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
;
1022 if (ir
->mode
== ir_var_uniform
&& strncmp(ir
->name
, "gl_", 3) == 0) {
1024 const ir_state_slot
*const slots
= ir
->state_slots
;
1025 assert(ir
->state_slots
!= NULL
);
1027 /* Check if this statevar's setup in the STATE file exactly
1028 * matches how we'll want to reference it as a
1029 * struct/array/whatever. If not, then we need to move it into
1030 * temporary storage and hope that it'll get copy-propagated
1033 for (i
= 0; i
< ir
->num_state_slots
; i
++) {
1034 if (slots
[i
].swizzle
!= SWIZZLE_XYZW
) {
1039 variable_storage
*storage
;
1041 if (i
== ir
->num_state_slots
) {
1042 /* We'll set the index later. */
1043 storage
= new(mem_ctx
) variable_storage(ir
, PROGRAM_STATE_VAR
, -1);
1044 this->variables
.push_tail(storage
);
1048 /* The variable_storage constructor allocates slots based on the size
1049 * of the type. However, this had better match the number of state
1050 * elements that we're going to copy into the new temporary.
1052 assert((int) ir
->num_state_slots
== type_size(ir
->type
));
1054 storage
= new(mem_ctx
) variable_storage(ir
, PROGRAM_TEMPORARY
,
1056 this->variables
.push_tail(storage
);
1057 this->next_temp
+= type_size(ir
->type
);
1059 dst
= st_dst_reg(st_src_reg(PROGRAM_TEMPORARY
, storage
->index
,
1060 native_integers
? ir
->type
->base_type
: GLSL_TYPE_FLOAT
));
1064 for (unsigned int i
= 0; i
< ir
->num_state_slots
; i
++) {
1065 int index
= _mesa_add_state_reference(this->prog
->Parameters
,
1066 (gl_state_index
*)slots
[i
].tokens
);
1068 if (storage
->file
== PROGRAM_STATE_VAR
) {
1069 if (storage
->index
== -1) {
1070 storage
->index
= index
;
1072 assert(index
== storage
->index
+ (int)i
);
1075 st_src_reg
src(PROGRAM_STATE_VAR
, index
,
1076 native_integers
? ir
->type
->base_type
: GLSL_TYPE_FLOAT
);
1077 src
.swizzle
= slots
[i
].swizzle
;
1078 emit(ir
, TGSI_OPCODE_MOV
, dst
, src
);
1079 /* even a float takes up a whole vec4 reg in a struct/array. */
1084 if (storage
->file
== PROGRAM_TEMPORARY
&&
1085 dst
.index
!= storage
->index
+ (int) ir
->num_state_slots
) {
1086 fail_link(this->shader_program
,
1087 "failed to load builtin uniform `%s' (%d/%d regs loaded)\n",
1088 ir
->name
, dst
.index
- storage
->index
,
1089 type_size(ir
->type
));
1095 glsl_to_tgsi_visitor::visit(ir_loop
*ir
)
1097 ir_dereference_variable
*counter
= NULL
;
1099 if (ir
->counter
!= NULL
)
1100 counter
= new(ir
) ir_dereference_variable(ir
->counter
);
1102 if (ir
->from
!= NULL
) {
1103 assert(ir
->counter
!= NULL
);
1105 ir_assignment
*a
= new(ir
) ir_assignment(counter
, ir
->from
, NULL
);
1111 emit(NULL
, TGSI_OPCODE_BGNLOOP
);
1115 new(ir
) ir_expression(ir
->cmp
, glsl_type::bool_type
,
1117 ir_if
*if_stmt
= new(ir
) ir_if(e
);
1119 ir_loop_jump
*brk
= new(ir
) ir_loop_jump(ir_loop_jump::jump_break
);
1121 if_stmt
->then_instructions
.push_tail(brk
);
1123 if_stmt
->accept(this);
1130 visit_exec_list(&ir
->body_instructions
, this);
1132 if (ir
->increment
) {
1134 new(ir
) ir_expression(ir_binop_add
, counter
->type
,
1135 counter
, ir
->increment
);
1137 ir_assignment
*a
= new(ir
) ir_assignment(counter
, e
, NULL
);
1144 emit(NULL
, TGSI_OPCODE_ENDLOOP
);
1148 glsl_to_tgsi_visitor::visit(ir_loop_jump
*ir
)
1151 case ir_loop_jump::jump_break
:
1152 emit(NULL
, TGSI_OPCODE_BRK
);
1154 case ir_loop_jump::jump_continue
:
1155 emit(NULL
, TGSI_OPCODE_CONT
);
1162 glsl_to_tgsi_visitor::visit(ir_function_signature
*ir
)
1169 glsl_to_tgsi_visitor::visit(ir_function
*ir
)
1171 /* Ignore function bodies other than main() -- we shouldn't see calls to
1172 * them since they should all be inlined before we get to glsl_to_tgsi.
1174 if (strcmp(ir
->name
, "main") == 0) {
1175 const ir_function_signature
*sig
;
1178 sig
= ir
->matching_signature(&empty
);
1182 foreach_iter(exec_list_iterator
, iter
, sig
->body
) {
1183 ir_instruction
*ir
= (ir_instruction
*)iter
.get();
1191 glsl_to_tgsi_visitor::try_emit_mad(ir_expression
*ir
, int mul_operand
)
1193 int nonmul_operand
= 1 - mul_operand
;
1195 st_dst_reg result_dst
;
1197 ir_expression
*expr
= ir
->operands
[mul_operand
]->as_expression();
1198 if (!expr
|| expr
->operation
!= ir_binop_mul
)
1201 expr
->operands
[0]->accept(this);
1203 expr
->operands
[1]->accept(this);
1205 ir
->operands
[nonmul_operand
]->accept(this);
1208 this->result
= get_temp(ir
->type
);
1209 result_dst
= st_dst_reg(this->result
);
1210 result_dst
.writemask
= (1 << ir
->type
->vector_elements
) - 1;
1211 emit(ir
, TGSI_OPCODE_MAD
, result_dst
, a
, b
, c
);
1217 * Emit MAD(a, -b, a) instead of AND(a, NOT(b))
1219 * The logic values are 1.0 for true and 0.0 for false. Logical-and is
1220 * implemented using multiplication, and logical-or is implemented using
1221 * addition. Logical-not can be implemented as (true - x), or (1.0 - x).
1222 * As result, the logical expression (a & !b) can be rewritten as:
1226 * - (a * 1) - (a * b)
1230 * This final expression can be implemented as a single MAD(a, -b, a)
1234 glsl_to_tgsi_visitor::try_emit_mad_for_and_not(ir_expression
*ir
, int try_operand
)
1236 const int other_operand
= 1 - try_operand
;
1239 ir_expression
*expr
= ir
->operands
[try_operand
]->as_expression();
1240 if (!expr
|| expr
->operation
!= ir_unop_logic_not
)
1243 ir
->operands
[other_operand
]->accept(this);
1245 expr
->operands
[0]->accept(this);
1248 b
.negate
= ~b
.negate
;
1250 this->result
= get_temp(ir
->type
);
1251 emit(ir
, TGSI_OPCODE_MAD
, st_dst_reg(this->result
), a
, b
, a
);
1257 glsl_to_tgsi_visitor::try_emit_sat(ir_expression
*ir
)
1259 /* Saturates were only introduced to vertex programs in
1260 * NV_vertex_program3, so don't give them to drivers in the VP.
1262 if (this->prog
->Target
== GL_VERTEX_PROGRAM_ARB
)
1265 ir_rvalue
*sat_src
= ir
->as_rvalue_to_saturate();
1269 sat_src
->accept(this);
1270 st_src_reg src
= this->result
;
1272 /* If we generated an expression instruction into a temporary in
1273 * processing the saturate's operand, apply the saturate to that
1274 * instruction. Otherwise, generate a MOV to do the saturate.
1276 * Note that we have to be careful to only do this optimization if
1277 * the instruction in question was what generated src->result. For
1278 * example, ir_dereference_array might generate a MUL instruction
1279 * to create the reladdr, and return us a src reg using that
1280 * reladdr. That MUL result is not the value we're trying to
1283 ir_expression
*sat_src_expr
= sat_src
->as_expression();
1284 if (sat_src_expr
&& (sat_src_expr
->operation
== ir_binop_mul
||
1285 sat_src_expr
->operation
== ir_binop_add
||
1286 sat_src_expr
->operation
== ir_binop_dot
)) {
1287 glsl_to_tgsi_instruction
*new_inst
;
1288 new_inst
= (glsl_to_tgsi_instruction
*)this->instructions
.get_tail();
1289 new_inst
->saturate
= true;
1291 this->result
= get_temp(ir
->type
);
1292 st_dst_reg result_dst
= st_dst_reg(this->result
);
1293 result_dst
.writemask
= (1 << ir
->type
->vector_elements
) - 1;
1294 glsl_to_tgsi_instruction
*inst
;
1295 inst
= emit(ir
, TGSI_OPCODE_MOV
, result_dst
, src
);
1296 inst
->saturate
= true;
1303 glsl_to_tgsi_visitor::reladdr_to_temp(ir_instruction
*ir
,
1304 st_src_reg
*reg
, int *num_reladdr
)
1309 emit_arl(ir
, address_reg
, *reg
->reladdr
);
1311 if (*num_reladdr
!= 1) {
1312 st_src_reg temp
= get_temp(glsl_type::vec4_type
);
1314 emit(ir
, TGSI_OPCODE_MOV
, st_dst_reg(temp
), *reg
);
1322 glsl_to_tgsi_visitor::visit(ir_expression
*ir
)
1324 unsigned int operand
;
1325 st_src_reg op
[Elements(ir
->operands
)];
1326 st_src_reg result_src
;
1327 st_dst_reg result_dst
;
1329 /* Quick peephole: Emit MAD(a, b, c) instead of ADD(MUL(a, b), c)
1331 if (ir
->operation
== ir_binop_add
) {
1332 if (try_emit_mad(ir
, 1))
1334 if (try_emit_mad(ir
, 0))
1338 /* Quick peephole: Emit OPCODE_MAD(-a, -b, a) instead of AND(a, NOT(b))
1340 if (ir
->operation
== ir_binop_logic_and
) {
1341 if (try_emit_mad_for_and_not(ir
, 1))
1343 if (try_emit_mad_for_and_not(ir
, 0))
1347 if (try_emit_sat(ir
))
1350 if (ir
->operation
== ir_quadop_vector
)
1351 assert(!"ir_quadop_vector should have been lowered");
1353 for (operand
= 0; operand
< ir
->get_num_operands(); operand
++) {
1354 this->result
.file
= PROGRAM_UNDEFINED
;
1355 ir
->operands
[operand
]->accept(this);
1356 if (this->result
.file
== PROGRAM_UNDEFINED
) {
1358 printf("Failed to get tree for expression operand:\n");
1359 ir
->operands
[operand
]->accept(&v
);
1362 op
[operand
] = this->result
;
1364 /* Matrix expression operands should have been broken down to vector
1365 * operations already.
1367 assert(!ir
->operands
[operand
]->type
->is_matrix());
1370 int vector_elements
= ir
->operands
[0]->type
->vector_elements
;
1371 if (ir
->operands
[1]) {
1372 vector_elements
= MAX2(vector_elements
,
1373 ir
->operands
[1]->type
->vector_elements
);
1376 this->result
.file
= PROGRAM_UNDEFINED
;
1378 /* Storage for our result. Ideally for an assignment we'd be using
1379 * the actual storage for the result here, instead.
1381 result_src
= get_temp(ir
->type
);
1382 /* convenience for the emit functions below. */
1383 result_dst
= st_dst_reg(result_src
);
1384 /* Limit writes to the channels that will be used by result_src later.
1385 * This does limit this temp's use as a temporary for multi-instruction
1388 result_dst
.writemask
= (1 << ir
->type
->vector_elements
) - 1;
1390 switch (ir
->operation
) {
1391 case ir_unop_logic_not
:
1392 if (result_dst
.type
!= GLSL_TYPE_FLOAT
)
1393 emit(ir
, TGSI_OPCODE_NOT
, result_dst
, op
[0]);
1395 /* Previously 'SEQ dst, src, 0.0' was used for this. However, many
1396 * older GPUs implement SEQ using multiple instructions (i915 uses two
1397 * SGE instructions and a MUL instruction). Since our logic values are
1398 * 0.0 and 1.0, 1-x also implements !x.
1400 op
[0].negate
= ~op
[0].negate
;
1401 emit(ir
, TGSI_OPCODE_ADD
, result_dst
, op
[0], st_src_reg_for_float(1.0));
1405 assert(result_dst
.type
== GLSL_TYPE_FLOAT
|| result_dst
.type
== GLSL_TYPE_INT
);
1406 if (result_dst
.type
== GLSL_TYPE_INT
)
1407 emit(ir
, TGSI_OPCODE_INEG
, result_dst
, op
[0]);
1409 op
[0].negate
= ~op
[0].negate
;
1414 assert(result_dst
.type
== GLSL_TYPE_FLOAT
);
1415 emit(ir
, TGSI_OPCODE_ABS
, result_dst
, op
[0]);
1418 emit(ir
, TGSI_OPCODE_SSG
, result_dst
, op
[0]);
1421 emit_scalar(ir
, TGSI_OPCODE_RCP
, result_dst
, op
[0]);
1425 emit_scalar(ir
, TGSI_OPCODE_EX2
, result_dst
, op
[0]);
1429 assert(!"not reached: should be handled by ir_explog_to_explog2");
1432 emit_scalar(ir
, TGSI_OPCODE_LG2
, result_dst
, op
[0]);
1435 emit_scalar(ir
, TGSI_OPCODE_SIN
, result_dst
, op
[0]);
1438 emit_scalar(ir
, TGSI_OPCODE_COS
, result_dst
, op
[0]);
1440 case ir_unop_sin_reduced
:
1441 emit_scs(ir
, TGSI_OPCODE_SIN
, result_dst
, op
[0]);
1443 case ir_unop_cos_reduced
:
1444 emit_scs(ir
, TGSI_OPCODE_COS
, result_dst
, op
[0]);
1448 emit(ir
, TGSI_OPCODE_DDX
, result_dst
, op
[0]);
1451 op
[0].negate
= ~op
[0].negate
;
1452 emit(ir
, TGSI_OPCODE_DDY
, result_dst
, op
[0]);
1455 case ir_unop_noise
: {
1456 /* At some point, a motivated person could add a better
1457 * implementation of noise. Currently not even the nvidia
1458 * binary drivers do anything more than this. In any case, the
1459 * place to do this is in the GL state tracker, not the poor
1462 emit(ir
, TGSI_OPCODE_MOV
, result_dst
, st_src_reg_for_float(0.5));
1467 emit(ir
, TGSI_OPCODE_ADD
, result_dst
, op
[0], op
[1]);
1470 emit(ir
, TGSI_OPCODE_SUB
, result_dst
, op
[0], op
[1]);
1474 emit(ir
, TGSI_OPCODE_MUL
, result_dst
, op
[0], op
[1]);
1477 if (result_dst
.type
== GLSL_TYPE_FLOAT
)
1478 assert(!"not reached: should be handled by ir_div_to_mul_rcp");
1480 emit(ir
, TGSI_OPCODE_DIV
, result_dst
, op
[0], op
[1]);
1483 if (result_dst
.type
== GLSL_TYPE_FLOAT
)
1484 assert(!"ir_binop_mod should have been converted to b * fract(a/b)");
1486 emit(ir
, TGSI_OPCODE_MOD
, result_dst
, op
[0], op
[1]);
1490 emit(ir
, TGSI_OPCODE_SLT
, result_dst
, op
[0], op
[1]);
1492 case ir_binop_greater
:
1493 emit(ir
, TGSI_OPCODE_SLT
, result_dst
, op
[1], op
[0]);
1495 case ir_binop_lequal
:
1496 emit(ir
, TGSI_OPCODE_SGE
, result_dst
, op
[1], op
[0]);
1498 case ir_binop_gequal
:
1499 emit(ir
, TGSI_OPCODE_SGE
, result_dst
, op
[0], op
[1]);
1501 case ir_binop_equal
:
1502 emit(ir
, TGSI_OPCODE_SEQ
, result_dst
, op
[0], op
[1]);
1504 case ir_binop_nequal
:
1505 emit(ir
, TGSI_OPCODE_SNE
, result_dst
, op
[0], op
[1]);
1507 case ir_binop_all_equal
:
1508 /* "==" operator producing a scalar boolean. */
1509 if (ir
->operands
[0]->type
->is_vector() ||
1510 ir
->operands
[1]->type
->is_vector()) {
1511 st_src_reg temp
= get_temp(native_integers
?
1512 glsl_type::get_instance(ir
->operands
[0]->type
->base_type
, 4, 1) :
1513 glsl_type::vec4_type
);
1515 if (native_integers
) {
1516 st_dst_reg temp_dst
= st_dst_reg(temp
);
1517 st_src_reg temp1
= st_src_reg(temp
), temp2
= st_src_reg(temp
);
1519 emit(ir
, TGSI_OPCODE_SEQ
, st_dst_reg(temp
), op
[0], op
[1]);
1521 /* Emit 1-3 AND operations to combine the SEQ results. */
1522 switch (ir
->operands
[0]->type
->vector_elements
) {
1526 temp_dst
.writemask
= WRITEMASK_Y
;
1527 temp1
.swizzle
= SWIZZLE_YYYY
;
1528 temp2
.swizzle
= SWIZZLE_ZZZZ
;
1529 emit(ir
, TGSI_OPCODE_AND
, temp_dst
, temp1
, temp2
);
1532 temp_dst
.writemask
= WRITEMASK_X
;
1533 temp1
.swizzle
= SWIZZLE_XXXX
;
1534 temp2
.swizzle
= SWIZZLE_YYYY
;
1535 emit(ir
, TGSI_OPCODE_AND
, temp_dst
, temp1
, temp2
);
1536 temp_dst
.writemask
= WRITEMASK_Y
;
1537 temp1
.swizzle
= SWIZZLE_ZZZZ
;
1538 temp2
.swizzle
= SWIZZLE_WWWW
;
1539 emit(ir
, TGSI_OPCODE_AND
, temp_dst
, temp1
, temp2
);
1542 temp1
.swizzle
= SWIZZLE_XXXX
;
1543 temp2
.swizzle
= SWIZZLE_YYYY
;
1544 emit(ir
, TGSI_OPCODE_AND
, result_dst
, temp1
, temp2
);
1546 emit(ir
, TGSI_OPCODE_SNE
, st_dst_reg(temp
), op
[0], op
[1]);
1548 /* After the dot-product, the value will be an integer on the
1549 * range [0,4]. Zero becomes 1.0, and positive values become zero.
1551 emit_dp(ir
, result_dst
, temp
, temp
, vector_elements
);
1553 /* Negating the result of the dot-product gives values on the range
1554 * [-4, 0]. Zero becomes 1.0, and negative values become zero.
