEMASK_X
),
4835 ureg_src(t
->outputs
[t
->pointSizeOutIndex
]),
4836 ureg_swizzle(t
->pointSizeConst
, 2,2,2,2));
4838 t
->prevInstWrotePointSize
= GL_FALSE
;
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
]);
4853 FREE(t
->immediates
);
4856 debug_printf("%s: translate error flag set\n", __FUNCTION__
);
4864 /* ----------------------------- End TGSI code ------------------------------ */
4867 * Convert a shader's GLSL IR into a Mesa gl_program, although without
4868 * generating Mesa IR.
4870 static struct gl_program
*
4871 get_mesa_program(struct gl_context
*ctx
,
4872 struct gl_shader_program
*shader_program
,
4873 struct gl_shader
*shader
)
4875 glsl_to_tgsi_visitor
* v
= new glsl_to_tgsi_visitor();
4876 struct gl_program
*prog
;
4878 const char *target_string
;
4880 struct gl_shader_compiler_options
*options
=
4881 &ctx
->ShaderCompilerOptions
[_mesa_shader_type_to_index(shader
->Type
)];
4883 switch (shader
->Type
) {
4884 case GL_VERTEX_SHADER
:
4885 target
= GL_VERTEX_PROGRAM_ARB
;
4886 target_string
= "vertex";
4888 case GL_FRAGMENT_SHADER
:
4889 target
= GL_FRAGMENT_PROGRAM_ARB
;
4890 target_string
= "fragment";
4892 case GL_GEOMETRY_SHADER
:
4893 target
= GL_GEOMETRY_PROGRAM_NV
;
4894 target_string
= "geometry";
4897 assert(!"should not be reached");
4901 validate_ir_tree(shader
->ir
);
4903 prog
= ctx
->Driver
.NewProgram(ctx
, target
, shader_program
->Name
);
4906 prog
->Parameters
= _mesa_new_parameter_list();
4909 v
->shader_program
= shader_program
;
4910 v
->options
= options
;
4911 v
->glsl_version
= ctx
->Const
.GLSLVersion
;
4912 v
->native_integers
= ctx
->Const
.NativeIntegers
;
4914 _mesa_generate_parameters_list_for_uniforms(shader_program
, shader
,
4917 /* Emit intermediate IR for main(). */
4918 visit_exec_list(shader
->ir
, v
);
4920 /* Now emit bodies for any functions that were used. */
4922 progress
= GL_FALSE
;
4924 foreach_iter(exec_list_iterator
, iter
, v
->function_signatures
) {
4925 function_entry
*entry
= (function_entry
*)iter
.get();
4927 if (!entry
->bgn_inst
) {
4928 v
->current_function
= entry
;
4930 entry
->bgn_inst
= v
->emit(NULL
, TGSI_OPCODE_BGNSUB
);
4931 entry
->bgn_inst
->function
= entry
;
4933 visit_exec_list(&entry
->sig
->body
, v
);
4935 glsl_to_tgsi_instruction
*last
;
4936 last
= (glsl_to_tgsi_instruction
*)v
->instructions
.get_tail();
4937 if (last
->op
!= TGSI_OPCODE_RET
)
4938 v
->emit(NULL
, TGSI_OPCODE_RET
);
4940 glsl_to_tgsi_instruction
*end
;
4941 end
= v
->emit(NULL
, TGSI_OPCODE_ENDSUB
);
4942 end
->function
= entry
;
4950 /* Print out some information (for debugging purposes) used by the
4951 * optimization passes. */
4952 for (i
=0; i
< v
->next_temp
; i
++) {
4953 int fr
= v
->get_first_temp_read(i
);
4954 int fw
= v
->get_first_temp_write(i
);
4955 int lr
= v
->get_last_temp_read(i
);
4956 int lw
= v
->get_last_temp_write(i
);
4958 printf("Temp %d: FR=%3d FW=%3d LR=%3d LW=%3d\n", i
, fr
, fw
, lr
, lw
);
4963 /* Remove reads to output registers, and to varyings in vertex shaders. */
4964 v
->remove_output_reads(PROGRAM_OUTPUT
);
4965 if (target
== GL_VERTEX_PROGRAM_ARB
)
4966 v
->remove_output_reads(PROGRAM_VARYING
);
4968 /* Perform optimizations on the instructions in the glsl_to_tgsi_visitor. */
4970 v
->copy_propagate();
4971 while (v
->eliminate_dead_code_advanced());
4973 /* FIXME: These passes to optimize temporary registers don't work when there
4974 * is indirect addressing of the temporary register space. We need proper
4975 * array support so that we don't have to give up these passes in every
4976 * shader that uses arrays.
4978 if (!v
->indirect_addr_temps
) {
4979 v
->eliminate_dead_code();
4980 v
->merge_registers();
4981 v
->renumber_registers();
4984 /* Write the END instruction. */
4985 v
->emit(NULL
, TGSI_OPCODE_END
);
4987 if (ctx
->Shader
.Flags
& GLSL_DUMP
) {
4989 printf("GLSL IR for linked %s program %d:\n", target_string
,
4990 shader_program
->Name
);
4991 _mesa_print_ir(shader
->ir
, NULL
);
4996 prog
->Instructions
= NULL
;
4997 prog
->NumInstructions
= 0;
4999 do_set_program_inouts(shader
->ir
, prog
, shader
->Type
== GL_FRAGMENT_SHADER
);
5000 count_resources(v
, prog
);
5002 check_resources(ctx
, shader_program
, v
, prog
);
5004 _mesa_reference_program(ctx
, &shader
->Program
, prog
);
5006 /* This has to be done last. Any operation the can cause
5007 * prog->ParameterValues to get reallocated (e.g., anything that adds a
5008 * program constant) has to happen before creating this linkage.
