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
;
359 unsigned num_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 simplify_cmp(void);
418 void rename_temp_register(int index
, int new_index
);
419 int get_first_temp_read(int index
);
420 int get_first_temp_write(int index
);
421 int get_last_temp_read(int index
);
422 int get_last_temp_write(int index
);
424 void copy_propagate(void);
425 void eliminate_dead_code(void);
426 int eliminate_dead_code_advanced(void);
427 void merge_registers(void);
428 void renumber_registers(void);
433 static st_src_reg undef_src
= st_src_reg(PROGRAM_UNDEFINED
, 0, GLSL_TYPE_ERROR
);
435 static st_dst_reg undef_dst
= st_dst_reg(PROGRAM_UNDEFINED
, SWIZZLE_NOOP
, GLSL_TYPE_ERROR
);
437 static st_dst_reg address_reg
= st_dst_reg(PROGRAM_ADDRESS
, WRITEMASK_X
, GLSL_TYPE_FLOAT
);
440 fail_link(struct gl_shader_program
*prog
, const char *fmt
, ...) PRINTFLIKE(2, 3);
443 fail_link(struct gl_shader_program
*prog
, const char *fmt
, ...)
447 ralloc_vasprintf_append(&prog
->InfoLog
, fmt
, args
);
450 prog
->LinkStatus
= GL_FALSE
;
454 swizzle_for_size(int size
)
456 int size_swizzles
[4] = {
457 MAKE_SWIZZLE4(SWIZZLE_X
, SWIZZLE_X
, SWIZZLE_X
, SWIZZLE_X
),
458 MAKE_SWIZZLE4(SWIZZLE_X
, SWIZZLE_Y
, SWIZZLE_Y
, SWIZZLE_Y
),
459 MAKE_SWIZZLE4(SWIZZLE_X
, SWIZZLE_Y
, SWIZZLE_Z
, SWIZZLE_Z
),
460 MAKE_SWIZZLE4(SWIZZLE_X
, SWIZZLE_Y
, SWIZZLE_Z
, SWIZZLE_W
),
463 assert((size
>= 1) && (size
<= 4));
464 return size_swizzles
[size
- 1];
468 is_tex_instruction(unsigned opcode
)
470 const tgsi_opcode_info
* info
= tgsi_get_opcode_info(opcode
);
475 num_inst_dst_regs(unsigned opcode
)
477 const tgsi_opcode_info
* info
= tgsi_get_opcode_info(opcode
);
478 return info
->num_dst
;
482 num_inst_src_regs(unsigned opcode
)
484 const tgsi_opcode_info
* info
= tgsi_get_opcode_info(opcode
);
485 return info
->is_tex
? info
->num_src
- 1 : info
->num_src
;
488 glsl_to_tgsi_instruction
*
489 glsl_to_tgsi_visitor::emit(ir_instruction
*ir
, unsigned op
,
491 st_src_reg src0
, st_src_reg src1
, st_src_reg src2
)
493 glsl_to_tgsi_instruction
*inst
= new(mem_ctx
) glsl_to_tgsi_instruction();
494 int num_reladdr
= 0, i
;
496 op
= get_opcode(ir
, op
, dst
, src0
, src1
);
498 /* If we have to do relative addressing, we want to load the ARL
499 * reg directly for one of the regs, and preload the other reladdr
500 * sources into temps.