1555 * This is achieved using SGE.
1557 st_src_reg sge_src
= result_src
;
1558 sge_src
.negate
= ~sge_src
.negate
;
1559 emit(ir
, TGSI_OPCODE_SGE
, result_dst
, sge_src
, st_src_reg_for_float(0.0));
1562 emit(ir
, TGSI_OPCODE_SEQ
, result_dst
, op
[0], op
[1]);
1565 case ir_binop_any_nequal
:
1566 /* "!=" operator producing a scalar boolean. */
1567 if (ir
->operands
[0]->type
->is_vector() ||
1568 ir
->operands
[1]->type
->is_vector()) {
1569 st_src_reg temp
= get_temp(native_integers
?
1570 glsl_type::get_instance(ir
->operands
[0]->type
->base_type
, 4, 1) :
1571 glsl_type::vec4_type
);
1572 emit(ir
, TGSI_OPCODE_SNE
, st_dst_reg(temp
), op
[0], op
[1]);
1574 if (native_integers
) {
1575 st_dst_reg temp_dst
= st_dst_reg(temp
);
1576 st_src_reg temp1
= st_src_reg(temp
), temp2
= st_src_reg(temp
);
1578 /* Emit 1-3 OR operations to combine the SNE results. */
1579 switch (ir
->operands
[0]->type
->vector_elements
) {
1583 temp_dst
.writemask
= WRITEMASK_Y
;
1584 temp1
.swizzle
= SWIZZLE_YYYY
;
1585 temp2
.swizzle
= SWIZZLE_ZZZZ
;
1586 emit(ir
, TGSI_OPCODE_OR
, temp_dst
, temp1
, temp2
);
1589 temp_dst
.writemask
= WRITEMASK_X
;
1590 temp1
.swizzle
= SWIZZLE_XXXX
;
1591 temp2
.swizzle
= SWIZZLE_YYYY
;
1592 emit(ir
, TGSI_OPCODE_OR
, temp_dst
, temp1
, temp2
);
1593 temp_dst
.writemask
= WRITEMASK_Y
;
1594 temp1
.swizzle
= SWIZZLE_ZZZZ
;
1595 temp2
.swizzle
= SWIZZLE_WWWW
;
1596 emit(ir
, TGSI_OPCODE_OR
, temp_dst
, temp1
, temp2
);
1599 temp1
.swizzle
= SWIZZLE_XXXX
;
1600 temp2
.swizzle
= SWIZZLE_YYYY
;
1601 emit(ir
, TGSI_OPCODE_OR
, result_dst
, temp1
, temp2
);
1603 /* After the dot-product, the value will be an integer on the
1604 * range [0,4]. Zero stays zero, and positive values become 1.0.
1606 glsl_to_tgsi_instruction
*const dp
=
1607 emit_dp(ir
, result_dst
, temp
, temp
, vector_elements
);
1608 if (this->prog
->Target
== GL_FRAGMENT_PROGRAM_ARB
) {
1609 /* The clamping to [0,1] can be done for free in the fragment
1610 * shader with a saturate.
1612 dp
->saturate
= true;
1614 /* Negating the result of the dot-product gives values on the range
1615 * [-4, 0]. Zero stays zero, and negative values become 1.0. This
1616 * achieved using SLT.
1618 st_src_reg slt_src
= result_src
;
1619 slt_src
.negate
= ~slt_src
.negate
;
1620 emit(ir
, TGSI_OPCODE_SLT
, result_dst
, slt_src
, st_src_reg_for_float(0.0));
1624 emit(ir
, TGSI_OPCODE_SNE
, result_dst
, op
[0], op
[1]);
1629 assert(ir
->operands
[0]->type
->is_vector());
1631 /* After the dot-product, the value will be an integer on the
1632 * range [0,4]. Zero stays zero, and positive values become 1.0.
1634 glsl_to_tgsi_instruction
*const dp
=
1635 emit_dp(ir
, result_dst
, op
[0], op
[0],
1636 ir
->operands
[0]->type
->vector_elements
);
1637 if (this->prog
->Target
== GL_FRAGMENT_PROGRAM_ARB
&&
1638 result_dst
.type
== GLSL_TYPE_FLOAT
) {
1639 /* The clamping to [0,1] can be done for free in the fragment
1640 * shader with a saturate.
1642 dp
->saturate
= true;
1643 } else if (result_dst
.type
== GLSL_TYPE_FLOAT
) {
1644 /* Negating the result of the dot-product gives values on the range
1645 * [-4, 0]. Zero stays zero, and negative values become 1.0. This
1646 * is achieved using SLT.
1648 st_src_reg slt_src
= result_src
;
1649 slt_src
.negate
= ~slt_src
.negate
;
1650 emit(ir
, TGSI_OPCODE_SLT
, result_dst
, slt_src
, st_src_reg_for_float(0.0));
1653 /* Use SNE 0 if integers are being used as boolean values. */
1654 emit(ir
, TGSI_OPCODE_SNE
, result_dst
, result_src
, st_src_reg_for_int(0));
1659 case ir_binop_logic_xor
:
1660 if (native_integers
)
1661 emit(ir
, TGSI_OPCODE_XOR
, result_dst
, op
[0], op
[1]);
1663 emit(ir
, TGSI_OPCODE_SNE
, result_dst
, op
[0], op
[1]);
1666 case ir_binop_logic_or
: {
1667 if (native_integers
) {
1668 /* If integers are used as booleans, we can use an actual "or"
1671 assert(native_integers
);
1672 emit(ir
, TGSI_OPCODE_OR
, result_dst
, op
[0], op
[1]);
1674 /* After the addition, the value will be an integer on the
1675 * range [0,2]. Zero stays zero, and positive values become 1.0.
1677 glsl_to_tgsi_instruction
*add
=
1678 emit(ir
, TGSI_OPCODE_ADD
, result_dst
, op
[0], op
[1]);
1679 if (this->prog
->Target
== GL_FRAGMENT_PROGRAM_ARB
) {
1680 /* The clamping to [0,1] can be done for free in the fragment
1681 * shader with a saturate if floats are being used as boolean values.
1683 add
->saturate
= true;
1685 /* Negating the result of the addition gives values on the range
1686 * [-2, 0]. Zero stays zero, and negative values become 1.0. This
1687 * is achieved using SLT.
1689 st_src_reg slt_src
= result_src
;
1690 slt_src
.negate
= ~slt_src
.negate
;
1691 emit(ir
, TGSI_OPCODE_SLT
, result_dst
, slt_src
, st_src_reg_for_float(0.0));
1697 case ir_binop_logic_and
:
1698 /* If native integers are disabled, the bool args are stored as float 0.0
1699 * or 1.0, so "mul" gives us "and". If they're enabled, just use the
1700 * actual AND opcode.
1702 if (native_integers
)
1703 emit(ir
, TGSI_OPCODE_AND
, result_dst
, op
[0], op
[1]);
1705 emit(ir
, TGSI_OPCODE_MUL
, result_dst
, op
[0], op
[1]);
1709 assert(ir
->operands
[0]->type
->is_vector());
1710 assert(ir
->operands
[0]->type
== ir
->operands
[1]->type
);
1711 emit_dp(ir
, result_dst
, op
[0], op
[1],
1712 ir
->operands
[0]->type
->vector_elements
);
1716 /* sqrt(x) = x * rsq(x). */
1717 emit_scalar(ir
, TGSI_OPCODE_RSQ
, result_dst
, op
[0]);
1718 emit(ir
, TGSI_OPCODE_MUL
, result_dst
, result_src
, op
[0]);
1719 /* For incoming channels <= 0, set the result to 0. */
1720 op
[0].negate
= ~op
[0].negate
;
1721 emit(ir
, TGSI_OPCODE_CMP
, result_dst
,
1722 op
[0], result_src
, st_src_reg_for_float(0.0));
1725 emit_scalar(ir
, TGSI_OPCODE_RSQ
, result_dst
, op
[0]);
1728 if (native_integers
) {
1729 emit(ir
, TGSI_OPCODE_I2F
, result_dst
, op
[0]);
1732 /* fallthrough to next case otherwise */
1734 if (native_integers
) {
1735 emit(ir
, TGSI_OPCODE_AND
, result_dst
, op
[0], st_src_reg_for_float(1.0));
1738 /* fallthrough to next case otherwise */
1741 /* Converting between signed and unsigned integers is a no-op. */
1745 if (native_integers
) {
1746 /* Booleans are stored as integers using ~0 for true and 0 for false.
1747 * GLSL requires that int(bool) return 1 for true and 0 for false.
1748 * This conversion is done with AND, but it could be done with NEG.
1750 emit(ir
, TGSI_OPCODE_AND
, result_dst
, op
[0], st_src_reg_for_int(1));
1752 /* Booleans and integers are both stored as floats when native
1753 * integers are disabled.
1759 if (native_integers
)
1760 emit(ir
, TGSI_OPCODE_F2I
, result_dst
, op
[0]);
1762 emit(ir
, TGSI_OPCODE_TRUNC
, result_dst
, op
[0]);
1765 emit(ir
, TGSI_OPCODE_SNE
, result_dst
, op
[0], st_src_reg_for_float(0.0));
1768 if (native_integers
)
1769 emit(ir
, TGSI_OPCODE_INEG
, result_dst
, op
[0]);
1771 emit(ir
, TGSI_OPCODE_SNE
, result_dst
, op
[0], st_src_reg_for_float(0.0));
1774 emit(ir
, TGSI_OPCODE_TRUNC
, result_dst
, op
[0]);
1777 op
[0].negate
= ~op
[0].negate
;
1778 emit(ir
, TGSI_OPCODE_FLR
, result_dst
, op
[0]);
1779 result_src
.negate
= ~result_src
.negate
;
1782 emit(ir
, TGSI_OPCODE_FLR
, result_dst
, op
[0]);
1785 emit(ir
, TGSI_OPCODE_FRC
, result_dst
, op
[0]);
1789 emit(ir
, TGSI_OPCODE_MIN
, result_dst
, op
[0], op
[1]);
1792 emit(ir
, TGSI_OPCODE_MAX
, result_dst
, op
[0], op
[1]);
1795 emit_scalar(ir
, TGSI_OPCODE_POW
, result_dst
, op
[0], op
[1]);
1798 case ir_unop_bit_not
:
1799 if (native_integers
) {
1800 emit(ir
, TGSI_OPCODE_NOT
, result_dst
, op
[0]);
1804 if (native_integers
) {
1805 emit(ir
, TGSI_OPCODE_U2F
, result_dst
, op
[0]);
1808 case ir_binop_lshift
:
1809 if (native_integers
) {
1810 emit(ir
, TGSI_OPCODE_SHL
, result_dst
, op
[0]);
1813 case ir_binop_rshift
:
1814 if (native_integers
) {
1815 emit(ir
, TGSI_OPCODE_ISHR
, result_dst
, op
[0]);
1818 case ir_binop_bit_and
:
1819 if (native_integers
) {
1820 emit(ir
, TGSI_OPCODE_AND
, result_dst
, op
[0]);
1823 case ir_binop_bit_xor
:
1824 if (native_integers
) {
1825 emit(ir
, TGSI_OPCODE_XOR
, result_dst
, op
[0]);
1828 case ir_binop_bit_or
:
1829 if (native_integers
) {
1830 emit(ir
, TGSI_OPCODE_OR
, result_dst
, op
[0]);
1833 case ir_unop_round_even
:
1834 assert(!"GLSL 1.30 features unsupported");
1837 case ir_quadop_vector
:
1838 /* This operation should have already been handled.
1840 assert(!"Should not get here.");
1844 this->result
= result_src
;
1849 glsl_to_tgsi_visitor::visit(ir_swizzle
*ir
)
1855 /* Note that this is only swizzles in expressions, not those on the left
1856 * hand side of an assignment, which do write masking. See ir_assignment
1860 ir
->val
->accept(this);
1862 assert(src
.file
!= PROGRAM_UNDEFINED
);
1864 for (i
= 0; i
< 4; i
++) {
1865 if (i
< ir
->type
->vector_elements
) {
1868 swizzle
[i
] = GET_SWZ(src
.swizzle
, ir
->mask
.x
);
1871 swizzle
[i
] = GET_SWZ(src
.swizzle
, ir
->mask
.y
);
1874 swizzle
[i
] = GET_SWZ(src
.swizzle
, ir
->mask
.z
);
1877 swizzle
[i
] = GET_SWZ(src
.swizzle
, ir
->mask
.w
);
1881 /* If the type is smaller than a vec4, replicate the last
1884 swizzle
[i
] = swizzle
[ir
->type
->vector_elements
- 1];
1888 src
.swizzle
= MAKE_SWIZZLE4(swizzle
[0], swizzle
[1], swizzle
[2], swizzle
[3]);
1894 glsl_to_tgsi_visitor::visit(ir_dereference_variable
*ir
)
1896 variable_storage
*entry
= find_variable_storage(ir
->var
);
1897 ir_variable
*var
= ir
->var
;
1900 switch (var
->mode
) {
1901 case ir_var_uniform
:
1902 entry
= new(mem_ctx
) variable_storage(var
, PROGRAM_UNIFORM
,
1904 this->variables
.push_tail(entry
);
1908 /* The linker assigns locations for varyings and attributes,
1909 * including deprecated builtins (like gl_Color), user-assign
1910 * generic attributes (glBindVertexLocation), and
1911 * user-defined varyings.
1913 * FINISHME: We would hit this path for function arguments. Fix!
1915 assert(var
->location
!= -1);
1916 entry
= new(mem_ctx
) variable_storage(var
,
1921 assert(var
->location
!= -1);
1922 entry
= new(mem_ctx
) variable_storage(var
,
1926 case ir_var_system_value
:
1927 entry
= new(mem_ctx
) variable_storage(var
,
1928 PROGRAM_SYSTEM_VALUE
,
1932 case ir_var_temporary
:
1933 entry
= new(mem_ctx
) variable_storage(var
, PROGRAM_TEMPORARY
,
1935 this->variables
.push_tail(entry
);
1937 next_temp
+= type_size(var
->type
);
1942 printf("Failed to make storage for %s\n", var
->name
);
1947 this->result
= st_src_reg(entry
->file
, entry
->index
, var
->type
);
1948 if (!native_integers
)
1949 this->result
.type
= GLSL_TYPE_FLOAT
;
1953 glsl_to_tgsi_visitor::visit(ir_dereference_array
*ir
)
1957 int element_size
= type_size(ir
->type
);
1959 index
= ir
->array_index
->constant_expression_value();
1961 ir
->array
->accept(this);
1965 src
.index
+= index
->value
.i
[0] * element_size
;
1967 /* Variable index array dereference. It eats the "vec4" of the
1968 * base of the array and an index that offsets the TGSI register
1971 ir
->array_index
->accept(this);
1973 st_src_reg index_reg
;
1975 if (element_size
== 1) {
1976 index_reg
= this->result
;
1978 index_reg
= get_temp(native_integers
?
1979 glsl_type::int_type
: glsl_type::float_type
);
1981 emit(ir
, TGSI_OPCODE_MUL
, st_dst_reg(index_reg
),
1982 this->result
, st_src_reg_for_type(index_reg
.type
, element_size
));
1985 /* If there was already a relative address register involved, add the
1986 * new and the old together to get the new offset.
1988 if (src
.reladdr
!= NULL
) {
1989 st_src_reg accum_reg
= get_temp(native_integers
?
1990 glsl_type::int_type
: glsl_type::float_type
);
1992 emit(ir
, TGSI_OPCODE_ADD
, st_dst_reg(accum_reg
),
1993 index_reg
, *src
.reladdr
);
1995 index_reg
= accum_reg
;
1998 src
.reladdr
= ralloc(mem_ctx
, st_src_reg
);
1999 memcpy(src
.reladdr
, &index_reg
, sizeof(index_reg
));
2002 /* If the type is smaller than a vec4, replicate the last channel out. */
2003 if (ir
->type
->is_scalar() || ir
->type
->is_vector())
2004 src
.swizzle
= swizzle_for_size(ir
->type
->vector_elements
);
2006 src
.swizzle
= SWIZZLE_NOOP
;
2012 glsl_to_tgsi_visitor::visit(ir_dereference_record
*ir
)
2015 const glsl_type
*struct_type
= ir
->record
->type
;
2018 ir
->record
->accept(this);
2020 for (i
= 0; i
< struct_type
->length
; i
++) {
2021 if (strcmp(struct_type
->fields
.structure
[i
].name
, ir
->field
) == 0)
2023 offset
+= type_size(struct_type
->fields
.structure
[i
].type
);
2026 /* If the type is smaller than a vec4, replicate the last channel out. */
2027 if (ir
->type
->is_scalar() || ir
->type
->is_vector())
2028 this->result
.swizzle
= swizzle_for_size(ir
->type
->vector_elements
);
2030 this->result
.swizzle
= SWIZZLE_NOOP
;
2032 this->result
.index
+= offset
;
2036 * We want to be careful in assignment setup to hit the actual storage
2037 * instead of potentially using a temporary like we might with the
2038 * ir_dereference handler.
2041 get_assignment_lhs(ir_dereference
*ir
, glsl_to_tgsi_visitor
*v
)
2043 /* The LHS must be a dereference. If the LHS is a variable indexed array
2044 * access of a vector, it must be separated into a series conditional moves
2045 * before reaching this point (see ir_vec_index_to_cond_assign).
2047 assert(ir
->as_dereference());
2048 ir_dereference_array
*deref_array
= ir
->as_dereference_array();
2050 assert(!deref_array
->array
->type
->is_vector());
2053 /* Use the rvalue deref handler for the most part. We'll ignore
2054 * swizzles in it and write swizzles using writemask, though.
2057 return st_dst_reg(v
->result
);
2061 * Process the condition of a conditional assignment
2063 * Examines the condition of a conditional assignment to generate the optimal
2064 * first operand of a \c CMP instruction. If the condition is a relational
2065 * operator with 0 (e.g., \c ir_binop_less), the value being compared will be
2066 * used as the source for the \c CMP instruction. Otherwise the comparison
2067 * is processed to a boolean result, and the boolean result is used as the
2068 * operand to the CMP instruction.
2071 glsl_to_tgsi_visitor::process_move_condition(ir_rvalue
*ir
)
2073 ir_rvalue
*src_ir
= ir
;
2075 bool switch_order
= false;
2077 ir_expression
*const expr
= ir
->as_expression();
2078 if ((expr
!= NULL
) && (expr
->get_num_operands() == 2)) {
2079 bool zero_on_left
= false;
2081 if (expr
->operands
[0]->is_zero()) {
2082 src_ir
= expr
->operands
[1];
2083 zero_on_left
= true;
2084 } else if (expr
->operands
[1]->is_zero()) {
2085 src_ir
= expr
->operands
[0];
2086 zero_on_left
= false;
2090 * (a < 0) T F F ( a < 0) T F F
2091 * (0 < a) F F T (-a < 0) F F T
2092 * (a <= 0) T T F (-a < 0) F F T (swap order of other operands)
2093 * (0 <= a) F T T ( a < 0) T F F (swap order of other operands)
2094 * (a > 0) F F T (-a < 0) F F T
2095 * (0 > a) T F F ( a < 0) T F F
2096 * (a >= 0) F T T ( a < 0) T F F (swap order of other operands)
2097 * (0 >= a) T T F (-a < 0) F F T (swap order of other operands)
2099 * Note that exchanging the order of 0 and 'a' in the comparison simply
2100 * means that the value of 'a' should be negated.