5010 _mesa_associate_uniform_storage(ctx
, shader_program
, prog
->Parameters
);
5011 if (!shader_program
->LinkStatus
) {
5015 struct st_vertex_program
*stvp
;
5016 struct st_fragment_program
*stfp
;
5017 struct st_geometry_program
*stgp
;
5019 switch (shader
->Type
) {
5020 case GL_VERTEX_SHADER
:
5021 stvp
= (struct st_vertex_program
*)prog
;
5022 stvp
->glsl_to_tgsi
= v
;
5024 case GL_FRAGMENT_SHADER
:
5025 stfp
= (struct st_fragment_program
*)prog
;
5026 stfp
->glsl_to_tgsi
= v
;
5028 case GL_GEOMETRY_SHADER
:
5029 stgp
= (struct st_geometry_program
*)prog
;
5030 stgp
->glsl_to_tgsi
= v
;
5033 assert(!"should not be reached");
5043 st_new_shader(struct gl_context
*ctx
, GLuint name
, GLuint type
)
5045 struct gl_shader
*shader
;
5046 assert(type
== GL_FRAGMENT_SHADER
|| type
== GL_VERTEX_SHADER
||
5047 type
== GL_GEOMETRY_SHADER_ARB
);
5048 shader
= rzalloc(NULL
, struct gl_shader
);
5050 shader
->Type
= type
;
5051 shader
->Name
= name
;
5052 _mesa_init_shader(ctx
, shader
);
5057 struct gl_shader_program
*
5058 st_new_shader_program(struct gl_context
*ctx
, GLuint name
)
5060 struct gl_shader_program
*shProg
;
5061 shProg
= rzalloc(NULL
, struct gl_shader_program
);
5063 shProg
->Name
= name
;
5064 _mesa_init_shader_program(ctx
, shProg
);
5071 * Called via ctx->Driver.LinkShader()
5072 * This actually involves converting GLSL IR into an intermediate TGSI-like IR
5073 * with code lowering and other optimizations.
5076 st_link_shader(struct gl_context
*ctx
, struct gl_shader_program
*prog
)
5078 assert(prog
->LinkStatus
);
5080 for (unsigned i
= 0; i
< MESA_SHADER_TYPES
; i
++) {
5081 if (prog
->_LinkedShaders
[i
] == NULL
)
5085 exec_list
*ir
= prog
->_LinkedShaders
[i
]->ir
;
5086 const struct gl_shader_compiler_options
*options
=
5087 &ctx
->ShaderCompilerOptions
[_mesa_shader_type_to_index(prog
->_LinkedShaders
[i
]->Type
)];
5093 do_mat_op_to_vec(ir
);
5094 lower_instructions(ir
, (MOD_TO_FRACT
| DIV_TO_MUL_RCP
| EXP_TO_EXP2
5095 | LOG_TO_LOG2
| INT_DIV_TO_MUL_RCP
5096 | ((options
->EmitNoPow
) ? POW_TO_EXP2
: 0)));
5098 progress
= do_lower_jumps(ir
, true, true, options
->EmitNoMainReturn
, options
->EmitNoCont
, options
->EmitNoLoops
) || progress
;
5100 progress
= do_common_optimization(ir
, true, true,
5101 options
->MaxUnrollIterations
)
5104 progress
= lower_quadop_vector(ir
, false) || progress
;
5106 if (options
->MaxIfDepth
== 0)
5107 progress
= lower_discard(ir
) || progress
;
5109 progress
= lower_if_to_cond_assign(ir
, options
->MaxIfDepth
) || progress
;
5111 if (options
->EmitNoNoise
)
5112 progress
= lower_noise(ir
) || progress
;
5114 /* If there are forms of indirect addressing that the driver
5115 * cannot handle, perform the lowering pass.
5117 if (options
->EmitNoIndirectInput
|| options
->EmitNoIndirectOutput
5118 || options
->EmitNoIndirectTemp
|| options
->EmitNoIndirectUniform
)
5120 lower_variable_index_to_cond_assign(ir
,
5121 options
->EmitNoIndirectInput
,
5122 options
->EmitNoIndirectOutput
,
5123 options
->EmitNoIndirectTemp
,
5124 options
->EmitNoIndirectUniform
)
5127 progress
= do_vec_index_to_cond_assign(ir
) || progress
;
5130 validate_ir_tree(ir
);
5133 for (unsigned i
= 0; i
< MESA_SHADER_TYPES
; i
++) {
5134 struct gl_program
*linked_prog
;
5136 if (prog
->_LinkedShaders
[i
] == NULL
)
5139 linked_prog
= get_mesa_program(ctx
, prog
, prog
->_LinkedShaders
[i
]);
5142 static const GLenum targets
[] = {
5143 GL_VERTEX_PROGRAM_ARB
,
5144 GL_FRAGMENT_PROGRAM_ARB
,
5145 GL_GEOMETRY_PROGRAM_NV
5148 _mesa_reference_program(ctx
, &prog
->_LinkedShaders
[i
]->Program
,
5150 if (!ctx
->Driver
.ProgramStringNotify(ctx
, targets
[i
], linked_prog
)) {
5151 _mesa_reference_program(ctx
, &prog
->_LinkedShaders
[i
]->Program
,
5153 _mesa_reference_program(ctx
, &linked_prog
, NULL
);
5158 _mesa_reference_program(ctx
, &linked_prog
, NULL
);