502 num_reladdr
+= dst
.reladdr
!= NULL
;
503 num_reladdr
+= src0
.reladdr
!= NULL
;
504 num_reladdr
+= src1
.reladdr
!= NULL
;
505 num_reladdr
+= src2
.reladdr
!= NULL
;
507 reladdr_to_temp(ir
, &src2
, &num_reladdr
);
508 reladdr_to_temp(ir
, &src1
, &num_reladdr
);
509 reladdr_to_temp(ir
, &src0
, &num_reladdr
);
512 emit_arl(ir
, address_reg
, *dst
.reladdr
);
515 assert(num_reladdr
== 0);
525 inst
->function
= NULL
;
527 if (op
== TGSI_OPCODE_ARL
|| op
== TGSI_OPCODE_UARL
)
528 this->num_address_regs
= 1;
530 /* Update indirect addressing status used by TGSI */
533 case PROGRAM_TEMPORARY
:
534 this->indirect_addr_temps
= true;
536 case PROGRAM_LOCAL_PARAM
:
537 case PROGRAM_ENV_PARAM
:
538 case PROGRAM_STATE_VAR
:
539 case PROGRAM_NAMED_PARAM
:
540 case PROGRAM_CONSTANT
:
541 case PROGRAM_UNIFORM
:
542 this->indirect_addr_consts
= true;
544 case PROGRAM_IMMEDIATE
:
545 assert(!"immediates should not have indirect addressing");
552 for (i
=0; i
<3; i
++) {
553 if(inst
->src
[i
].reladdr
) {
554 switch(inst
->src
[i
].file
) {
555 case PROGRAM_TEMPORARY
:
556 this->indirect_addr_temps
= true;
558 case PROGRAM_LOCAL_PARAM
:
559 case PROGRAM_ENV_PARAM
:
560 case PROGRAM_STATE_VAR
:
561 case PROGRAM_NAMED_PARAM
:
562 case PROGRAM_CONSTANT
:
563 case PROGRAM_UNIFORM
:
564 this->indirect_addr_consts
= true;
566 case PROGRAM_IMMEDIATE
:
567 assert(!"immediates should not have indirect addressing");
576 this->instructions
.push_tail(inst
);
579 try_emit_float_set(ir
, op
, dst
);
585 glsl_to_tgsi_instruction
*
586 glsl_to_tgsi_visitor::emit(ir_instruction
*ir
, unsigned op
,
587 st_dst_reg dst
, st_src_reg src0
, st_src_reg src1
)
589 return emit(ir
, op
, dst
, src0
, src1
, undef_src
);
592 glsl_to_tgsi_instruction
*
593 glsl_to_tgsi_visitor::emit(ir_instruction
*ir
, unsigned op
,
594 st_dst_reg dst
, st_src_reg src0
)
596 assert(dst
.writemask
!= 0);
597 return emit(ir
, op
, dst
, src0
, undef_src
, undef_src
);
600 glsl_to_tgsi_instruction
*
601 glsl_to_tgsi_visitor::emit(ir_instruction
*ir
, unsigned op
)
603 return emit(ir
, op
, undef_dst
, undef_src
, undef_src
, undef_src
);
607 * Emits the code to convert the result of float SET instructions to integers.
610 glsl_to_tgsi_visitor::try_emit_float_set(ir_instruction
*ir
, unsigned op
,
613 if ((op
== TGSI_OPCODE_SEQ
||
614 op
== TGSI_OPCODE_SNE
||
615 op
== TGSI_OPCODE_SGE
||
616 op
== TGSI_OPCODE_SLT
))
618 st_src_reg src
= st_src_reg(dst
);
619 src
.negate
= ~src
.negate
;
620 dst
.type
= GLSL_TYPE_FLOAT
;
621 emit(ir
, TGSI_OPCODE_F2I
, dst
, src
);
626 * Determines whether to use an integer, unsigned integer, or float opcode
627 * based on the operands and input opcode, then emits the result.
630 glsl_to_tgsi_visitor::get_opcode(ir_instruction
*ir
, unsigned op
,
632 st_src_reg src0
, st_src_reg src1
)
634 int type
= GLSL_TYPE_FLOAT
;
636 if (src0
.type
== GLSL_TYPE_FLOAT
|| src1
.type
== GLSL_TYPE_FLOAT
)
637 type
= GLSL_TYPE_FLOAT
;
638 else if (native_integers
)
639 type
= src0
.type
== GLSL_TYPE_BOOL
? GLSL_TYPE_INT
: src0
.type
;
641 #define case4(c, f, i, u) \
642 case TGSI_OPCODE_##c: \
643 if (type == GLSL_TYPE_INT) op = TGSI_OPCODE_##i; \
644 else if (type == GLSL_TYPE_UINT) op = TGSI_OPCODE_##u; \
645 else op = TGSI_OPCODE_##f; \
647 #define case3(f, i, u) case4(f, f, i, u)
648 #define case2fi(f, i) case4(f, f, i, i)
649 #define case2iu(i, u) case4(i, LAST, i, u)
655 case3(DIV
, IDIV
, UDIV
);
656 case3(MAX
, IMAX
, UMAX
);
657 case3(MIN
, IMIN
, UMIN
);
662 case3(SGE
, ISGE
, USGE
);
663 case3(SLT
, ISLT
, USLT
);
668 case3(ABS
, IABS
, IABS
);
673 assert(op
!= TGSI_OPCODE_LAST
);
677 glsl_to_tgsi_instruction
*
678 glsl_to_tgsi_visitor::emit_dp(ir_instruction
*ir
,
679 st_dst_reg dst
, st_src_reg src0
, st_src_reg src1
,
682 static const unsigned dot_opcodes
[] = {
683 TGSI_OPCODE_DP2
, TGSI_OPCODE_DP3
, TGSI_OPCODE_DP4
686 return emit(ir
, dot_opcodes
[elements
- 2], dst
, src0
, src1
);
690 * Emits TGSI scalar opcodes to produce unique answers across channels.