2103 switch (expr
->operation
) {
2105 switch_order
= false;
2106 negate
= zero_on_left
;
2109 case ir_binop_greater
:
2110 switch_order
= false;
2111 negate
= !zero_on_left
;
2114 case ir_binop_lequal
:
2115 switch_order
= true;
2116 negate
= !zero_on_left
;
2119 case ir_binop_gequal
:
2120 switch_order
= true;
2121 negate
= zero_on_left
;
2125 /* This isn't the right kind of comparison afterall, so make sure
2126 * the whole condition is visited.
2134 src_ir
->accept(this);
2136 /* We use the TGSI_OPCODE_CMP (a < 0 ? b : c) for conditional moves, and the
2137 * condition we produced is 0.0 or 1.0. By flipping the sign, we can
2138 * choose which value TGSI_OPCODE_CMP produces without an extra instruction
2139 * computing the condition.
2142 this->result
.negate
= ~this->result
.negate
;
2144 return switch_order
;
2148 glsl_to_tgsi_visitor::visit(ir_assignment
*ir
)
2154 ir
->rhs
->accept(this);
2157 l
= get_assignment_lhs(ir
->lhs
, this);
2159 /* FINISHME: This should really set to the correct maximal writemask for each
2160 * FINISHME: component written (in the loops below). This case can only
2161 * FINISHME: occur for matrices, arrays, and structures.
2163 if (ir
->write_mask
== 0) {
2164 assert(!ir
->lhs
->type
->is_scalar() && !ir
->lhs
->type
->is_vector());
2165 l
.writemask
= WRITEMASK_XYZW
;
2166 } else if (ir
->lhs
->type
->is_scalar() &&
2167 ir
->lhs
->variable_referenced()->mode
== ir_var_out
) {
2168 /* FINISHME: This hack makes writing to gl_FragDepth, which lives in the
2169 * FINISHME: W component of fragment shader output zero, work correctly.
2171 l
.writemask
= WRITEMASK_XYZW
;
2174 int first_enabled_chan
= 0;
2177 l
.writemask
= ir
->write_mask
;
2179 for (int i
= 0; i
< 4; i
++) {
2180 if (l
.writemask
& (1 << i
)) {
2181 first_enabled_chan
= GET_SWZ(r
.swizzle
, i
);
2186 /* Swizzle a small RHS vector into the channels being written.
2188 * glsl ir treats write_mask as dictating how many channels are
2189 * present on the RHS while TGSI treats write_mask as just
2190 * showing which channels of the vec4 RHS get written.
2192 for (int i
= 0; i
< 4; i
++) {
2193 if (l
.writemask
& (1 << i
))
2194 swizzles
[i
] = GET_SWZ(r
.swizzle
, rhs_chan
++);
2196 swizzles
[i
] = first_enabled_chan
;
2198 r
.swizzle
= MAKE_SWIZZLE4(swizzles
[0], swizzles
[1],
2199 swizzles
[2], swizzles
[3]);
2202 assert(l
.file
!= PROGRAM_UNDEFINED
);
2203 assert(r
.file
!= PROGRAM_UNDEFINED
);
2205 if (ir
->condition
) {
2206 const bool switch_order
= this->process_move_condition(ir
->condition
);
2207 st_src_reg condition
= this->result
;
2209 for (i
= 0; i
< type_size(ir
->lhs
->type
); i
++) {
2210 st_src_reg l_src
= st_src_reg(l
);
2211 st_src_reg condition_temp
= condition
;
2212 l_src
.swizzle
= swizzle_for_size(ir
->lhs
->type
->vector_elements
);
2214 if (native_integers
) {
2215 /* This is necessary because TGSI's CMP instruction expects the
2216 * condition to be a float, and we store booleans as integers.
2217 * If TGSI had a UCMP instruction or similar, this extra
2218 * instruction would not be necessary.
2220 condition_temp
= get_temp(glsl_type::vec4_type
);
2221 condition
.negate
= 0;
2222 emit(ir
, TGSI_OPCODE_I2F
, st_dst_reg(condition_temp
), condition
);
2223 condition_temp
.swizzle
= condition
.swizzle
;
2227 emit(ir
, TGSI_OPCODE_CMP
, l
, condition_temp
, l_src
, r
);
2229 emit(ir
, TGSI_OPCODE_CMP
, l
, condition_temp
, r
, l_src
);
2235 } else if (ir
->rhs
->as_expression() &&
2236 this->instructions
.get_tail() &&
2237 ir
->rhs
== ((glsl_to_tgsi_instruction
*)this->instructions
.get_tail())->ir
&&
2238 type_size(ir
->lhs
->type
) == 1 &&
2239 l
.writemask
== ((glsl_to_tgsi_instruction
*)this->instructions
.get_tail())->dst
.writemask
) {
2240 /* To avoid emitting an extra MOV when assigning an expression to a
2241 * variable, emit the last instruction of the expression again, but
2242 * replace the destination register with the target of the assignment.
2243 * Dead code elimination will remove the original instruction.
2245 glsl_to_tgsi_instruction
*inst
, *new_inst
;
2246 inst
= (glsl_to_tgsi_instruction
*)this->instructions
.get_tail();
2247 new_inst
= emit(ir
, inst
->op
, l
, inst
->src
[0], inst
->src
[1], inst
->src
[2]);
2248 new_inst
->saturate
= inst
->saturate
;
2249 inst
->dead_mask
= inst
->dst
.writemask
;
2251 for (i
= 0; i
< type_size(ir
->lhs
->type
); i
++) {
2252 emit(ir
, TGSI_OPCODE_MOV
, l
, r
);
2261 glsl_to_tgsi_visitor::visit(ir_constant
*ir
)
2264 GLfloat stack_vals
[4] = { 0 };
2265 gl_constant_value
*values
= (gl_constant_value
*) stack_vals
;
2266 GLenum gl_type
= GL_NONE
;
2268 static int in_array
= 0;
2269 gl_register_file file
= in_array
? PROGRAM_CONSTANT
: PROGRAM_IMMEDIATE
;
2271 /* Unfortunately, 4 floats is all we can get into
2272 * _mesa_add_typed_unnamed_constant. So, make a temp to store an
2273 * aggregate constant and move each constant value into it. If we
2274 * get lucky, copy propagation will eliminate the extra moves.
2276 if (ir
->type
->base_type
== GLSL_TYPE_STRUCT
) {
2277 st_src_reg temp_base
= get_temp(ir
->type
);
2278 st_dst_reg temp
= st_dst_reg(temp_base
);
2280 foreach_iter(exec_list_iterator
, iter
, ir
->components
) {
2281 ir_constant
*field_value
= (ir_constant
*)iter
.get();
2282 int size
= type_size(field_value
->type
);
2286 field_value
->accept(this);
2289 for (i
= 0; i
< (unsigned int)size
; i
++) {
2290 emit(ir
, TGSI_OPCODE_MOV
, temp
, src
);
2296 this->result
= temp_base
;
2300 if (ir
->type
->is_array()) {
2301 st_src_reg temp_base
= get_temp(ir
->type
);
2302 st_dst_reg temp
= st_dst_reg(temp_base
);
2303 int size
= type_size(ir
->type
->fields
.array
);
2308 for (i
= 0; i
< ir
->type
->length
; i
++) {
2309 ir
->array_elements
[i
]->accept(this);
2311 for (int j
= 0; j
< size
; j
++) {
2312 emit(ir
, TGSI_OPCODE_MOV
, temp
, src
);
2318 this->result
= temp_base
;
2323 if (ir
->type
->is_matrix()) {
2324 st_src_reg mat
= get_temp(ir
->type
);
2325 st_dst_reg mat_column
= st_dst_reg(mat
);
2327 for (i
= 0; i
< ir
->type
->matrix_columns
; i
++) {
2328 assert(ir
->type
->base_type
== GLSL_TYPE_FLOAT
);
2329 values
= (gl_constant_value
*) &ir
->value
.f
[i
* ir
->type
->vector_elements
];
2331 src
= st_src_reg(file
, -1, ir
->type
->base_type
);
2332 src
.index
= add_constant(file
,
2334 ir
->type
->vector_elements
,
2337 emit(ir
, TGSI_OPCODE_MOV
, mat_column
, src
);
2346 switch (ir
->type
->base_type
) {
2347 case GLSL_TYPE_FLOAT
:
2349 for (i
= 0; i
< ir
->type
->vector_elements
; i
++) {
2350 values
[i
].f
= ir
->value
.f
[i
];
2353 case GLSL_TYPE_UINT
:
2354 gl_type
= native_integers
? GL_UNSIGNED_INT
: GL_FLOAT
;
2355 for (i
= 0; i
< ir
->type
->vector_elements
; i
++) {
2356 if (native_integers
)
2357 values
[i
].u
= ir
->value
.u
[i
];
2359 values
[i
].f
= ir
->value
.u
[i
];
2363 gl_type
= native_integers
? GL_INT
: GL_FLOAT
;
2364 for (i
= 0; i
< ir
->type
->vector_elements
; i
++) {
2365 if (native_integers
)
2366 values
[i
].i
= ir
->value
.i
[i
];
2368 values
[i
].f
= ir
->value
.i
[i
];
2371 case GLSL_TYPE_BOOL
:
2372 gl_type
= native_integers
? GL_BOOL
: GL_FLOAT
;
2373 for (i
= 0; i
< ir
->type
->vector_elements
; i
++) {
2374 if (native_integers
)
2375 values
[i
].b
= ir
->value
.b
[i
];
2377 values
[i
].f
= ir
->value
.b
[i
];
2381 assert(!"Non-float/uint/int/bool constant");
2384 this->result
= st_src_reg(file
, -1, ir
->type
);
2385 this->result
.index
= add_constant(file
,
2387 ir
->type
->vector_elements
,
2389 &this->result
.swizzle
);
2393 glsl_to_tgsi_visitor::get_function_signature(ir_function_signature
*sig
)
2395 function_entry
*entry
;
2397 foreach_iter(exec_list_iterator
, iter
, this->function_signatures
) {
2398 entry
= (function_entry
*)iter
.get();
2400 if (entry
->sig
== sig
)
2404 entry
= ralloc(mem_ctx
, function_entry
);
2406 entry
->sig_id
= this->next_signature_id
++;
2407 entry
->bgn_inst
= NULL
;
2409 /* Allocate storage for all the parameters. */
2410 foreach_iter(exec_list_iterator
, iter
, sig
->parameters
) {
2411 ir_variable
*param
= (ir_variable
*)iter
.get();
2412 variable_storage
*storage
;
2414 storage
= find_variable_storage(param
);
2417 storage
= new(mem_ctx
) variable_storage(param
, PROGRAM_TEMPORARY
,
2419 this->variables
.push_tail(storage
);
2421 this->next_temp
+= type_size(param
->type
);
2424 if (!sig
->return_type
->is_void()) {
2425 entry
->return_reg
= get_temp(sig
->return_type
);
2427 entry
->return_reg
= undef_src
;
2430 this->function_signatures
.push_tail(entry
);
2435 glsl_to_tgsi_visitor::visit(ir_call
*ir
)
2437 glsl_to_tgsi_instruction
*call_inst
;
2438 ir_function_signature
*sig
= ir
->get_callee();
2439 function_entry
*entry
= get_function_signature(sig
);
2442 /* Process in parameters. */
2443 exec_list_iterator sig_iter
= sig
->parameters
.iterator();
2444 foreach_iter(exec_list_iterator
, iter
, *ir
) {
2445 ir_rvalue
*param_rval
= (ir_rvalue
*)iter
.get();
2446 ir_variable
*param
= (ir_variable
*)sig_iter
.get();
2448 if (param
->mode
== ir_var_in
||
2449 param
->mode
== ir_var_inout
) {
2450 variable_storage
*storage
= find_variable_storage(param
);
2453 param_rval
->accept(this);
2454 st_src_reg r
= this->result
;
2457 l
.file
= storage
->file
;
2458 l
.index
= storage
->index
;
2460 l
.writemask
= WRITEMASK_XYZW
;
2461 l
.cond_mask
= COND_TR
;
2463 for (i
= 0; i
< type_size(param
->type
); i
++) {
2464 emit(ir
, TGSI_OPCODE_MOV
, l
, r
);
2472 assert(!sig_iter
.has_next());
2474 /* Emit call instruction */
2475 call_inst
= emit(ir
, TGSI_OPCODE_CAL
);
2476 call_inst
->function
= entry
;
2478 /* Process out parameters. */
2479 sig_iter
= sig
->parameters
.iterator();
2480 foreach_iter(exec_list_iterator
, iter
, *ir
) {
2481 ir_rvalue
*param_rval
= (ir_rvalue
*)iter
.get();
2482 ir_variable
*param
= (ir_variable
*)sig_iter
.get();
2484 if (param
->mode
== ir_var_out
||
2485 param
->mode
== ir_var_inout
) {
2486 variable_storage
*storage
= find_variable_storage(param
);
2490 r
.file
= storage
->file
;
2491 r
.index
= storage
->index
;
2493 r
.swizzle
= SWIZZLE_NOOP
;
2496 param_rval
->accept(this);
2497 st_dst_reg l
= st_dst_reg(this->result
);
2499 for (i
= 0; i
< type_size(param
->type
); i
++) {
2500 emit(ir
, TGSI_OPCODE_MOV
, l
, r
);
2508 assert(!sig_iter
.has_next());
2510 /* Process return value. */
2511 this->result
= entry
->return_reg
;
2515 glsl_to_tgsi_visitor::visit(ir_texture
*ir
)
2517 st_src_reg result_src
, coord
, lod_info
, projector
, dx
, dy
, offset
;
2518 st_dst_reg result_dst
, coord_dst
;
2519 glsl_to_tgsi_instruction
*inst
= NULL
;
2520 unsigned opcode
= TGSI_OPCODE_NOP
;
2522 if (ir
->coordinate
) {
2523 ir
->coordinate
->accept(this);
2525 /* Put our coords in a temp. We'll need to modify them for shadow,
2526 * projection, or LOD, so the only case we'd use it as is is if
2527 * we're doing plain old texturing. The optimization passes on
2528 * glsl_to_tgsi_visitor should handle cleaning up our mess in that case.
2530 coord
= get_temp(glsl_type::vec4_type
);
2531 coord_dst
= st_dst_reg(coord
);
2532 emit(ir
, TGSI_OPCODE_MOV
, coord_dst
, this->result
);
2535 if (ir
->projector
) {
2536 ir
->projector
->accept(this);
2537 projector
= this->result
;
2540 /* Storage for our result. Ideally for an assignment we'd be using
2541 * the actual storage for the result here, instead.
2543 result_src
= get_temp(glsl_type::vec4_type
);
2544 result_dst
= st_dst_reg(result_src
);
2548 opcode
= TGSI_OPCODE_TEX
;
2551 opcode
= TGSI_OPCODE_TXB
;
2552 ir
->lod_info
.bias
->accept(this);
2553 lod_info
= this->result
;
2556 opcode
= TGSI_OPCODE_TXL
;
2557 ir
->lod_info
.lod
->accept(this);
2558 lod_info
= this->result
;
2561 opcode
= TGSI_OPCODE_TXD
;
2562 ir
->lod_info
.grad
.dPdx
->accept(this);
2564 ir
->lod_info
.grad
.dPdy
->accept(this);
2568 opcode
= TGSI_OPCODE_TXQ
;
2569 ir
->lod_info
.lod
->accept(this);
2570 lod_info
= this->result
;
2573 opcode
= TGSI_OPCODE_TXF
;
2574 ir
->lod_info
.lod
->accept(this);
2575 lod_info
= this->result
;
2577 ir
->offset
->accept(this);
2578 offset
= this->result
;
2583 const glsl_type
*sampler_type
= ir
->sampler
->type
;
2585 if (ir
->projector
) {
2586 if (opcode
== TGSI_OPCODE_TEX
) {
2587 /* Slot the projector in as the last component of the coord. */
2588 coord_dst
.writemask
= WRITEMASK_W
;
2589 emit(ir
, TGSI_OPCODE_MOV
, coord_dst
, projector
);
2590 coord_dst
.writemask
= WRITEMASK_XYZW
;
2591 opcode
= TGSI_OPCODE_TXP
;
2593 st_src_reg coord_w
= coord
;
2594 coord_w
.swizzle
= SWIZZLE_WWWW
;
2596 /* For the other TEX opcodes there's no projective version
2597 * since the last slot is taken up by LOD info. Do the
2598 * projective divide now.
2600 coord_dst
.writemask
= WRITEMASK_W
;
2601 emit(ir
, TGSI_OPCODE_RCP
, coord_dst
, projector
);
2603 /* In the case where we have to project the coordinates "by hand,"
2604 * the shadow comparator value must also be projected.
2606 st_src_reg tmp_src
= coord
;
2607 if (ir
->shadow_comparitor
) {
2608 /* Slot the shadow value in as the second to last component of the
2611 ir
->shadow_comparitor
->accept(this);
2613 tmp_src
= get_temp(glsl_type::vec4_type
);
2614 st_dst_reg tmp_dst
= st_dst_reg(tmp_src
);
2616 /* Projective division not allowed for array samplers. */
2617 assert(!sampler_type
->sampler_array
);
2619 tmp_dst
.writemask
= WRITEMASK_Z
;
2620 emit(ir
, TGSI_OPCODE_MOV
, tmp_dst
, this->result
);
2622 tmp_dst
.writemask
= WRITEMASK_XY
;
2623 emit(ir
, TGSI_OPCODE_MOV
, tmp_dst
, coord
);
2626 coord_dst
.writemask
= WRITEMASK_XYZ
;
2627 emit(ir
, TGSI_OPCODE_MUL
, coord_dst
, tmp_src
, coord_w
);
2629 coord_dst
.writemask
= WRITEMASK_XYZW
;
2630 coord
.swizzle
= SWIZZLE_XYZW
;
2634 /* If projection is done and the opcode is not TGSI_OPCODE_TXP, then the shadow
2635 * comparator was put in the correct place (and projected) by the code,
2636 * above, that handles by-hand projection.
2638 if (ir
->shadow_comparitor
&& (!ir
->projector
|| opcode
== TGSI_OPCODE_TXP
)) {
2639 /* Slot the shadow value in as the second to last component of the
2642 ir
->shadow_comparitor
->accept(this);
2644 /* XXX This will need to be updated for cubemap array samplers. */
2645 if (sampler_type
->sampler_dimensionality
== GLSL_SAMPLER_DIM_2D
&&
2646 sampler_type
->sampler_array
) {
2647 coord_dst
.writemask
= WRITEMASK_W
;
2649 coord_dst
.writemask
= WRITEMASK_Z
;
2652 emit(ir
, TGSI_OPCODE_MOV
, coord_dst
, this->result
);
2653 coord_dst
.writemask
= WRITEMASK_XYZW
;
2656 if (opcode
== TGSI_OPCODE_TXL
|| opcode
== TGSI_OPCODE_TXB
||
2657 opcode
== TGSI_OPCODE_TXF
) {
2658 /* TGSI stores LOD or LOD bias in the last channel of the coords. */
2659 coord_dst
.writemask
= WRITEMASK_W
;
2660 emit(ir
, TGSI_OPCODE_MOV
, coord_dst
, lod_info
);
2661 coord_dst
.writemask
= WRITEMASK_XYZW
;
2664 if (opcode
== TGSI_OPCODE_TXD
)
2665 inst
= emit(ir
, opcode
, result_dst
, coord
, dx
, dy
);
2666 else if (opcode
== TGSI_OPCODE_TXQ
)
2667 inst
= emit(ir
, opcode
, result_dst
, lod_info
);
2668 else if (opcode
== TGSI_OPCODE_TXF
) {
2669 inst
= emit(ir
, opcode
, result_dst
, coord
);
2671 inst
= emit(ir
, opcode
, result_dst
, coord
);
2673 if (ir
->shadow_comparitor
)
2674 inst
->tex_shadow
= GL_TRUE
;
2676 inst
->sampler
= _mesa_get_sampler_uniform_value(ir
->sampler
,
2677 this->shader_program
,
2681 inst
->tex_offset_num_offset
= 1;
2682 inst
->tex_offsets
[0].Index
= offset
.index
;
2683 inst
->tex_offsets
[0].File
= offset
.file
;
2684 inst
->tex_offsets
[0].SwizzleX
= GET_SWZ(offset
.swizzle
, 0);
2685 inst
->tex_offsets
[0].SwizzleY
= GET_SWZ(offset
.swizzle
, 1);
2686 inst
->tex_offsets
[0].SwizzleZ
= GET_SWZ(offset
.swizzle
, 2);
2689 switch (sampler_type
->sampler_dimensionality
) {
2690 case GLSL_SAMPLER_DIM_1D
:
2691 inst
->tex_target
= (sampler_type
->sampler_array
)
2692 ? TEXTURE_1D_ARRAY_INDEX
: TEXTURE_1D_INDEX
;
2694 case GLSL_SAMPLER_DIM_2D
:
2695 inst
->tex_target
= (sampler_type
->sampler_array
)
2696 ? TEXTURE_2D_ARRAY_INDEX
: TEXTURE_2D_INDEX
;
2698 case GLSL_SAMPLER_DIM_3D
:
2699 inst
->tex_target
= TEXTURE_3D_INDEX
;
2701 case GLSL_SAMPLER_DIM_CUBE
:
2702 inst
->tex_target
= TEXTURE_CUBE_INDEX
;
2704 case GLSL_SAMPLER_DIM_RECT
:
2705 inst
->tex_target
= TEXTURE_RECT_INDEX
;
2707 case GLSL_SAMPLER_DIM_BUF
:
2708 assert(!"FINISHME: Implement ARB_texture_buffer_object");
2710 case GLSL_SAMPLER_DIM_EXTERNAL
:
2711 inst
->tex_target
= TEXTURE_EXTERNAL_INDEX
;
2714 assert(!"Should not get here.");
2717 this->result
= result_src
;
2721 glsl_to_tgsi_visitor::visit(ir_return
*ir
)
2723 if (ir
->get_value()) {
2727 assert(current_function
);
2729 ir
->get_value()->accept(this);
2730 st_src_reg r
= this->result
;
2732 l
= st_dst_reg(current_function
->return_reg
);
2734 for (i
= 0; i
< type_size(current_function
->sig
->return_type
); i
++) {
2735 emit(ir
, TGSI_OPCODE_MOV
, l
, r
);
2741 emit(ir
, TGSI_OPCODE_RET
);
2745 glsl_to_tgsi_visitor::visit(ir_discard
*ir
)
2747 struct gl_fragment_program
*fp
= (struct gl_fragment_program
*)this->prog
;
2749 if (ir
->condition
) {
2750 ir
->condition
->accept(this);
2751 this->result
.negate
= ~this->result
.negate
;
2752 emit(ir
, TGSI_OPCODE_KIL
, undef_dst
, this->result
);
2754 emit(ir
, TGSI_OPCODE_KILP
);
2757 fp
->UsesKill
= GL_TRUE
;
2761 glsl_to_tgsi_visitor::visit(ir_if
*ir
)
2763 glsl_to_tgsi_instruction
*cond_inst
, *if_inst
;
2764 glsl_to_tgsi_instruction
*prev_inst
;
2766 prev_inst
= (glsl_to_tgsi_instruction
*)this->instructions
.get_tail();
2768 ir
->condition
->accept(this);
2769 assert(this->result
.file
!= PROGRAM_UNDEFINED
);
2771 if (this->options
->EmitCondCodes
) {
2772 cond_inst
= (glsl_to_tgsi_instruction
*)this->instructions
.get_tail();
2774 /* See if we actually generated any instruction for generating
2775 * the condition. If not, then cook up a move to a temp so we
2776 * have something to set cond_update on.
2778 if (cond_inst
== prev_inst
) {
2779 st_src_reg temp
= get_temp(glsl_type::bool_type
);
2780 cond_inst
= emit(ir
->condition
, TGSI_OPCODE_MOV
, st_dst_reg(temp
), result
);
2782 cond_inst
->cond_update
= GL_TRUE
;
2784 if_inst
= emit(ir
->condition
, TGSI_OPCODE_IF
);
2785 if_inst
->dst
.cond_mask
= COND_NE
;
2787 if_inst
= emit(ir
->condition
, TGSI_OPCODE_IF
, undef_dst
, this->result
);
2790 this->instructions
.push_tail(if_inst
);
2792 visit_exec_list(&ir
->then_instructions
, this);
2794 if (!ir
->else_instructions
.is_empty()) {
2795 emit(ir
->condition
, TGSI_OPCODE_ELSE
);
2796 visit_exec_list(&ir
->else_instructions
, this);
2799 if_inst
= emit(ir
->condition
, TGSI_OPCODE_ENDIF
);
2802 glsl_to_tgsi_visitor::glsl_to_tgsi_visitor()
2804 result
.file
= PROGRAM_UNDEFINED
;
2806 next_signature_id
= 1;
2808 current_function
= NULL
;
2809 num_address_regs
= 0;
2810 indirect_addr_temps
= false;
2811 indirect_addr_consts
= false;
2812 mem_ctx
= ralloc_context(NULL
);
2815 glsl_to_tgsi_visitor::~glsl_to_tgsi_visitor()
2817 ralloc_free(mem_ctx
);
2820 extern "C" void free_glsl_to_tgsi_visitor(glsl_to_tgsi_visitor
*v
)
2827 * Count resources used by the given gpu program (number of texture
2831 count_resources(glsl_to_tgsi_visitor
*v
, gl_program
*prog
)
2833 v
->samplers_used
= 0;
2835 foreach_iter(exec_list_iterator
, iter
, v
->instructions
) {
2836 glsl_to_tgsi_instruction
*inst
= (glsl_to_tgsi_instruction
*)iter
.get();
2838 if (is_tex_instruction(inst
->op
)) {
2839 v
->samplers_used
|= 1 << inst
->sampler
;
2841 prog
->SamplerTargets
[inst
->sampler
] =
2842 (gl_texture_index
)inst
->tex_target
;
2843 if (inst
->tex_shadow
) {
2844 prog
->ShadowSamplers
|= 1 << inst
->sampler
;
2849 prog
->SamplersUsed
= v
->samplers_used
;
2850 _mesa_update_shader_textures_used(prog
);
2854 set_uniform_initializer(struct gl_context
*ctx
, void *mem_ctx
,
2855 struct gl_shader_program
*shader_program
,
2856 const char *name
, const glsl_type
*type
,
2859 if (type
->is_record()) {
2860 ir_constant
*field_constant
;
2862 field_constant
= (ir_constant
*)val
->components
.get_head();
2864 for (unsigned int i
= 0; i
< type
->length
; i
++) {
2865 const glsl_type
*field_type
= type
->fields
.structure
[i
].type
;
2866 const char *field_name
= ralloc_asprintf(mem_ctx
, "%s.%s", name
,
2867 type
->fields
.structure
[i
].name
);
2868 set_uniform_initializer(ctx
, mem_ctx
, shader_program
, field_name
,
2869 field_type
, field_constant
);
2870 field_constant
= (ir_constant
*)field_constant
->next
;
2875 int loc
= _mesa_get_uniform_location(ctx
, shader_program
, name
);
2878 fail_link(shader_program
,
2879 "Couldn't find uniform for initializer %s\n", name
);
2883 for (unsigned int i
= 0; i
< (type
->is_array() ? type
->length
: 1); i
++) {
2884 ir_constant
*element
;
2885 const glsl_type
*element_type
;
2886 if (type
->is_array()) {
2887 element
= val
->array_elements
[i
];
2888 element_type
= type
->fields
.array
;
2891 element_type
= type
;
2896 if (element_type
->base_type
== GLSL_TYPE_BOOL
) {
2897 int *conv
= ralloc_array(mem_ctx
, int, element_type
->components());
2898 for (unsigned int j
= 0; j
< element_type
->components(); j
++) {
2899 conv
[j
] = element
->value
.b
[j
];
2901 values
= (void *)conv
;
2902 element_type
= glsl_type::get_instance(GLSL_TYPE_INT
,
2903 element_type
->vector_elements
,
2906 values
= &element
->value
;
2909 if (element_type
->is_matrix()) {
2910 _mesa_uniform_matrix(ctx
, shader_program
,
2911 element_type
->matrix_columns
,
2912 element_type
->vector_elements
,
2913 loc
, 1, GL_FALSE
, (GLfloat
*)values
);
2915 _mesa_uniform(ctx
, shader_program
, loc
, element_type
->matrix_columns
,
2916 values
, element_type
->gl_type
);
2924 * Scan/rewrite program to remove reads of custom (output) registers.
2925 * The passed type has to be either PROGRAM_OUTPUT or PROGRAM_VARYING
2926 * (for vertex shaders).
2927 * In GLSL shaders, varying vars can be read and written.
2928 * On some hardware, trying to read an output register causes trouble.
2929 * So, rewrite the program to use a temporary register in this case.
2931 * Based on _mesa_remove_output_reads from programopt.c.
2934 glsl_to_tgsi_visitor::remove_output_reads(gl_register_file type
)
2937 GLint outputMap
[VERT_RESULT_MAX
];
2938 GLint outputTypes
[VERT_RESULT_MAX
];
2939 GLuint numVaryingReads
= 0;
2940 GLboolean
*usedTemps
;
2941 GLuint firstTemp
= 0;
2943 usedTemps
= new GLboolean
[MAX_TEMPS
];
2947 _mesa_find_used_registers(prog
, PROGRAM_TEMPORARY
,
2948 usedTemps
, MAX_TEMPS
);
2950 assert(type
== PROGRAM_VARYING
|| type
== PROGRAM_OUTPUT
);
2951 assert(prog
->Target
== GL_VERTEX_PROGRAM_ARB
|| type
!= PROGRAM_VARYING
);
2953 for (i
= 0; i
< VERT_RESULT_MAX
; i
++)
2956 /* look for instructions which read from varying vars */
2957 foreach_iter(exec_list_iterator
, iter
, this->instructions
) {
2958 glsl_to_tgsi_instruction
*inst
= (glsl_to_tgsi_instruction
*)iter
.get();
2959 const GLuint numSrc
= num_inst_src_regs(inst
->op
);
2961 for (j
= 0; j
< numSrc
; j
++) {
2962 if (inst
->src
[j
].file
== type
) {
2963 /* replace the read with a temp reg */
2964 const GLuint var
= inst
->src
[j
].index
;
2965 if (outputMap
[var
] == -1) {
2967 outputMap
[var
] = _mesa_find_free_register(usedTemps
,
2970 outputTypes
[var
] = inst
->src
[j
].type
;
2971 firstTemp
= outputMap
[var
] + 1;
2973 inst
->src
[j
].file
= PROGRAM_TEMPORARY
;
2974 inst
->src
[j
].index
= outputMap
[var
];
2979 delete [] usedTemps
;
2981 if (numVaryingReads
== 0)
2982 return; /* nothing to be done */
2984 /* look for instructions which write to the varying vars identified above */
2985 foreach_iter(exec_list_iterator
, iter
, this->instructions
) {
2986 glsl_to_tgsi_instruction
*inst
= (glsl_to_tgsi_instruction
*)iter
.get();
2987 if (inst
->dst
.file
== type
&& outputMap
[inst
->dst
.index
] >= 0) {
2988 /* change inst to write to the temp reg, instead of the varying */
2989 inst
->dst
.file
= PROGRAM_TEMPORARY
;
2990 inst
->dst
.index
= outputMap
[inst
->dst
.index
];
2994 /* insert new MOV instructions at the end */
2995 for (i
= 0; i
< VERT_RESULT_MAX
; i
++) {
2996 if (outputMap
[i
] >= 0) {
2997 /* MOV VAR[i], TEMP[tmp]; */
2998 st_src_reg src
= st_src_reg(PROGRAM_TEMPORARY
, outputMap
[i
], outputTypes
[i
]);
2999 st_dst_reg dst
= st_dst_reg(type
, WRITEMASK_XYZW
, outputTypes
[i
]);
3001 this->emit(NULL
, TGSI_OPCODE_MOV
, dst
, src
);
3007 * Returns the mask of channels (bitmask of WRITEMASK_X,Y,Z,W) which
3008 * are read from the given src in this instruction
3011 get_src_arg_mask(st_dst_reg dst
, st_src_reg src
)
3013 int read_mask
= 0, comp
;
3015 /* Now, given the src swizzle and the written channels, find which
3016 * components are actually read
3018 for (comp
= 0; comp
< 4; ++comp
) {
3019 const unsigned coord
= GET_SWZ(src
.swizzle
, comp
);
3021 if (dst
.writemask
& (1 << comp
) && coord
<= SWIZZLE_W
)
3022 read_mask
|= 1 << coord
;
3029 * This pass replaces CMP T0, T1 T2 T0 with MOV T0, T2 when the CMP
3030 * instruction is the first instruction to write to register T0. There are
3031 * several lowering passes done in GLSL IR (e.g. branches and
3032 * relative addressing) that create a large number of conditional assignments
3033 * that ir_to_mesa converts to CMP instructions like the one mentioned above.
3035 * Here is why this conversion is safe:
3036 * CMP T0, T1 T2 T0 can be expanded to:
3042 * If (T1 < 0.0) evaluates to true then our replacement MOV T0, T2 is the same
3043 * as the original program. If (T1 < 0.0) evaluates to false, executing
3044 * MOV T0, T0 will store a garbage value in T0 since T0 is uninitialized.
3045 * Therefore, it doesn't matter that we are replacing MOV T0, T0 with MOV T0, T2
3046 * because any instruction that was going to read from T0 after this was going
3047 * to read a garbage value anyway.
3050 glsl_to_tgsi_visitor::simplify_cmp(void)
3052 unsigned *tempWrites
;
3053 unsigned outputWrites
[MAX_PROGRAM_OUTPUTS
];
3055 tempWrites
= new unsigned[MAX_TEMPS
];
3059 memset(tempWrites
, 0, sizeof(tempWrites
));
3060 memset(outputWrites
, 0, sizeof(outputWrites
));
3062 foreach_iter(exec_list_iterator
, iter
, this->instructions
) {
3063 glsl_to_tgsi_instruction
*inst
= (glsl_to_tgsi_instruction
*)iter
.get();
3064 unsigned prevWriteMask
= 0;
3066 /* Give up if we encounter relative addressing or flow control. */
3067 if (inst
->dst
.reladdr
||
3068 tgsi_get_opcode_info(inst
->op
)->is_branch
||
3069 inst
->op
== TGSI_OPCODE_BGNSUB
||
3070 inst
->op
== TGSI_OPCODE_CONT
||
3071 inst
->op
== TGSI_OPCODE_END
||
3072 inst
->op
== TGSI_OPCODE_ENDSUB
||
3073 inst
->op
== TGSI_OPCODE_RET
) {
3077 if (inst
->dst
.file
== PROGRAM_OUTPUT
) {
3078 assert(inst
->dst
.index
< MAX_PROGRAM_OUTPUTS
);
3079 prevWriteMask
= outputWrites
[inst
->dst
.index
];
3080 outputWrites
[inst
->dst
.index
] |= inst
->dst
.writemask
;
3081 } else if (inst
->dst
.file
== PROGRAM_TEMPORARY
) {
3082 assert(inst
->dst
.index
< MAX_TEMPS
);
3083 prevWriteMask
= tempWrites
[inst
->dst
.index
];
3084 tempWrites
[inst
->dst
.index
] |= inst
->dst
.writemask
;
3087 /* For a CMP to be considered a conditional write, the destination
3088 * register and source register two must be the same. */
3089 if (inst
->op
== TGSI_OPCODE_CMP
3090 && !(inst
->dst
.writemask
& prevWriteMask
)
3091 && inst
->src
[2].file
== inst
->dst
.file
3092 && inst
->src
[2].index
== inst
->dst
.index
3093 && inst
->dst
.writemask
== get_src_arg_mask(inst
->dst
, inst
->src
[2])) {
3095 inst
->op
= TGSI_OPCODE_MOV
;
3096 inst
->src
[0] = inst
->src
[1];
3100 delete [] tempWrites
;
3103 /* Replaces all references to a temporary register index with another index. */
3105 glsl_to_tgsi_visitor::rename_temp_register(int index
, int new_index
)
3107 foreach_iter(exec_list_iterator
, iter
, this->instructions
) {
3108 glsl_to_tgsi_instruction
*inst
= (glsl_to_tgsi_instruction
*)iter
.get();
3111 for (j
=0; j
< num_inst_src_regs(inst
->op
); j
++) {
3112 if (inst
->src
[j
].file
== PROGRAM_TEMPORARY
&&
3113 inst
->src
[j
].index
== index
) {
3114 inst
->src
[j
].index
= new_index
;
3118 if (inst
->dst
.file
== PROGRAM_TEMPORARY
&& inst
->dst
.index
== index
) {
3119 inst
->dst
.index
= new_index
;
3125 glsl_to_tgsi_visitor::get_first_temp_read(int index
)
3127 int depth
= 0; /* loop depth */
3128 int loop_start
= -1; /* index of the first active BGNLOOP (if any) */
3131 foreach_iter(exec_list_iterator
, iter
, this->instructions
) {
3132 glsl_to_tgsi_instruction
*inst
= (glsl_to_tgsi_instruction
*)iter
.get();
3134 for (j
=0; j
< num_inst_src_regs(inst
->op
); j
++) {
3135 if (inst
->src
[j
].file
== PROGRAM_TEMPORARY
&&
3136 inst
->src
[j
].index
== index
) {
3137 return (depth
== 0) ? i
: loop_start
;
3141 if (inst
->op
== TGSI_OPCODE_BGNLOOP
) {
3144 } else if (inst
->op
== TGSI_OPCODE_ENDLOOP
) {
3157 glsl_to_tgsi_visitor::get_first_temp_write(int index
)
3159 int depth
= 0; /* loop depth */
3160 int loop_start
= -1; /* index of the first active BGNLOOP (if any) */
3163 foreach_iter(exec_list_iterator
, iter
, this->instructions
) {
3164 glsl_to_tgsi_instruction
*inst
= (glsl_to_tgsi_instruction
*)iter
.get();
3166 if (inst
->dst
.file
== PROGRAM_TEMPORARY
&& inst
->dst
.index
== index
) {
3167 return (depth
== 0) ? i
: loop_start
;
3170 if (inst
->op
== TGSI_OPCODE_BGNLOOP
) {
3173 } else if (inst
->op
== TGSI_OPCODE_ENDLOOP
) {
3186 glsl_to_tgsi_visitor::get_last_temp_read(int index
)
3188 int depth
= 0; /* loop depth */
3189 int last
= -1; /* index of last instruction that reads the temporary */
3192 foreach_iter(exec_list_iterator
, iter
, this->instructions
) {
3193 glsl_to_tgsi_instruction
*inst
= (glsl_to_tgsi_instruction
*)iter
.get();
3195 for (j
=0; j
< num_inst_src_regs(inst
->op
); j
++) {
3196 if (inst
->src
[j
].file
== PROGRAM_TEMPORARY
&&
3197 inst
->src
[j
].index
== index
) {
3198 last
= (depth
== 0) ? i
: -2;
3202 if (inst
->op
== TGSI_OPCODE_BGNLOOP
)
3204 else if (inst
->op
== TGSI_OPCODE_ENDLOOP
)
3205 if (--depth
== 0 && last
== -2)
3217 glsl_to_tgsi_visitor::get_last_temp_write(int index
)
3219 int depth
= 0; /* loop depth */
3220 int last
= -1; /* index of last instruction that writes to the temporary */
3223 foreach_iter(exec_list_iterator
, iter
, this->instructions
) {
3224 glsl_to_tgsi_instruction
*inst
= (glsl_to_tgsi_instruction
*)iter
.get();
3226 if (inst
->dst
.file
== PROGRAM_TEMPORARY
&& inst
->dst
.index
== index
)
3227 last
= (depth
== 0) ? i
: -2;
3229 if (inst
->op
== TGSI_OPCODE_BGNLOOP
)
3231 else if (inst
->op
== TGSI_OPCODE_ENDLOOP
)
3232 if (--depth
== 0 && last
== -2)
3244 * On a basic block basis, tracks available PROGRAM_TEMPORARY register
3245 * channels for copy propagation and updates following instructions to
3246 * use the original versions.