692 * Some TGSI opcodes are scalar-only, like ARB_fp/vp. The src X
693 * channel determines the result across all channels. So to do a vec4
694 * of this operation, we want to emit a scalar per source channel used
695 * to produce dest channels.
698 glsl_to_tgsi_visitor::emit_scalar(ir_instruction
*ir
, unsigned op
,
700 st_src_reg orig_src0
, st_src_reg orig_src1
)
703 int done_mask
= ~dst
.writemask
;
705 /* TGSI RCP is a scalar operation splatting results to all channels,
706 * like ARB_fp/vp. So emit as many RCPs as necessary to cover our
709 for (i
= 0; i
< 4; i
++) {
710 GLuint this_mask
= (1 << i
);
711 glsl_to_tgsi_instruction
*inst
;
712 st_src_reg src0
= orig_src0
;
713 st_src_reg src1
= orig_src1
;
715 if (done_mask
& this_mask
)
718 GLuint src0_swiz
= GET_SWZ(src0
.swizzle
, i
);
719 GLuint src1_swiz
= GET_SWZ(src1
.swizzle
, i
);
720 for (j
= i
+ 1; j
< 4; j
++) {
721 /* If there is another enabled component in the destination that is
722 * derived from the same inputs, generate its value on this pass as
725 if (!(done_mask
& (1 << j
)) &&
726 GET_SWZ(src0
.swizzle
, j
) == src0_swiz
&&
727 GET_SWZ(src1
.swizzle
, j
) == src1_swiz
) {
728 this_mask
|= (1 << j
);
731 src0
.swizzle
= MAKE_SWIZZLE4(src0_swiz
, src0_swiz
,
732 src0_swiz
, src0_swiz
);
733 src1
.swizzle
= MAKE_SWIZZLE4(src1_swiz
, src1_swiz
,
734 src1_swiz
, src1_swiz
);
736 inst
= emit(ir
, op
, dst
, src0
, src1
);
737 inst
->dst
.writemask
= this_mask
;
738 done_mask
|= this_mask
;
743 glsl_to_tgsi_visitor::emit_scalar(ir_instruction
*ir
, unsigned op
,
744 st_dst_reg dst
, st_src_reg src0
)
746 st_src_reg undef
= undef_src
;
748 undef
.swizzle
= SWIZZLE_XXXX
;
750 emit_scalar(ir
, op
, dst
, src0
, undef
);
754 glsl_to_tgsi_visitor::emit_arl(ir_instruction
*ir
,
755 st_dst_reg dst
, st_src_reg src0
)
757 int op
= TGSI_OPCODE_ARL
;
759 if (src0
.type
== GLSL_TYPE_INT
|| src0
.type
== GLSL_TYPE_UINT
)
760 op
= TGSI_OPCODE_UARL
;
762 emit(NULL
, op
, dst
, src0
);
766 * Emit an TGSI_OPCODE_SCS instruction
768 * The \c SCS opcode functions a bit differently than the other TGSI opcodes.
769 * Instead of splatting its result across all four components of the
770 * destination, it writes one value to the \c x component and another value to
771 * the \c y component.
773 * \param ir IR instruction being processed
774 * \param op Either \c TGSI_OPCODE_SIN or \c TGSI_OPCODE_COS depending
775 * on which value is desired.
776 * \param dst Destination register
777 * \param src Source register
780 glsl_to_tgsi_visitor::emit_scs(ir_instruction
*ir
, unsigned op
,
782 const st_src_reg
&src
)
784 /* Vertex programs cannot use the SCS opcode.