3248 * The glsl_to_tgsi_visitor lazily produces code assuming that this pass
3249 * will occur. As an example, a TXP production before this pass:
3251 * 0: MOV TEMP[1], INPUT[4].xyyy;
3252 * 1: MOV TEMP[1].w, INPUT[4].wwww;
3253 * 2: TXP TEMP[2], TEMP[1], texture[0], 2D;
3257 * 0: MOV TEMP[1], INPUT[4].xyyy;
3258 * 1: MOV TEMP[1].w, INPUT[4].wwww;
3259 * 2: TXP TEMP[2], INPUT[4].xyyw, texture[0], 2D;
3261 * which allows for dead code elimination on TEMP[1]'s writes.
3264 glsl_to_tgsi_visitor::copy_propagate(void)
3266 glsl_to_tgsi_instruction
**acp
= rzalloc_array(mem_ctx
,
3267 glsl_to_tgsi_instruction
*,
3268 this->next_temp
* 4);
3269 int *acp_level
= rzalloc_array(mem_ctx
, int, this->next_temp
* 4);
3272 foreach_iter(exec_list_iterator
, iter
, this->instructions
) {
3273 glsl_to_tgsi_instruction
*inst
= (glsl_to_tgsi_instruction
*)iter
.get();
3275 assert(inst
->dst
.file
!= PROGRAM_TEMPORARY
3276 || inst
->dst
.index
< this->next_temp
);
3278 /* First, do any copy propagation possible into the src regs. */
3279 for (int r
= 0; r
< 3; r
++) {
3280 glsl_to_tgsi_instruction
*first
= NULL
;
3282 int acp_base
= inst
->src
[r
].index
* 4;
3284 if (inst
->src
[r
].file
!= PROGRAM_TEMPORARY
||
3285 inst
->src
[r
].reladdr
)
3288 /* See if we can find entries in the ACP consisting of MOVs
3289 * from the same src register for all the swizzled channels
3290 * of this src register reference.
3292 for (int i
= 0; i
< 4; i
++) {
3293 int src_chan
= GET_SWZ(inst
->src
[r
].swizzle
, i
);
3294 glsl_to_tgsi_instruction
*copy_chan
= acp
[acp_base
+ src_chan
];
3301 assert(acp_level
[acp_base
+ src_chan
] <= level
);
3306 if (first
->src
[0].file
!= copy_chan
->src
[0].file
||
3307 first
->src
[0].index
!= copy_chan
->src
[0].index
) {
3315 /* We've now validated that we can copy-propagate to
3316 * replace this src register reference. Do it.
3318 inst
->src
[r
].file
= first
->src
[0].file
;
3319 inst
->src
[r
].index
= first
->src
[0].index
;
3322 for (int i
= 0; i
< 4; i
++) {
3323 int src_chan
= GET_SWZ(inst
->src
[r
].swizzle
, i
);
3324 glsl_to_tgsi_instruction
*copy_inst
= acp
[acp_base
+ src_chan
];
3325 swizzle
|= (GET_SWZ(copy_inst
->src
[0].swizzle
, src_chan
) <<
3328 inst
->src
[r
].swizzle
= swizzle
;
3333 case TGSI_OPCODE_BGNLOOP
:
3334 case TGSI_OPCODE_ENDLOOP
:
3335 /* End of a basic block, clear the ACP entirely. */
3336 memset(acp
, 0, sizeof(*acp
) * this->next_temp
* 4);
3339 case TGSI_OPCODE_IF
:
3343 case TGSI_OPCODE_ENDIF
:
3344 case TGSI_OPCODE_ELSE
:
3345 /* Clear all channels written inside the block from the ACP, but
3346 * leaving those that were not touched.
3348 for (int r
= 0; r
< this->next_temp
; r
++) {
3349 for (int c
= 0; c
< 4; c
++) {
3350 if (!acp
[4 * r
+ c
])
3353 if (acp_level
[4 * r
+ c
] >= level
)
3354 acp
[4 * r
+ c
] = NULL
;
3357 if (inst
->op
== TGSI_OPCODE_ENDIF
)
3362 /* Continuing the block, clear any written channels from
3365 if (inst
->dst
.file
== PROGRAM_TEMPORARY
&& inst
->dst
.reladdr
) {
3366 /* Any temporary might be written, so no copy propagation
3367 * across this instruction.
3369 memset(acp
, 0, sizeof(*acp
) * this->next_temp
* 4);
3370 } else if (inst
->dst
.file
== PROGRAM_OUTPUT
&&
3371 inst
->dst
.reladdr
) {
3372 /* Any output might be written, so no copy propagation
3373 * from outputs across this instruction.
3375 for (int r
= 0; r
< this->next_temp
; r
++) {
3376 for (int c
= 0; c
< 4; c
++) {
3377 if (!acp
[4 * r
+ c
])
3380 if (acp
[4 * r
+ c
]->src
[0].file
== PROGRAM_OUTPUT
)
3381 acp
[4 * r
+ c
] = NULL
;
3384 } else if (inst
->dst
.file
== PROGRAM_TEMPORARY
||
3385 inst
->dst
.file
== PROGRAM_OUTPUT
) {
3386 /* Clear where it's used as dst. */
3387 if (inst
->dst
.file
== PROGRAM_TEMPORARY
) {
3388 for (int c
= 0; c
< 4; c
++) {
3389 if (inst
->dst
.writemask
& (1 << c
)) {
3390 acp
[4 * inst
->dst
.index
+ c
] = NULL
;
3395 /* Clear where it's used as src. */
3396 for (int r
= 0; r
< this->next_temp
; r
++) {
3397 for (int c
= 0; c
< 4; c
++) {
3398 if (!acp
[4 * r
+ c
])
3401 int src_chan
= GET_SWZ(acp
[4 * r
+ c
]->src
[0].swizzle
, c
);
3403 if (acp
[4 * r
+ c
]->src
[0].file
== inst
->dst
.file
&&
3404 acp
[4 * r
+ c
]->src
[0].index
== inst
->dst
.index
&&
3405 inst
->dst
.writemask
& (1 << src_chan
))
3407 acp
[4 * r
+ c
] = NULL
;
3415 /* If this is a copy, add it to the ACP. */
3416 if (inst
->op
== TGSI_OPCODE_MOV
&&
3417 inst
->dst
.file
== PROGRAM_TEMPORARY
&&
3418 !inst
->dst
.reladdr
&&
3420 !inst
->src
[0].reladdr
&&
3421 !inst
->src
[0].negate
) {
3422 for (int i
= 0; i
< 4; i
++) {
3423 if (inst
->dst
.writemask
& (1 << i
)) {
3424 acp
[4 * inst
->dst
.index
+ i
] = inst
;
3425 acp_level
[4 * inst
->dst
.index
+ i
] = level
;
3431 ralloc_free(acp_level
);
3436 * Tracks available PROGRAM_TEMPORARY registers for dead code elimination.
3438 * The glsl_to_tgsi_visitor lazily produces code assuming that this pass
3439 * will occur. As an example, a TXP production after copy propagation but
3442 * 0: MOV TEMP[1], INPUT[4].xyyy;
3443 * 1: MOV TEMP[1].w, INPUT[4].wwww;
3444 * 2: TXP TEMP[2], INPUT[4].xyyw, texture[0], 2D;
3446 * and after this pass:
3448 * 0: TXP TEMP[2], INPUT[4].xyyw, texture[0], 2D;
3450 * FIXME: assumes that all functions are inlined (no support for BGNSUB/ENDSUB)
3451 * FIXME: doesn't eliminate all dead code inside of loops; it steps around them
3454 glsl_to_tgsi_visitor::eliminate_dead_code(void)
3458 for (i
=0; i
< this->next_temp
; i
++) {
3459 int last_read
= get_last_temp_read(i
);
3462 foreach_iter(exec_list_iterator
, iter
, this->instructions
) {
3463 glsl_to_tgsi_instruction
*inst
= (glsl_to_tgsi_instruction
*)iter
.get();
3465 if (inst
->dst
.file
== PROGRAM_TEMPORARY
&& inst
->dst
.index
== i
&&
3478 * On a basic block basis, tracks available PROGRAM_TEMPORARY registers for dead
3479 * code elimination. This is less primitive than eliminate_dead_code(), as it
3480 * is per-channel and can detect consecutive writes without a read between them
3481 * as dead code. However, there is some dead code that can be eliminated by
3482 * eliminate_dead_code() but not this function - for example, this function
3483 * cannot eliminate an instruction writing to a register that is never read and
3484 * is the only instruction writing to that register.
3486 * The glsl_to_tgsi_visitor lazily produces code assuming that this pass
3490 glsl_to_tgsi_visitor::eliminate_dead_code_advanced(void)
3492 glsl_to_tgsi_instruction
**writes
= rzalloc_array(mem_ctx
,
3493 glsl_to_tgsi_instruction
*,
3494 this->next_temp
* 4);
3495 int *write_level
= rzalloc_array(mem_ctx
, int, this->next_temp
* 4);
3499 foreach_iter(exec_list_iterator
, iter
, this->instructions
) {
3500 glsl_to_tgsi_instruction
*inst
= (glsl_to_tgsi_instruction
*)iter
.get();
3502 assert(inst
->dst
.file
!= PROGRAM_TEMPORARY
3503 || inst
->dst
.index
< this->next_temp
);
3506 case TGSI_OPCODE_BGNLOOP
:
3507 case TGSI_OPCODE_ENDLOOP
:
3508 /* End of a basic block, clear the write array entirely.
3509 * FIXME: This keeps us from killing dead code when the writes are
3510 * on either side of a loop, even when the register isn't touched
3513 memset(writes
, 0, sizeof(*writes
) * this->next_temp
* 4);
3516 case TGSI_OPCODE_ENDIF
:
3517 case TGSI_OPCODE_ELSE
:
3518 /* Promote the recorded level all channels written inside the preceding
3519 * if or else block to the level above the if/else block.
3521 for (int r
= 0; r
< this->next_temp
; r
++) {
3522 for (int c
= 0; c
< 4; c
++) {
3523 if (!writes
[4 * r
+ c
])
3526 if (write_level
[4 * r
+ c
] == level
)
3527 write_level
[4 * r
+ c
] = level
-1;
3531 if(inst
->op
== TGSI_OPCODE_ENDIF
)
3536 case TGSI_OPCODE_IF
:
3538 /* fallthrough to default case to mark the condition as read */
3541 /* Continuing the block, clear any channels from the write array that
3542 * are read by this instruction.
3544 for (unsigned i
= 0; i
< Elements(inst
->src
); i
++) {
3545 if (inst
->src
[i
].file
== PROGRAM_TEMPORARY
&& inst
->src
[i
].reladdr
){
3546 /* Any temporary might be read, so no dead code elimination
3547 * across this instruction.
3549 memset(writes
, 0, sizeof(*writes
) * this->next_temp
* 4);
3550 } else if (inst
->src
[i
].file
== PROGRAM_TEMPORARY
) {
3551 /* Clear where it's used as src. */
3552 int src_chans
= 1 << GET_SWZ(inst
->src
[i
].swizzle
, 0);
3553 src_chans
|= 1 << GET_SWZ(inst
->src
[i
].swizzle
, 1);
3554 src_chans
|= 1 << GET_SWZ(inst
->src
[i
].swizzle
, 2);
3555 src_chans
|= 1 << GET_SWZ(inst
->src
[i
].swizzle
, 3);
3557 for (int c
= 0; c
< 4; c
++) {
3558 if (src_chans
& (1 << c
)) {
3559 writes
[4 * inst
->src
[i
].index
+ c
] = NULL
;
3567 /* If this instruction writes to a temporary, add it to the write array.
3568 * If there is already an instruction in the write array for one or more
3569 * of the channels, flag that channel write as dead.
3571 if (inst
->dst
.file
== PROGRAM_TEMPORARY
&&
3572 !inst
->dst
.reladdr
&&
3574 for (int c
= 0; c
< 4; c
++) {
3575 if (inst
->dst
.writemask
& (1 << c
)) {
3576 if (writes
[4 * inst
->dst
.index
+ c
]) {
3577 if (write_level
[4 * inst
->dst
.index
+ c
] < level
)
3580 writes
[4 * inst
->dst
.index
+ c
]->dead_mask
|= (1 << c
);
3582 writes
[4 * inst
->dst
.index
+ c
] = inst
;
3583 write_level
[4 * inst
->dst
.index
+ c
] = level
;
3589 /* Anything still in the write array at this point is dead code. */
3590 for (int r
= 0; r
< this->next_temp
; r
++) {
3591 for (int c
= 0; c
< 4; c
++) {
3592 glsl_to_tgsi_instruction
*inst
= writes
[4 * r
+ c
];
3594 inst
->dead_mask
|= (1 << c
);
3598 /* Now actually remove the instructions that are completely dead and update
3599 * the writemask of other instructions with dead channels.
3601 foreach_iter(exec_list_iterator
, iter
, this->instructions
) {
3602 glsl_to_tgsi_instruction
*inst
= (glsl_to_tgsi_instruction
*)iter
.get();
3604 if (!inst
->dead_mask
|| !inst
->dst
.writemask
)
3606 else if ((inst
->dst
.writemask
& ~inst
->dead_mask
) == 0) {
3611 inst
->dst
.writemask
&= ~(inst
->dead_mask
);
3614 ralloc_free(write_level
);
3615 ralloc_free(writes
);
3620 /* Merges temporary registers together where possible to reduce the number of
3621 * registers needed to run a program.
3623 * Produces optimal code only after copy propagation and dead code elimination
3626 glsl_to_tgsi_visitor::merge_registers(void)
3628 int *last_reads
= rzalloc_array(mem_ctx
, int, this->next_temp
);
3629 int *first_writes
= rzalloc_array(mem_ctx
, int, this->next_temp
);
3632 /* Read the indices of the last read and first write to each temp register
3633 * into an array so that we don't have to traverse the instruction list as
3635 for (i
=0; i
< this->next_temp
; i
++) {
3636 last_reads
[i
] = get_last_temp_read(i
);
3637 first_writes
[i
] = get_first_temp_write(i
);
3640 /* Start looking for registers with non-overlapping usages that can be
3641 * merged together. */
3642 for (i
=0; i
< this->next_temp
; i
++) {
3643 /* Don't touch unused registers. */
3644 if (last_reads
[i
] < 0 || first_writes
[i
] < 0) continue;
3646 for (j
=0; j
< this->next_temp
; j
++) {
3647 /* Don't touch unused registers. */
3648 if (last_reads
[j
] < 0 || first_writes
[j
] < 0) continue;
3650 /* We can merge the two registers if the first write to j is after or
3651 * in the same instruction as the last read from i. Note that the
3652 * register at index i will always be used earlier or at the same time
3653 * as the register at index j. */
3654 if (first_writes
[i
] <= first_writes
[j
] &&
3655 last_reads
[i
] <= first_writes
[j
])
3657 rename_temp_register(j
, i
); /* Replace all references to j with i.*/
3659 /* Update the first_writes and last_reads arrays with the new
3660 * values for the merged register index, and mark the newly unused
3661 * register index as such. */
3662 last_reads
[i
] = last_reads
[j
];
3663 first_writes
[j
] = -1;
3669 ralloc_free(last_reads
);
3670 ralloc_free(first_writes
);
3673 /* Reassign indices to temporary registers by reusing unused indices created
3674 * by optimization passes. */
3676 glsl_to_tgsi_visitor::renumber_registers(void)
3681 for (i
=0; i
< this->next_temp
; i
++) {
3682 if (get_first_temp_read(i
) < 0) continue;
3684 rename_temp_register(i
, new_index
);
3688 this->next_temp
= new_index
;
3692 * Returns a fragment program which implements the current pixel transfer ops.
3693 * Based on get_pixel_transfer_program in st_atom_pixeltransfer.c.
3696 get_pixel_transfer_visitor(struct st_fragment_program
*fp
,
3697 glsl_to_tgsi_visitor
*original
,
3698 int scale_and_bias
, int pixel_maps
)
3700 glsl_to_tgsi_visitor
*v
= new glsl_to_tgsi_visitor();
3701 struct st_context
*st
= st_context(original
->ctx
);
3702 struct gl_program
*prog
= &fp
->Base
.Base
;
3703 struct gl_program_parameter_list
*params
= _mesa_new_parameter_list();
3704 st_src_reg coord
, src0
;
3706 glsl_to_tgsi_instruction
*inst
;
3708 /* Copy attributes of the glsl_to_tgsi_visitor in the original shader. */
3709 v
->ctx
= original
->ctx
;
3711 v
->glsl_version
= original
->glsl_version
;
3712 v
->native_integers
= original
->native_integers
;
3713 v
->options
= original
->options
;
3714 v
->next_temp
= original
->next_temp
;
3715 v
->num_address_regs
= original
->num_address_regs
;
3716 v
->samplers_used
= prog
->SamplersUsed
= original
->samplers_used
;
3717 v
->indirect_addr_temps
= original
->indirect_addr_temps
;
3718 v
->indirect_addr_consts
= original
->indirect_addr_consts
;
3719 memcpy(&v
->immediates
, &original
->immediates
, sizeof(v
->immediates
));
3722 * Get initial pixel color from the texture.
3723 * TEX colorTemp, fragment.texcoord[0], texture[0], 2D;
3725 coord
= st_src_reg(PROGRAM_INPUT
, FRAG_ATTRIB_TEX0
, glsl_type::vec2_type
);
3726 src0
= v
->get_temp(glsl_type::vec4_type
);
3727 dst0
= st_dst_reg(src0
);
3728 inst
= v
->emit(NULL
, TGSI_OPCODE_TEX
, dst0
, coord
);
3730 inst
->tex_target
= TEXTURE_2D_INDEX
;
3732 prog
->InputsRead
|= FRAG_BIT_TEX0
;
3733 prog
->SamplersUsed
|= (1 << 0); /* mark sampler 0 as used */
3734 v
->samplers_used
|= (1 << 0);
3736 if (scale_and_bias
) {
3737 static const gl_state_index scale_state
[STATE_LENGTH
] =
3738 { STATE_INTERNAL
, STATE_PT_SCALE
,
3739 (gl_state_index
) 0, (gl_state_index
) 0, (gl_state_index
) 0 };
3740 static const gl_state_index bias_state
[STATE_LENGTH
] =
3741 { STATE_INTERNAL
, STATE_PT_BIAS
,
3742 (gl_state_index
) 0, (gl_state_index
) 0, (gl_state_index
) 0 };
3743 GLint scale_p
, bias_p
;
3744 st_src_reg scale
, bias
;
3746 scale_p
= _mesa_add_state_reference(params
, scale_state
);
3747 bias_p
= _mesa_add_state_reference(params
, bias_state
);
3749 /* MAD colorTemp, colorTemp, scale, bias; */
3750 scale
= st_src_reg(PROGRAM_STATE_VAR
, scale_p
, GLSL_TYPE_FLOAT
);
3751 bias
= st_src_reg(PROGRAM_STATE_VAR
, bias_p
, GLSL_TYPE_FLOAT
);
3752 inst
= v
->emit(NULL
, TGSI_OPCODE_MAD
, dst0
, src0
, scale
, bias
);
3756 st_src_reg temp
= v
->get_temp(glsl_type::vec4_type
);
3757 st_dst_reg temp_dst
= st_dst_reg(temp
);
3759 assert(st
->pixel_xfer
.pixelmap_texture
);
3761 /* With a little effort, we can do four pixel map look-ups with
3762 * two TEX instructions:
3765 /* TEX temp.rg, colorTemp.rgba, texture[1], 2D; */
3766 temp_dst
.writemask
= WRITEMASK_XY
; /* write R,G */
3767 inst
= v
->emit(NULL
, TGSI_OPCODE_TEX
, temp_dst
, src0
);
3769 inst
->tex_target
= TEXTURE_2D_INDEX
;
3771 /* TEX temp.ba, colorTemp.baba, texture[1], 2D; */
3772 src0
.swizzle
= MAKE_SWIZZLE4(SWIZZLE_Z
, SWIZZLE_W
, SWIZZLE_Z
, SWIZZLE_W
);
3773 temp_dst
.writemask
= WRITEMASK_ZW
; /* write B,A */
3774 inst
= v
->emit(NULL
, TGSI_OPCODE_TEX
, temp_dst
, src0
);
3776 inst
->tex_target
= TEXTURE_2D_INDEX
;
3778 prog
->SamplersUsed
|= (1 << 1); /* mark sampler 1 as used */
3779 v
->samplers_used
|= (1 << 1);
3781 /* MOV colorTemp, temp; */
3782 inst
= v
->emit(NULL
, TGSI_OPCODE_MOV
, dst0
, temp
);
3785 /* Now copy the instructions from the original glsl_to_tgsi_visitor into the
3787 foreach_iter(exec_list_iterator
, iter
, original
->instructions
) {
3788 glsl_to_tgsi_instruction
*inst
= (glsl_to_tgsi_instruction
*)iter
.get();
3789 st_src_reg src_regs
[3];
3791 if (inst
->dst
.file
== PROGRAM_OUTPUT
)
3792 prog
->OutputsWritten
|= BITFIELD64_BIT(inst
->dst
.index
);
3794 for (int i
=0; i
<3; i
++) {
3795 src_regs
[i
] = inst
->src
[i
];
3796 if (src_regs
[i
].file
== PROGRAM_INPUT
&&
3797 src_regs
[i
].index
== FRAG_ATTRIB_COL0
)
3799 src_regs
[i
].file
= PROGRAM_TEMPORARY
;
3800 src_regs
[i
].index
= src0
.index
;
3802 else if (src_regs
[i
].file
== PROGRAM_INPUT
)
3803 prog
->InputsRead
|= BITFIELD64_BIT(src_regs
[i
].index
);
3806 v
->emit(NULL
, inst
->op
, inst
->dst
, src_regs
[0], src_regs
[1], src_regs
[2]);
3809 /* Make modifications to fragment program info. */
3810 prog
->Parameters
= _mesa_combine_parameter_lists(params
,
3811 original
->prog
->Parameters
);
3812 _mesa_free_parameter_list(params
);
3813 count_resources(v
, prog
);
3814 fp
->glsl_to_tgsi
= v
;
3818 * Make fragment program for glBitmap:
3819 * Sample the texture and kill the fragment if the bit is 0.
3820 * This program will be combined with the user's fragment program.
3822 * Based on make_bitmap_fragment_program in st_cb_bitmap.c.
3825 get_bitmap_visitor(struct st_fragment_program
*fp
,
3826 glsl_to_tgsi_visitor
*original
, int samplerIndex
)
3828 glsl_to_tgsi_visitor
*v
= new glsl_to_tgsi_visitor();
3829 struct st_context
*st
= st_context(original
->ctx
);
3830 struct gl_program
*prog
= &fp
->Base
.Base
;
3831 st_src_reg coord
, src0
;
3833 glsl_to_tgsi_instruction
*inst
;
3835 /* Copy attributes of the glsl_to_tgsi_visitor in the original shader. */
3836 v
->ctx
= original
->ctx
;
3838 v
->glsl_version
= original
->glsl_version
;
3839 v
->native_integers
= original
->native_integers
;
3840 v
->options
= original
->options
;
3841 v
->next_temp
= original
->next_temp
;
3842 v
->num_address_regs
= original
->num_address_regs
;
3843 v
->samplers_used
= prog
->SamplersUsed
= original
->samplers_used
;
3844 v
->indirect_addr_temps
= original
->indirect_addr_temps
;
3845 v
->indirect_addr_consts
= original
->indirect_addr_consts
;
3846 memcpy(&v
->immediates
, &original
->immediates
, sizeof(v
->immediates
));
3848 /* TEX tmp0, fragment.texcoord[0], texture[0], 2D; */
3849 coord
= st_src_reg(PROGRAM_INPUT
, FRAG_ATTRIB_TEX0
, glsl_type::vec2_type
);
3850 src0
= v
->get_temp(glsl_type::vec4_type
);
3851 dst0
= st_dst_reg(src0
);
3852 inst
= v
->emit(NULL
, TGSI_OPCODE_TEX
, dst0
, coord
);
3853 inst
->sampler
= samplerIndex
;
3854 inst
->tex_target
= TEXTURE_2D_INDEX
;
3856 prog
->InputsRead
|= FRAG_BIT_TEX0
;
3857 prog
->SamplersUsed
|= (1 << samplerIndex
); /* mark sampler as used */
3858 v
->samplers_used
|= (1 << samplerIndex
);
3860 /* KIL if -tmp0 < 0 # texel=0 -> keep / texel=0 -> discard */
3861 src0
.negate
= NEGATE_XYZW
;
3862 if (st
->bitmap
.tex_format
== PIPE_FORMAT_L8_UNORM
)
3863 src0
.swizzle
= SWIZZLE_XXXX
;
3864 inst
= v
->emit(NULL
, TGSI_OPCODE_KIL
, undef_dst
, src0
);
3866 /* Now copy the instructions from the original glsl_to_tgsi_visitor into the
3868 foreach_iter(exec_list_iterator
, iter
, original
->instructions
) {
3869 glsl_to_tgsi_instruction
*inst
= (glsl_to_tgsi_instruction
*)iter
.get();
3870 st_src_reg src_regs
[3];
3872 if (inst
->dst
.file
== PROGRAM_OUTPUT
)
3873 prog
->OutputsWritten
|= BITFIELD64_BIT(inst
->dst
.index
);
3875 for (int i
=0; i
<3; i
++) {
3876 src_regs
[i
] = inst
->src
[i
];
3877 if (src_regs
[i
].file
== PROGRAM_INPUT
)
3878 prog
->InputsRead
|= BITFIELD64_BIT(src_regs
[i
].index
);
3881 v
->emit(NULL
, inst
->op
, inst
->dst
, src_regs
[0], src_regs
[1], src_regs
[2]);
3884 /* Make modifications to fragment program info. */
3885 prog
->Parameters
= _mesa_clone_parameter_list(original
->prog
->Parameters
);
3886 count_resources(v
, prog
);
3887 fp
->glsl_to_tgsi
= v
;
3890 /* ------------------------- TGSI conversion stuff -------------------------- */
3892 unsigned branch_target
;
3897 * Intermediate state used during shader translation.
3899 struct st_translate
{
3900 struct ureg_program
*ureg
;
3902 struct ureg_dst temps
[MAX_TEMPS
];
3903 struct ureg_src
*constants
;
3904 struct ureg_src
*immediates
;
3905 struct ureg_dst outputs
[PIPE_MAX_SHADER_OUTPUTS
];
3906 struct ureg_src inputs
[PIPE_MAX_SHADER_INPUTS
];
3907 struct ureg_dst address
[1];
3908 struct ureg_src samplers
[PIPE_MAX_SAMPLERS
];
3909 struct ureg_src systemValues
[SYSTEM_VALUE_MAX
];
3911 /* Extra info for handling point size clamping in vertex shader */
3912 struct ureg_dst pointSizeResult
; /**< Actual point size output register */
3913 struct ureg_src pointSizeConst
; /**< Point size range constant register */
3914 GLint pointSizeOutIndex
; /**< Temp point size output register */
3915 GLboolean prevInstWrotePointSize
;
3917 const GLuint
*inputMapping
;
3918 const GLuint
*outputMapping
;
3920 /* For every instruction that contains a label (eg CALL), keep
3921 * details so that we can go back afterwards and emit the correct
3922 * tgsi instruction number for each label.
3924 struct label
*labels
;
3925 unsigned labels_size
;
3926 unsigned labels_count
;
3928 /* Keep a record of the tgsi instruction number that each mesa
3929 * instruction starts at, will be used to fix up labels after
3934 unsigned insn_count
;
3936 unsigned procType
; /**< TGSI_PROCESSOR_VERTEX/FRAGMENT */
3941 /** Map Mesa's SYSTEM_VALUE_x to TGSI_SEMANTIC_x */
3942 static unsigned mesa_sysval_to_semantic
[SYSTEM_VALUE_MAX
] = {
3944 TGSI_SEMANTIC_VERTEXID
,
3945 TGSI_SEMANTIC_INSTANCEID
3949 * Make note of a branch to a label in the TGSI code.
3950 * After we've emitted all instructions, we'll go over the list
3951 * of labels built here and patch the TGSI code with the actual
3952 * location of each label.
3954 static unsigned *get_label(struct st_translate
*t
, unsigned branch_target
)
3958 if (t
->labels_count
+ 1 >= t
->labels_size
) {
3959 t
->labels_size
= 1 << (util_logbase2(t
->labels_size
) + 1);
3960 t
->labels
= (struct label
*)realloc(t
->labels
,
3961 t
->labels_size
* sizeof(struct label
));
3962 if (t
->labels
== NULL
) {
3963 static unsigned dummy
;
3969 i
= t
->labels_count
++;
3970 t
->labels
[i
].branch_target
= branch_target
;
3971 return &t
->labels
[i
].token
;
3975 * Called prior to emitting the TGSI code for each instruction.
3976 * Allocate additional space for instructions if needed.
3977 * Update the insn[] array so the next glsl_to_tgsi_instruction points to
3978 * the next TGSI instruction.
3980 static void set_insn_start(struct st_translate
*t
, unsigned start
)
3982 if (t
->insn_count
+ 1 >= t
->insn_size
) {
3983 t
->insn_size
= 1 << (util_logbase2(t
->insn_size
) + 1);
3984 t
->insn
= (unsigned *)realloc(t
->insn
, t
->insn_size
* sizeof(t
->insn
[0]));
3985 if (t
->insn
== NULL
) {
3991 t
->insn
[t
->insn_count
++] = start
;
3995 * Map a glsl_to_tgsi constant/immediate to a TGSI immediate.
3997 static struct ureg_src
3998 emit_immediate(struct st_translate
*t
,
3999 gl_constant_value values
[4],
4002 struct ureg_program
*ureg
= t
->ureg
;
4007 return ureg_DECL_immediate(ureg
, &values
[0].f
, size
);
4009 return ureg_DECL_immediate_int(ureg
, &values
[0].i
, size
);
4010 case GL_UNSIGNED_INT
:
4012 return ureg_DECL_immediate_uint(ureg
, &values
[0].u
, size
);
4014 assert(!"should not get here - type must be float, int, uint, or bool");
4015 return ureg_src_undef();
4020 * Map a glsl_to_tgsi dst register to a TGSI ureg_dst register.
4022 static struct ureg_dst
4023 dst_register(struct st_translate
*t
,
4024 gl_register_file file
,
4028 case PROGRAM_UNDEFINED
:
4029 return ureg_dst_undef();
4031 case PROGRAM_TEMPORARY
:
4032 if (ureg_dst_is_undef(t
->temps
[index
]))
4033 t
->temps
[index
] = ureg_DECL_temporary(t
->ureg
);
4035 return t
->temps
[index
];
4037 case PROGRAM_OUTPUT
:
4038 if (t
->procType
== TGSI_PROCESSOR_VERTEX
&& index
== VERT_RESULT_PSIZ
)
4039 t
->prevInstWrotePointSize
= GL_TRUE
;
4041 if (t
->procType
== TGSI_PROCESSOR_VERTEX
)
4042 assert(index
< VERT_RESULT_MAX
);
4043 else if (t
->procType
== TGSI_PROCESSOR_FRAGMENT
)
4044 assert(index
< FRAG_RESULT_MAX
);
4046 assert(index
< GEOM_RESULT_MAX
);
4048 assert(t
->outputMapping
[index
] < Elements(t
->outputs
));
4050 return t
->outputs
[t
->outputMapping
[index
]];
4052 case PROGRAM_ADDRESS
:
4053 return t
->address
[index
];
4056 assert(!"unknown dst register file");
4057 return ureg_dst_undef();
4062 * Map a glsl_to_tgsi src register to a TGSI ureg_src register.
4064 static struct ureg_src
4065 src_register(struct st_translate
*t
,
4066 gl_register_file file
,
4070 case PROGRAM_UNDEFINED
:
4071 return ureg_src_undef();
4073 case PROGRAM_TEMPORARY
:
4075 assert(index
< Elements(t
->temps
));
4076 if (ureg_dst_is_undef(t
->temps
[index
]))
4077 t
->temps
[index
] = ureg_DECL_temporary(t
->ureg
);
4078 return ureg_src(t
->temps
[index
]);
4080 case PROGRAM_NAMED_PARAM
:
4081 case PROGRAM_ENV_PARAM
:
4082 case PROGRAM_LOCAL_PARAM
:
4083 case PROGRAM_UNIFORM
:
4085 return t
->constants
[index
];
4086 case PROGRAM_STATE_VAR
:
4087 case PROGRAM_CONSTANT
: /* ie, immediate */
4089 return ureg_DECL_constant(t
->ureg
, 0);
4091 return t
->constants
[index
];
4093 case PROGRAM_IMMEDIATE
:
4094 return t
->immediates
[index
];
4097 assert(t
->inputMapping
[index
] < Elements(t
->inputs
));
4098 return t
->inputs
[t
->inputMapping
[index
]];
4100 case PROGRAM_OUTPUT
:
4101 assert(t
->outputMapping
[index
] < Elements(t
->outputs
));
4102 return ureg_src(t
->outputs
[t
->outputMapping
[index
]]); /* not needed? */
4104 case PROGRAM_ADDRESS
:
4105 return ureg_src(t
->address
[index
]);
4107 case PROGRAM_SYSTEM_VALUE
:
4108 assert(index
< Elements(t
->systemValues
));
4109 return t
->systemValues
[index
];
4112 assert(!"unknown src register file");
4113 return ureg_src_undef();
4118 * Create a TGSI ureg_dst register from an st_dst_reg.
4120 static struct ureg_dst
4121 translate_dst(struct st_translate
*t
,
4122 const st_dst_reg
*dst_reg
,
4125 struct ureg_dst dst
= dst_register(t
,
4129 dst
= ureg_writemask(dst
, dst_reg
->writemask
);
4132 dst
= ureg_saturate(dst
);
4134 if (dst_reg
->reladdr
!= NULL
)
4135 dst
= ureg_dst_indirect(dst
, ureg_src(t
->address
[0]));
4141 * Create a TGSI ureg_src register from an st_src_reg.