786 if (this->prog
->Target
== GL_VERTEX_PROGRAM_ARB
) {
787 emit_scalar(ir
, op
, dst
, src
);
791 const unsigned component
= (op
== TGSI_OPCODE_SIN
) ? 0 : 1;
792 const unsigned scs_mask
= (1U << component
);
793 int done_mask
= ~dst
.writemask
;
796 assert(op
== TGSI_OPCODE_SIN
|| op
== TGSI_OPCODE_COS
);
798 /* If there are compnents in the destination that differ from the component
799 * that will be written by the SCS instrution, we'll need a temporary.
801 if (scs_mask
!= unsigned(dst
.writemask
)) {
802 tmp
= get_temp(glsl_type::vec4_type
);
805 for (unsigned i
= 0; i
< 4; i
++) {
806 unsigned this_mask
= (1U << i
);
807 st_src_reg src0
= src
;
809 if ((done_mask
& this_mask
) != 0)
812 /* The source swizzle specified which component of the source generates
813 * sine / cosine for the current component in the destination. The SCS
814 * instruction requires that this value be swizzle to the X component.
815 * Replace the current swizzle with a swizzle that puts the source in
818 unsigned src0_swiz
= GET_SWZ(src
.swizzle
, i
);
820 src0
.swizzle
= MAKE_SWIZZLE4(src0_swiz
, src0_swiz
,
821 src0_swiz
, src0_swiz
);
822 for (unsigned j
= i
+ 1; j
< 4; j
++) {
823 /* If there is another enabled component in the destination that is
824 * derived from the same inputs, generate its value on this pass as
827 if (!(done_mask
& (1 << j
)) &&
828 GET_SWZ(src0
.swizzle
, j
) == src0_swiz
) {
829 this_mask
|= (1 << j
);
833 if (this_mask
!= scs_mask
) {
834 glsl_to_tgsi_instruction
*inst
;
835 st_dst_reg tmp_dst
= st_dst_reg(tmp
);
837 /* Emit the SCS instruction.
839 inst
= emit(ir
, TGSI_OPCODE_SCS
, tmp_dst
, src0
);
840 inst
->dst
.writemask
= scs_mask
;
842 /* Move the result of the SCS instruction to the desired location in
845 tmp
.swizzle
= MAKE_SWIZZLE4(component
, component
,
846 component
, component
);
847 inst
= emit(ir
, TGSI_OPCODE_SCS
, dst
, tmp
);
848 inst
->dst
.writemask
= this_mask
;
850 /* Emit the SCS instruction to write directly to the destination.
852 glsl_to_tgsi_instruction
*inst
= emit(ir
, TGSI_OPCODE_SCS
, dst
, src0
);
853 inst
->dst
.writemask
= scs_mask
;
856 done_mask
|= this_mask
;
861 glsl_to_tgsi_visitor::add_constant(gl_register_file file
,
862 gl_constant_value values
[4], int size
, int datatype
,
865 if (file
== PROGRAM_CONSTANT
) {
866 return _mesa_add_typed_unnamed_constant(this->prog
->Parameters
, values
,
867 size
, datatype
, swizzle_out
);
870 immediate_storage
*entry
;
871 assert(file
== PROGRAM_IMMEDIATE
);
873 /* Search immediate storage to see if we already have an identical
874 * immediate that we can use instead of adding a duplicate entry.