4143 static struct ureg_src
4144 translate_src(struct st_translate
*t
, const st_src_reg
*src_reg
)
4146 struct ureg_src src
= src_register(t
, src_reg
->file
, src_reg
->index
);
4148 src
= ureg_swizzle(src
,
4149 GET_SWZ(src_reg
->swizzle
, 0) & 0x3,
4150 GET_SWZ(src_reg
->swizzle
, 1) & 0x3,
4151 GET_SWZ(src_reg
->swizzle
, 2) & 0x3,
4152 GET_SWZ(src_reg
->swizzle
, 3) & 0x3);
4154 if ((src_reg
->negate
& 0xf) == NEGATE_XYZW
)
4155 src
= ureg_negate(src
);
4157 if (src_reg
->reladdr
!= NULL
) {
4158 /* Normally ureg_src_indirect() would be used here, but a stupid compiler
4159 * bug in g++ makes ureg_src_indirect (an inline C function) erroneously
4160 * set the bit for src.Negate. So we have to do the operation manually
4161 * here to work around the compiler's problems. */
4162 /*src = ureg_src_indirect(src, ureg_src(t->address[0]));*/
4163 struct ureg_src addr
= ureg_src(t
->address
[0]);
4165 src
.IndirectFile
= addr
.File
;
4166 src
.IndirectIndex
= addr
.Index
;
4167 src
.IndirectSwizzle
= addr
.SwizzleX
;
4169 if (src_reg
->file
!= PROGRAM_INPUT
&&
4170 src_reg
->file
!= PROGRAM_OUTPUT
) {
4171 /* If src_reg->index was negative, it was set to zero in
4172 * src_register(). Reassign it now. But don't do this
4173 * for input/output regs since they get remapped while
4174 * const buffers don't.
4176 src
.Index
= src_reg
->index
;
4183 static struct tgsi_texture_offset
4184 translate_tex_offset(struct st_translate
*t
,
4185 const struct tgsi_texture_offset
*in_offset
)
4187 struct tgsi_texture_offset offset
;
4189 assert(in_offset
->File
== PROGRAM_IMMEDIATE
);
4191 offset
.File
= TGSI_FILE_IMMEDIATE
;
4192 offset
.Index
= in_offset
->Index
;
4193 offset
.SwizzleX
= in_offset
->SwizzleX
;
4194 offset
.SwizzleY
= in_offset
->SwizzleY
;
4195 offset
.SwizzleZ
= in_offset
->SwizzleZ
;
4201 compile_tgsi_instruction(struct st_translate
*t
,
4202 const glsl_to_tgsi_instruction
*inst
)
4204 struct ureg_program
*ureg
= t
->ureg
;
4206 struct ureg_dst dst
[1];
4207 struct ureg_src src
[4];
4208 struct tgsi_texture_offset texoffsets
[MAX_GLSL_TEXTURE_OFFSET
];
4213 num_dst
= num_inst_dst_regs(inst
->op
);
4214 num_src
= num_inst_src_regs(inst
->op
);
4217 dst
[0] = translate_dst(t
,
4221 for (i
= 0; i
< num_src
; i
++)
4222 src
[i
] = translate_src(t
, &inst
->src
[i
]);
4225 case TGSI_OPCODE_BGNLOOP
:
4226 case TGSI_OPCODE_CAL
:
4227 case TGSI_OPCODE_ELSE
:
4228 case TGSI_OPCODE_ENDLOOP
:
4229 case TGSI_OPCODE_IF
:
4230 assert(num_dst
== 0);
4231 ureg_label_insn(ureg
,
4235 inst
->op
== TGSI_OPCODE_CAL
? inst
->function
->sig_id
: 0));
4238 case TGSI_OPCODE_TEX
:
4239 case TGSI_OPCODE_TXB
:
4240 case TGSI_OPCODE_TXD
:
4241 case TGSI_OPCODE_TXL
:
4242 case TGSI_OPCODE_TXP
:
4243 case TGSI_OPCODE_TXQ
:
4244 case TGSI_OPCODE_TXF
:
4245 src
[num_src
++] = t
->samplers
[inst
->sampler
];
4246 for (i
= 0; i
< inst
->tex_offset_num_offset
; i
++) {
4247 texoffsets
[i
] = translate_tex_offset(t
, &inst
->tex_offsets
[i
]);
4252 translate_texture_target(inst
->tex_target
, inst
->tex_shadow
),
4253 texoffsets
, inst
->tex_offset_num_offset
,
4257 case TGSI_OPCODE_SCS
:
4258 dst
[0] = ureg_writemask(dst
[0], TGSI_WRITEMASK_XY
);
4259 ureg_insn(ureg
, inst
->op
, dst
, num_dst
, src
, num_src
);
4272 * Emit the TGSI instructions for inverting and adjusting WPOS.
4273 * This code is unavoidable because it also depends on whether
4274 * a FBO is bound (STATE_FB_WPOS_Y_TRANSFORM).
4277 emit_wpos_adjustment( struct st_translate
*t
,
4278 const struct gl_program
*program
,
4280 GLfloat adjX
, GLfloat adjY
[2])
4282 struct ureg_program
*ureg
= t
->ureg
;
4284 /* Fragment program uses fragment position input.
4285 * Need to replace instances of INPUT[WPOS] with temp T
4286 * where T = INPUT[WPOS] by y is inverted.
4288 static const gl_state_index wposTransformState
[STATE_LENGTH
]
4289 = { STATE_INTERNAL
, STATE_FB_WPOS_Y_TRANSFORM
,
4290 (gl_state_index
)0, (gl_state_index
)0, (gl_state_index
)0 };
4292 /* XXX: note we are modifying the incoming shader here! Need to
4293 * do this before emitting the constant decls below, or this
4296 unsigned wposTransConst
= _mesa_add_state_reference(program
->Parameters
,
4297 wposTransformState
);
4299 struct ureg_src wpostrans
= ureg_DECL_constant( ureg
, wposTransConst
);
4300 struct ureg_dst wpos_temp
= ureg_DECL_temporary( ureg
);
4301 struct ureg_src wpos_input
= t
->inputs
[t
->inputMapping
[FRAG_ATTRIB_WPOS
]];
4303 /* First, apply the coordinate shift: */
4304 if (adjX
|| adjY
[0] || adjY
[1]) {
4305 if (adjY
[0] != adjY
[1]) {
4306 /* Adjust the y coordinate by adjY[1] or adjY[0] respectively
4307 * depending on whether inversion is actually going to be applied
4308 * or not, which is determined by testing against the inversion
4309 * state variable used below, which will be either +1 or -1.
4311 struct ureg_dst adj_temp
= ureg_DECL_temporary(ureg
);
4313 ureg_CMP(ureg
, adj_temp
,
4314 ureg_scalar(wpostrans
, invert
? 2 : 0),
4315 ureg_imm4f(ureg
, adjX
, adjY
[0], 0.0f
, 0.0f
),
4316 ureg_imm4f(ureg
, adjX
, adjY
[1], 0.0f
, 0.0f
));
4317 ureg_ADD(ureg
, wpos_temp
, wpos_input
, ureg_src(adj_temp
));
4319 ureg_ADD(ureg
, wpos_temp
, wpos_input
,
4320 ureg_imm4f(ureg
, adjX
, adjY
[0], 0.0f
, 0.0f
));
4322 wpos_input
= ureg_src(wpos_temp
);
4324 /* MOV wpos_temp, input[wpos]
4326 ureg_MOV( ureg
, wpos_temp
, wpos_input
);
4329 /* Now the conditional y flip: STATE_FB_WPOS_Y_TRANSFORM.xy/zw will be
4330 * inversion/identity, or the other way around if we're drawing to an FBO.
4333 /* MAD wpos_temp.y, wpos_input, wpostrans.xxxx, wpostrans.yyyy
4336 ureg_writemask(wpos_temp
, TGSI_WRITEMASK_Y
),
4338 ureg_scalar(wpostrans
, 0),
4339 ureg_scalar(wpostrans
, 1));
4341 /* MAD wpos_temp.y, wpos_input, wpostrans.zzzz, wpostrans.wwww
4344 ureg_writemask(wpos_temp
, TGSI_WRITEMASK_Y
),
4346 ureg_scalar(wpostrans
, 2),
4347 ureg_scalar(wpostrans
, 3));
4350 /* Use wpos_temp as position input from here on:
4352 t
->inputs
[t
->inputMapping
[FRAG_ATTRIB_WPOS
]] = ureg_src(wpos_temp
);
4357 * Emit fragment position/ooordinate code.
4360 emit_wpos(struct st_context
*st
,
4361 struct st_translate
*t
,
4362 const struct gl_program
*program
,
4363 struct ureg_program
*ureg
)
4365 const struct gl_fragment_program
*fp
=
4366 (const struct gl_fragment_program
*) program
;
4367 struct pipe_screen
*pscreen
= st
->pipe
->screen
;
4368 GLfloat adjX
= 0.0f
;
4369 GLfloat adjY
[2] = { 0.0f
, 0.0f
};
4370 boolean invert
= FALSE
;
4372 /* Query the pixel center conventions supported by the pipe driver and set
4373 * adjX, adjY to help out if it cannot handle the requested one internally.
4375 * The bias of the y-coordinate depends on whether y-inversion takes place
4376 * (adjY[1]) or not (adjY[0]), which is in turn dependent on whether we are
4377 * drawing to an FBO (causes additional inversion), and whether the the pipe
4378 * driver origin and the requested origin differ (the latter condition is
4379 * stored in the 'invert' variable).
4381 * For height = 100 (i = integer, h = half-integer, l = lower, u = upper):
4383 * center shift only:
4388 * l,i -> u,i: ( 0.0 + 1.0) * -1 + 100 = 99
4389 * l,h -> u,h: ( 0.5 + 0.0) * -1 + 100 = 99.5
4390 * u,i -> l,i: (99.0 + 1.0) * -1 + 100 = 0
4391 * u,h -> l,h: (99.5 + 0.0) * -1 + 100 = 0.5
4393 * inversion and center shift:
4394 * l,i -> u,h: ( 0.0 + 0.5) * -1 + 100 = 99.5
4395 * l,h -> u,i: ( 0.5 + 0.5) * -1 + 100 = 99
4396 * u,i -> l,h: (99.0 + 0.5) * -1 + 100 = 0.5
4397 * u,h -> l,i: (99.5 + 0.5) * -1 + 100 = 0
4399 if (fp
->OriginUpperLeft
) {
4400 /* Fragment shader wants origin in upper-left */
4401 if (pscreen
->get_param(pscreen
, PIPE_CAP_TGSI_FS_COORD_ORIGIN_UPPER_LEFT
)) {
4402 /* the driver supports upper-left origin */
4404 else if (pscreen
->get_param(pscreen
, PIPE_CAP_TGSI_FS_COORD_ORIGIN_LOWER_LEFT
)) {
4405 /* the driver supports lower-left origin, need to invert Y */
4406 ureg_property_fs_coord_origin(ureg
, TGSI_FS_COORD_ORIGIN_LOWER_LEFT
);
4413 /* Fragment shader wants origin in lower-left */
4414 if (pscreen
->get_param(pscreen
, PIPE_CAP_TGSI_FS_COORD_ORIGIN_LOWER_LEFT
))
4415 /* the driver supports lower-left origin */
4416 ureg_property_fs_coord_origin(ureg
, TGSI_FS_COORD_ORIGIN_LOWER_LEFT
);
4417 else if (pscreen
->get_param(pscreen
, PIPE_CAP_TGSI_FS_COORD_ORIGIN_UPPER_LEFT
))
4418 /* the driver supports upper-left origin, need to invert Y */
4424 if (fp
->PixelCenterInteger
) {
4425 /* Fragment shader wants pixel center integer */
4426 if (pscreen
->get_param(pscreen
, PIPE_CAP_TGSI_FS_COORD_PIXEL_CENTER_INTEGER
)) {
4427 /* the driver supports pixel center integer */
4429 ureg_property_fs_coord_pixel_center(ureg
, TGSI_FS_COORD_PIXEL_CENTER_INTEGER
);
4431 else if (pscreen
->get_param(pscreen
, PIPE_CAP_TGSI_FS_COORD_PIXEL_CENTER_HALF_INTEGER
)) {
4432 /* the driver supports pixel center half integer, need to bias X,Y */
4441 /* Fragment shader wants pixel center half integer */
4442 if (pscreen
->get_param(pscreen
, PIPE_CAP_TGSI_FS_COORD_PIXEL_CENTER_HALF_INTEGER
)) {
4443 /* the driver supports pixel center half integer */
4445 else if (pscreen
->get_param(pscreen
, PIPE_CAP_TGSI_FS_COORD_PIXEL_CENTER_INTEGER
)) {
4446 /* the driver supports pixel center integer, need to bias X,Y */
4447 adjX
= adjY
[0] = adjY
[1] = 0.5f
;
4448 ureg_property_fs_coord_pixel_center(ureg
, TGSI_FS_COORD_PIXEL_CENTER_INTEGER
);
4454 /* we invert after adjustment so that we avoid the MOV to temporary,
4455 * and reuse the adjustment ADD instead */
4456 emit_wpos_adjustment(t
, program
, invert
, adjX
, adjY
);
4460 * OpenGL's fragment gl_FrontFace input is 1 for front-facing, 0 for back.
4461 * TGSI uses +1 for front, -1 for back.
4462 * This function converts the TGSI value to the GL value. Simply clamping/
4463 * saturating the value to [0,1] does the job.
4466 emit_face_var(struct st_translate
*t
)
4468 struct ureg_program
*ureg
= t
->ureg
;
4469 struct ureg_dst face_temp
= ureg_DECL_temporary(ureg
);
4470 struct ureg_src face_input
= t
->inputs
[t
->inputMapping
[FRAG_ATTRIB_FACE
]];
4472 /* MOV_SAT face_temp, input[face] */
4473 face_temp
= ureg_saturate(face_temp
);
4474 ureg_MOV(ureg
, face_temp
, face_input
);
4476 /* Use face_temp as face input from here on: */
4477 t
->inputs
[t
->inputMapping
[FRAG_ATTRIB_FACE
]] = ureg_src(face_temp
);
4481 emit_edgeflags(struct st_translate
*t
)
4483 struct ureg_program
*ureg
= t
->ureg
;
4484 struct ureg_dst edge_dst
= t
->outputs
[t
->outputMapping
[VERT_RESULT_EDGE
]];
4485 struct ureg_src edge_src
= t
->inputs
[t
->inputMapping
[VERT_ATTRIB_EDGEFLAG
]];
4487 ureg_MOV(ureg
, edge_dst
, edge_src
);
4491 * Translate intermediate IR (glsl_to_tgsi_instruction) to TGSI format.
4492 * \param program the program to translate
4493 * \param numInputs number of input registers used
4494 * \param inputMapping maps Mesa fragment program inputs to TGSI generic
4496 * \param inputSemanticName the TGSI_SEMANTIC flag for each input
4497 * \param inputSemanticIndex the semantic index (ex: which texcoord) for
4499 * \param interpMode the TGSI_INTERPOLATE_LINEAR/PERSP mode for each input
4500 * \param numOutputs number of output registers used
4501 * \param outputMapping maps Mesa fragment program outputs to TGSI
4503 * \param outputSemanticName the TGSI_SEMANTIC flag for each output
4504 * \param outputSemanticIndex the semantic index (ex: which texcoord) for
4507 * \return PIPE_OK or PIPE_ERROR_OUT_OF_MEMORY
4509 extern "C" enum pipe_error
4510 st_translate_program(
4511 struct gl_context
*ctx
,
4513 struct ureg_program
*ureg
,
4514 glsl_to_tgsi_visitor
*program
,
4515 const struct gl_program
*proginfo
,
4517 const GLuint inputMapping
[],
4518 const ubyte inputSemanticName
[],
4519 const ubyte inputSemanticIndex
[],
4520 const GLuint interpMode
[],
4522 const GLuint outputMapping
[],
4523 const ubyte outputSemanticName
[],
4524 const ubyte outputSemanticIndex
[],
4525 boolean passthrough_edgeflags
)
4527 struct st_translate
*t
;
4529 enum pipe_error ret
= PIPE_OK
;
4531 assert(numInputs
<= Elements(t
->inputs
));
4532 assert(numOutputs
<= Elements(t
->outputs
));
4534 t
= CALLOC_STRUCT(st_translate
);
4536 ret
= PIPE_ERROR_OUT_OF_MEMORY
;
4540 memset(t
, 0, sizeof *t
);
4542 t
->procType
= procType
;
4543 t
->inputMapping
= inputMapping
;
4544 t
->outputMapping
= outputMapping
;
4546 t
->pointSizeOutIndex
= -1;
4547 t
->prevInstWrotePointSize
= GL_FALSE
;
4550 * Declare input attributes.
4552 if (procType
== TGSI_PROCESSOR_FRAGMENT
) {
4553 for (i
= 0; i
< numInputs
; i
++) {
4554 t
->inputs
[i
] = ureg_DECL_fs_input(ureg
,
4555 inputSemanticName
[i
],
4556 inputSemanticIndex
[i
],
4560 if (program
->shader_program
->FragDepthLayout
!= FRAG_DEPTH_LAYOUT_NONE
) {
4561 switch (program
->shader_program
->FragDepthLayout
) {
4562 case FRAG_DEPTH_LAYOUT_ANY
:
4563 ureg_property_fs_depth_layout(ureg
, TGSI_FS_DEPTH_LAYOUT_ANY
);
4565 case FRAG_DEPTH_LAYOUT_GREATER
:
4566 ureg_property_fs_depth_layout(ureg
, TGSI_FS_DEPTH_LAYOUT_GREATER
);
4568 case FRAG_DEPTH_LAYOUT_LESS
:
4569 ureg_property_fs_depth_layout(ureg
, TGSI_FS_DEPTH_LAYOUT_LESS
);
4571 case FRAG_DEPTH_LAYOUT_UNCHANGED
:
4572 ureg_property_fs_depth_layout(ureg
, TGSI_FS_DEPTH_LAYOUT_UNCHANGED
);
4579 if (proginfo
->InputsRead
& FRAG_BIT_WPOS
) {
4580 /* Must do this after setting up t->inputs, and before
4581 * emitting constant references, below:
4583 emit_wpos(st_context(ctx
), t
, proginfo
, ureg
);
4586 if (proginfo
->InputsRead
& FRAG_BIT_FACE
)
4590 * Declare output attributes.