876 foreach_iter(exec_list_iterator
, iter
, this->immediates
) {
877 entry
= (immediate_storage
*)iter
.get();
879 if (entry
->size
== size
&&
880 entry
->type
== datatype
&&
881 !memcmp(entry
->values
, values
, size
* sizeof(gl_constant_value
))) {
887 /* Add this immediate to the list. */
888 entry
= new(mem_ctx
) immediate_storage(values
, size
, datatype
);
889 this->immediates
.push_tail(entry
);
890 this->num_immediates
++;
896 glsl_to_tgsi_visitor::st_src_reg_for_float(float val
)
898 st_src_reg
src(PROGRAM_IMMEDIATE
, -1, GLSL_TYPE_FLOAT
);
899 union gl_constant_value uval
;
902 src
.index
= add_constant(src
.file
, &uval
, 1, GL_FLOAT
, &src
.swizzle
);
908 glsl_to_tgsi_visitor::st_src_reg_for_int(int val
)
910 st_src_reg
src(PROGRAM_IMMEDIATE
, -1, GLSL_TYPE_INT
);
911 union gl_constant_value uval
;
913 assert(native_integers
);
916 src
.index
= add_constant(src
.file
, &uval
, 1, GL_INT
, &src
.swizzle
);
922 glsl_to_tgsi_visitor::st_src_reg_for_type(int type
, int val
)
925 return type
== GLSL_TYPE_FLOAT
? st_src_reg_for_float(val
) :
926 st_src_reg_for_int(val
);
928 return st_src_reg_for_float(val
);
932 type_size(const struct glsl_type
*type
)
937 switch (type
->base_type
) {
940 case GLSL_TYPE_FLOAT
:
942 if (type
->is_matrix()) {
943 return type
->matrix_columns
;
945 /* Regardless of size of vector, it gets a vec4. This is bad
946 * packing for things like floats, but otherwise arrays become a
947 * mess. Hopefully a later pass over the code can pack scalars
948 * down if appropriate.
952 case GLSL_TYPE_ARRAY
:
953 assert(type
->length
> 0);
954 return type_size(type
->fields
.array
) * type
->length
;
955 case GLSL_TYPE_STRUCT
:
957 for (i
= 0; i
< type
->length
; i
++) {
958 size
+= type_size(type
->fields
.structure
[i
].type
);
961 case GLSL_TYPE_SAMPLER
:
962 /* Samplers take up one slot in UNIFORMS[], but they're baked in
973 * In the initial pass of codegen, we assign temporary numbers to
974 * intermediate results. (not SSA -- variable assignments will reuse
978 glsl_to_tgsi_visitor::get_temp(const glsl_type
*type
)
982 src
.type
= native_integers
? type
->base_type
: GLSL_TYPE_FLOAT
;
983 src
.file
= PROGRAM_TEMPORARY
;
984 src
.index
= next_temp
;
986 next_temp
+= type_size(type
);
988 if (type
->is_array() || type
->is_record()) {
989 src
.swizzle
= SWIZZLE_NOOP
;
991 src
.swizzle
= swizzle_for_size(type
->vector_elements
);
999 glsl_to_tgsi_visitor::find_variable_storage(ir_variable
*var
)
1002 variable_storage
*entry
;
1004 foreach_iter(exec_list_iterator
, iter
, this->variables
) {
1005 entry
= (variable_storage
*)iter
.get();
1007 if (entry
->var
== var
)
1015 glsl_to_tgsi_visitor::visit(ir_variable
*ir
)
1017 if (strcmp(ir
->name
, "gl_FragCoord") == 0) {
1018 struct gl_fragment_program
*fp
= (struct gl_fragment_program
*)this->prog
;
1020 fp
->OriginUpperLeft
= ir
->origin_upper_left
;
1021 fp
->PixelCenterInteger
= ir
->pixel_center_integer
;
1024 if (ir
->mode
== ir_var_uniform
&& strncmp(ir
->name
, "gl_", 3) == 0) {
1026 const ir_state_slot
*const slots
= ir
->state_slots
;
1027 assert(ir
->state_slots
!= NULL
);
1029 /* Check if this statevar's setup in the STATE file exactly
1030 * matches how we'll want to reference it as a
1031 * struct/array/whatever. If not, then we need to move it into
1032 * temporary storage and hope that it'll get copy-propagated
1035 for (i
= 0; i
< ir
->num_state_slots
; i
++) {
1036 if (slots
[i
].swizzle
!= SWIZZLE_XYZW
) {
1041 variable_storage
*storage
;
1043 if (i
== ir
->num_state_slots
) {
1044 /* We'll set the index later. */
1045 storage
= new(mem_ctx
) variable_storage(ir
, PROGRAM_STATE_VAR
, -1);
1046 this->variables
.push_tail(storage
);
1050 /* The variable_storage constructor allocates slots based on the size
1051 * of the type. However, this had better match the number of state
1052 * elements that we're going to copy into the new temporary.