4592 for (i
= 0; i
< numOutputs
; i
++) {
4593 switch (outputSemanticName
[i
]) {
4594 case TGSI_SEMANTIC_POSITION
:
4595 t
->outputs
[i
] = ureg_DECL_output(ureg
,
4596 TGSI_SEMANTIC_POSITION
, /* Z/Depth */
4597 outputSemanticIndex
[i
]);
4598 t
->outputs
[i
] = ureg_writemask(t
->outputs
[i
], TGSI_WRITEMASK_Z
);
4600 case TGSI_SEMANTIC_STENCIL
:
4601 t
->outputs
[i
] = ureg_DECL_output(ureg
,
4602 TGSI_SEMANTIC_STENCIL
, /* Stencil */
4603 outputSemanticIndex
[i
]);
4604 t
->outputs
[i
] = ureg_writemask(t
->outputs
[i
], TGSI_WRITEMASK_Y
);
4606 case TGSI_SEMANTIC_COLOR
:
4607 t
->outputs
[i
] = ureg_DECL_output(ureg
,
4608 TGSI_SEMANTIC_COLOR
,
4609 outputSemanticIndex
[i
]);
4612 assert(!"fragment shader outputs must be POSITION/STENCIL/COLOR");
4613 ret
= PIPE_ERROR_BAD_INPUT
;
4618 else if (procType
== TGSI_PROCESSOR_GEOMETRY
) {
4619 for (i
= 0; i
< numInputs
; i
++) {
4620 t
->inputs
[i
] = ureg_DECL_gs_input(ureg
,
4622 inputSemanticName
[i
],
4623 inputSemanticIndex
[i
]);
4626 for (i
= 0; i
< numOutputs
; i
++) {
4627 t
->outputs
[i
] = ureg_DECL_output(ureg
,
4628 outputSemanticName
[i
],
4629 outputSemanticIndex
[i
]);
4633 assert(procType
== TGSI_PROCESSOR_VERTEX
);
4635 for (i
= 0; i
< numInputs
; i
++) {
4636 t
->inputs
[i
] = ureg_DECL_vs_input(ureg
, i
);
4639 for (i
= 0; i
< numOutputs
; i
++) {
4640 t
->outputs
[i
] = ureg_DECL_output(ureg
,
4641 outputSemanticName
[i
],
4642 outputSemanticIndex
[i
]);
4643 if ((outputSemanticName
[i
] == TGSI_SEMANTIC_PSIZE
) && proginfo
->Id
) {
4644 /* Writing to the point size result register requires special
4645 * handling to implement clamping.
4647 static const gl_state_index pointSizeClampState
[STATE_LENGTH
]
4648 = { STATE_INTERNAL
, STATE_POINT_SIZE_IMPL_CLAMP
, (gl_state_index
)0, (gl_state_index
)0, (gl_state_index
)0 };
4649 /* XXX: note we are modifying the incoming shader here! Need to
4650 * do this before emitting the constant decls below, or this
4653 unsigned pointSizeClampConst
=
4654 _mesa_add_state_reference(proginfo
->Parameters
,
4655 pointSizeClampState
);
4656 struct ureg_dst psizregtemp
= ureg_DECL_temporary(ureg
);
4657 t
->pointSizeConst
= ureg_DECL_constant(ureg
, pointSizeClampConst
);
4658 t
->pointSizeResult
= t
->outputs
[i
];
4659 t
->pointSizeOutIndex
= i
;
4660 t
->outputs
[i
] = psizregtemp
;
4663 if (passthrough_edgeflags
)
4667 /* Declare address register.
4669 if (program
->num_address_regs
> 0) {
4670 assert(program
->num_address_regs
== 1);
4671 t
->address
[0] = ureg_DECL_address(ureg
);
4674 /* Declare misc input registers
4677 GLbitfield sysInputs
= proginfo
->SystemValuesRead
;
4678 unsigned numSys
= 0;
4679 for (i
= 0; sysInputs
; i
++) {
4680 if (sysInputs
& (1 << i
)) {
4681 unsigned semName
= mesa_sysval_to_semantic
[i
];
4682 t
->systemValues
[i
] = ureg_DECL_system_value(ureg
, numSys
, semName
, 0);
4684 sysInputs
&= ~(1 << i
);
4689 if (program
->indirect_addr_temps
) {
4690 /* If temps are accessed with indirect addressing, declare temporaries
4691 * in sequential order. Else, we declare them on demand elsewhere.
4692 * (Note: the number of temporaries is equal to program->next_temp)
4694 for (i
= 0; i
< (unsigned)program
->next_temp
; i
++) {
4695 /* XXX use TGSI_FILE_TEMPORARY_ARRAY when it's supported by ureg */
4696 t
->temps
[i
] = ureg_DECL_temporary(t
->ureg
);
4700 /* Emit constants and uniforms. TGSI uses a single index space for these,
4701 * so we put all the translated regs in t->constants.
4703 if (proginfo
->Parameters
) {
4704 t
->constants
= (struct ureg_src
*)CALLOC(proginfo
->Parameters
->NumParameters
* sizeof(t
->constants
[0]));
4705 if (t
->constants
== NULL
) {
4706 ret
= PIPE_ERROR_OUT_OF_MEMORY
;
4710 for (i
= 0; i
< proginfo
->Parameters
->NumParameters
; i
++) {
4711 switch (proginfo
->Parameters
->Parameters
[i
].Type
) {
4712 case PROGRAM_ENV_PARAM
:
4713 case PROGRAM_LOCAL_PARAM
:
4714 case PROGRAM_STATE_VAR
:
4715 case PROGRAM_NAMED_PARAM
:
4716 case PROGRAM_UNIFORM
:
4717 t
->constants
[i
] = ureg_DECL_constant(ureg
, i
);
4720 /* Emit immediates for PROGRAM_CONSTANT only when there's no indirect
4721 * addressing of the const buffer.
4722 * FIXME: Be smarter and recognize param arrays:
4723 * indirect addressing is only valid within the referenced
4726 case PROGRAM_CONSTANT
:
4727 if (program
->indirect_addr_consts
)
4728 t
->constants
[i
] = ureg_DECL_constant(ureg
, i
);
4730 t
->constants
[i
] = emit_immediate(t
,
4731 proginfo
->Parameters
->ParameterValues
[i
],
4732 proginfo
->Parameters
->Parameters
[i
].DataType
,
4741 /* Emit immediate values.
4743 t
->immediates
= (struct ureg_src
*)CALLOC(program
->num_immediates
* sizeof(struct ureg_src
));
4744 if (t
->immediates
== NULL
) {
4745 ret
= PIPE_ERROR_OUT_OF_MEMORY
;
4749 foreach_iter(exec_list_iterator
, iter
, program
->immediates
) {
4750 immediate_storage
*imm
= (immediate_storage
*)iter
.get();
4751 t
->immediates
[i
++] = emit_immediate(t
, imm
->values
, imm
->type
, imm
->size
);
4754 /* texture samplers */
4755 for (i
= 0; i
< ctx
->Const
.MaxTextureImageUnits
; i
++) {
4756 if (program
->samplers_used
& (1 << i
)) {
4757 t
->samplers
[i
] = ureg_DECL_sampler(ureg
, i
);
4761 /* Emit each instruction in turn:
4763 foreach_iter(exec_list_iterator
, iter
, program
->instructions
) {
4764 set_insn_start(t
, ureg_get_instruction_number(ureg
));
4765 compile_tgsi_instruction(t
, (glsl_to_tgsi_instruction
*)iter
.get());
4767 if (t
->prevInstWrotePointSize
&& proginfo
->Id
) {
4768 /* The previous instruction wrote to the (fake) vertex point size
4769 * result register. Now we need to clamp that value to the min/max
4770 * point size range, putting the result into the real point size
4772 * Note that we can't do this easily at the end of program due to
4773 * possible early return.
4775 set_insn_start(t
, ureg_get_instruction_number(ureg
));
4777 ureg_writemask(t
->outputs
[t
->pointSizeOutIndex
], WRITEMASK_X
),
4778 ureg_src(t
->outputs
[t
->pointSizeOutIndex
]),
4779 ureg_swizzle(t
->pointSizeConst
, 1,1,1,1));
4780 ureg_MIN(t
->ureg
, ureg_writemask(t
->pointSizeResult
, WRITEMASK_X
),
4781 ureg_src(t
->outputs
[t
->pointSizeOutIndex
]),
4782 ureg_swizzle(t
->pointSizeConst
, 2,2,2,2));
4784 t
->prevInstWrotePointSize
= GL_FALSE
;
4787 /* Fix up all emitted labels:
4789 for (i
= 0; i
< t
->labels_count
; i
++) {
4790 ureg_fixup_label(ureg
, t
->labels
[i
].token
,
4791 t
->insn
[t
->labels
[i
].branch_target
]);
4799 FREE(t
->immediates
);
4802 debug_printf("%s: translate error flag set\n", __FUNCTION__
);
4810 /* ----------------------------- End TGSI code ------------------------------ */
4813 * Convert a shader's GLSL IR into a Mesa gl_program, although without
4814 * generating Mesa IR.
4816 static struct gl_program
*
4817 get_mesa_program(struct gl_context
*ctx
,
4818 struct gl_shader_program
*shader_program
,
4819 struct gl_shader
*shader
)
4821 glsl_to_tgsi_visitor
* v
= new glsl_to_tgsi_visitor();
4822 struct gl_program
*prog
;
4823 struct pipe_screen
* screen
= st_context(ctx
)->pipe
->screen
;
4824 unsigned pipe_shader_type
;
4826 const char *target_string
;
4828 struct gl_shader_compiler_options
*options
=
4829 &ctx
->ShaderCompilerOptions
[_mesa_shader_type_to_index(shader
->Type
)];
4831 switch (shader
->Type
) {
4832 case GL_VERTEX_SHADER
:
4833 target
= GL_VERTEX_PROGRAM_ARB
;
4834 target_string
= "vertex";
4835 pipe_shader_type
= PIPE_SHADER_VERTEX
;
4837 case GL_FRAGMENT_SHADER
:
4838 target
= GL_FRAGMENT_PROGRAM_ARB
;
4839 target_string
= "fragment";
4840 pipe_shader_type
= PIPE_SHADER_FRAGMENT
;
4842 case GL_GEOMETRY_SHADER
:
4843 target
= GL_GEOMETRY_PROGRAM_NV
;
4844 target_string
= "geometry";
4845 pipe_shader_type
= PIPE_SHADER_GEOMETRY
;
4848 assert(!"should not be reached");
4852 validate_ir_tree(shader
->ir
);
4854 prog
= ctx
->Driver
.NewProgram(ctx
, target
, shader_program
->Name
);
4857 prog
->Parameters
= _mesa_new_parameter_list();
4860 v
->shader_program
= shader_program
;
4861 v
->options
= options
;
4862 v
->glsl_version
= ctx
->Const
.GLSLVersion
;
4863 v
->native_integers
= ctx
->Const
.NativeIntegers
;
4865 _mesa_generate_parameters_list_for_uniforms(shader_program
, shader
,
4868 /* Emit intermediate IR for main(). */
4869 visit_exec_list(shader
->ir
, v
);
4871 /* Now emit bodies for any functions that were used. */
4873 progress
= GL_FALSE
;
4875 foreach_iter(exec_list_iterator
, iter
, v
->function_signatures
) {
4876 function_entry
*entry
= (function_entry
*)iter
.get();
4878 if (!entry
->bgn_inst
) {
4879 v
->current_function
= entry
;
4881 entry
->bgn_inst
= v
->emit(NULL
, TGSI_OPCODE_BGNSUB
);
4882 entry
->bgn_inst
->function
= entry
;
4884 visit_exec_list(&entry
->sig
->body
, v
);
4886 glsl_to_tgsi_instruction
*last
;
4887 last
= (glsl_to_tgsi_instruction
*)v
->instructions
.get_tail();
4888 if (last
->op
!= TGSI_OPCODE_RET
)
4889 v
->emit(NULL
, TGSI_OPCODE_RET
);
4891 glsl_to_tgsi_instruction
*end
;
4892 end
= v
->emit(NULL
, TGSI_OPCODE_ENDSUB
);
4893 end
->function
= entry
;
4901 /* Print out some information (for debugging purposes) used by the
4902 * optimization passes. */
4903 for (i
=0; i
< v
->next_temp
; i
++) {
4904 int fr
= v
->get_first_temp_read(i
);
4905 int fw
= v
->get_first_temp_write(i
);
4906 int lr
= v
->get_last_temp_read(i
);
4907 int lw
= v
->get_last_temp_write(i
);
4909 printf("Temp %d: FR=%3d FW=%3d LR=%3d LW=%3d\n", i
, fr
, fw
, lr
, lw
);
4914 if (!screen
->get_shader_param(screen
, pipe_shader_type
,
4915 PIPE_SHADER_CAP_OUTPUT_READ
)) {
4916 /* Remove reads to output registers, and to varyings in vertex shaders. */
4917 v
->remove_output_reads(PROGRAM_OUTPUT
);
4918 if (target
== GL_VERTEX_PROGRAM_ARB
)
4919 v
->remove_output_reads(PROGRAM_VARYING
);
4922 /* Perform optimizations on the instructions in the glsl_to_tgsi_visitor. */
4924 v
->copy_propagate();
4925 while (v
->eliminate_dead_code_advanced());
4927 /* FIXME: These passes to optimize temporary registers don't work when there
4928 * is indirect addressing of the temporary register space. We need proper
4929 * array support so that we don't have to give up these passes in every
4930 * shader that uses arrays.
4932 if (!v
->indirect_addr_temps
) {
4933 v
->eliminate_dead_code();
4934 v
->merge_registers();
4935 v
->renumber_registers();
4938 /* Write the END instruction. */
4939 v
->emit(NULL
, TGSI_OPCODE_END
);
4941 if (ctx
->Shader
.Flags
& GLSL_DUMP
) {
4943 printf("GLSL IR for linked %s program %d:\n", target_string
,
4944 shader_program
->Name
);
4945 _mesa_print_ir(shader
->ir
, NULL
);
4951 prog
->Instructions
= NULL
;
4952 prog
->NumInstructions
= 0;
4954 do_set_program_inouts(shader
->ir
, prog
, shader
->Type
== GL_FRAGMENT_SHADER
);
4955 count_resources(v
, prog
);
4957 _mesa_reference_program(ctx
, &shader
->Program
, prog
);
4959 /* This has to be done last. Any operation the can cause
4960 * prog->ParameterValues to get reallocated (e.g., anything that adds a
4961 * program constant) has to happen before creating this linkage.
4963 _mesa_associate_uniform_storage(ctx
, shader_program
, prog
->Parameters
);
4964 if (!shader_program
->LinkStatus
) {
4968 struct st_vertex_program
*stvp
;
4969 struct st_fragment_program
*stfp
;
4970 struct st_geometry_program
*stgp
;
4972 switch (shader
->Type
) {
4973 case GL_VERTEX_SHADER
:
4974 stvp
= (struct st_vertex_program
*)prog
;
4975 stvp
->glsl_to_tgsi
= v
;
4977 case GL_FRAGMENT_SHADER
:
4978 stfp
= (struct st_fragment_program
*)prog
;
4979 stfp
->glsl_to_tgsi
= v
;
4981 case GL_GEOMETRY_SHADER
:
4982 stgp
= (struct st_geometry_program
*)prog
;
4983 stgp
->glsl_to_tgsi
= v
;
4986 assert(!"should not be reached");
4996 st_new_shader(struct gl_context
*ctx
, GLuint name
, GLuint type
)
4998 struct gl_shader
*shader
;
4999 assert(type
== GL_FRAGMENT_SHADER
|| type
== GL_VERTEX_SHADER
||
5000 type
== GL_GEOMETRY_SHADER_ARB
);
5001 shader
= rzalloc(NULL
, struct gl_shader
);
5003 shader
->Type
= type
;
5004 shader
->Name
= name
;
5005 _mesa_init_shader(ctx
, shader
);
5010 struct gl_shader_program
*
5011 st_new_shader_program(struct gl_context
*ctx
, GLuint name
)
5013 struct gl_shader_program
*shProg
;
5014 shProg
= rzalloc(NULL
, struct gl_shader_program
);
5016 shProg
->Name
= name
;
5017 _mesa_init_shader_program(ctx
, shProg
);
5024 * Called via ctx->Driver.LinkShader()
5025 * This actually involves converting GLSL IR into an intermediate TGSI-like IR
5026 * with code lowering and other optimizations.
5029 st_link_shader(struct gl_context
*ctx
, struct gl_shader_program
*prog
)
5031 assert(prog
->LinkStatus
);
5033 for (unsigned i
= 0; i
< MESA_SHADER_TYPES
; i
++) {
5034 if (prog
->_LinkedShaders
[i
] == NULL
)
5038 exec_list
*ir
= prog
->_LinkedShaders
[i
]->ir
;
5039 const struct gl_shader_compiler_options
*options
=
5040 &ctx
->ShaderCompilerOptions
[_mesa_shader_type_to_index(prog
->_LinkedShaders
[i
]->Type
)];
5046 do_mat_op_to_vec(ir
);
5047 lower_instructions(ir
, (MOD_TO_FRACT
| DIV_TO_MUL_RCP
| EXP_TO_EXP2
5048 | LOG_TO_LOG2
| INT_DIV_TO_MUL_RCP
5049 | ((options
->EmitNoPow
) ? POW_TO_EXP2
: 0)));
5051 progress
= do_lower_jumps(ir
, true, true, options
->EmitNoMainReturn
, options
->EmitNoCont
, options
->EmitNoLoops
) || progress
;
5053 progress
= do_common_optimization(ir
, true, true,
5054 options
->MaxUnrollIterations
)
5057 progress
= lower_quadop_vector(ir
, false) || progress
;
5059 if (options
->MaxIfDepth
== 0)
5060 progress
= lower_discard(ir
) || progress
;
5062 progress
= lower_if_to_cond_assign(ir
, options
->MaxIfDepth
) || progress
;
5064 if (options
->EmitNoNoise
)
5065 progress
= lower_noise(ir
) || progress
;
5067 /* If there are forms of indirect addressing that the driver
5068 * cannot handle, perform the lowering pass.
5070 if (options
->EmitNoIndirectInput
|| options
->EmitNoIndirectOutput
5071 || options
->EmitNoIndirectTemp
|| options
->EmitNoIndirectUniform
)
5073 lower_variable_index_to_cond_assign(ir
,
5074 options
->EmitNoIndirectInput
,
5075 options
->EmitNoIndirectOutput
,
5076 options
->EmitNoIndirectTemp
,
5077 options
->EmitNoIndirectUniform
)
5080 progress
= do_vec_index_to_cond_assign(ir
) || progress
;
5083 validate_ir_tree(ir
);
5086 for (unsigned i
= 0; i
< MESA_SHADER_TYPES
; i
++) {
5087 struct gl_program
*linked_prog
;
5089 if (prog
->_LinkedShaders
[i
] == NULL
)
5092 linked_prog
= get_mesa_program(ctx
, prog
, prog
->_LinkedShaders
[i
]);
5095 static const GLenum targets
[] = {
5096 GL_VERTEX_PROGRAM_ARB
,
5097 GL_FRAGMENT_PROGRAM_ARB
,
5098 GL_GEOMETRY_PROGRAM_NV
5101 _mesa_reference_program(ctx
, &prog
->_LinkedShaders
[i
]->Program
,
5103 if (!ctx
->Driver
.ProgramStringNotify(ctx
, targets
[i
], linked_prog
)) {
5104 _mesa_reference_program(ctx
, &prog
->_LinkedShaders
[i
]->Program
,
5106 _mesa_reference_program(ctx
, &linked_prog
, NULL
);
5111 _mesa_reference_program(ctx
, &linked_prog
, NULL
);