1054 assert((int) ir
->num_state_slots
== type_size(ir
->type
));
1056 storage
= new(mem_ctx
) variable_storage(ir
, PROGRAM_TEMPORARY
,
1058 this->variables
.push_tail(storage
);
1059 this->next_temp
+= type_size(ir
->type
);
1061 dst
= st_dst_reg(st_src_reg(PROGRAM_TEMPORARY
, storage
->index
,
1062 native_integers
? ir
->type
->base_type
: GLSL_TYPE_FLOAT
));
1066 for (unsigned int i
= 0; i
< ir
->num_state_slots
; i
++) {
1067 int index
= _mesa_add_state_reference(this->prog
->Parameters
,
1068 (gl_state_index
*)slots
[i
].tokens
);
1070 if (storage
->file
== PROGRAM_STATE_VAR
) {
1071 if (storage
->index
== -1) {
1072 storage
->index
= index
;
1074 assert(index
== storage
->index
+ (int)i
);
1077 st_src_reg
src(PROGRAM_STATE_VAR
, index
,
1078 native_integers
? ir
->type
->base_type
: GLSL_TYPE_FLOAT
);
1079 src
.swizzle
= slots
[i
].swizzle
;
1080 emit(ir
, TGSI_OPCODE_MOV
, dst
, src
);
1081 /* even a float takes up a whole vec4 reg in a struct/array. */
1086 if (storage
->file
== PROGRAM_TEMPORARY
&&
1087 dst
.index
!= storage
->index
+ (int) ir
->num_state_slots
) {
1088 fail_link(this->shader_program
,
1089 "failed to load builtin uniform `%s' (%d/%d regs loaded)\n",
1090 ir
->name
, dst
.index
- storage
->index
,
1091 type_size(ir
->type
));
1097 glsl_to_tgsi_visitor::visit(ir_loop
*ir
)
1099 ir_dereference_variable
*counter
= NULL
;
1101 if (ir
->counter
!= NULL
)
1102 counter
= new(ir
) ir_dereference_variable(ir
->counter
);
1104 if (ir
->from
!= NULL
) {
1105 assert(ir
->counter
!= NULL
);
1107 ir_assignment
*a
= new(ir
) ir_assignment(counter
, ir
->from
, NULL
);
1113 emit(NULL
, TGSI_OPCODE_BGNLOOP
);
1117 new(ir
) ir_expression(ir
->cmp
, glsl_type::bool_type
,
1119 ir_if
*if_stmt
= new(ir
) ir_if(e
);
1121 ir_loop_jump
*brk
= new(ir
) ir_loop_jump(ir_loop_jump::jump_break
);
1123 if_stmt
->then_instructions
.push_tail(brk
);
1125 if_stmt
->accept(this);
1132 visit_exec_list(&ir
->body_instructions
, this);
1134 if (ir
->increment
) {
1136 new(ir
) ir_expression(ir_binop_add
, counter
->type
,
1137 counter
, ir
->increment
);
1139 ir_assignment
*a
= new(ir
) ir_assignment(counter
, e
, NULL
);
1146 emit(NULL
, TGSI_OPCODE_ENDLOOP
);
1150 glsl_to_tgsi_visitor::visit(ir_loop_jump
*ir
)
1153 case ir_loop_jump::jump_break
:
1154 emit(NULL
, TGSI_OPCODE_BRK
);
1156 case ir_loop_jump::jump_continue
:
1157 emit(NULL
, TGSI_OPCODE_CONT
);
1164 glsl_to_tgsi_visitor::visit(ir_function_signature
*ir
)
1171 glsl_to_tgsi_visitor::visit(ir_function
*ir
)
1173 /* Ignore function bodies other than main() -- we shouldn't see calls to
1174 * them since they should all be inlined before we get to glsl_to_tgsi.
1176 if (strcmp(ir
->name
, "main") == 0) {
1177 const ir_function_signature
*sig
;
1180 sig
= ir
->matching_signature(&empty
);
1184 foreach_iter(exec_list_iterator
, iter
, sig
->body
) {
1185 ir_instruction
*ir
= (ir_instruction
*)iter
.get();
1193 glsl_to_tgsi_visitor::try_emit_mad(ir_expression
*ir
, int mul_operand
)
1195 int nonmul_operand
= 1 - mul_operand
;
1197 st_dst_reg result_dst
;
1199 ir_expression
*expr
= ir
->operands
[mul_operand
]->as_expression();
1200 if (!expr
|| expr
->operation
!= ir_binop_mul
)
1203 expr
->operands
[0]->accept(this);
1205 expr
->operands
[1]->accept(this);
1207 ir
->operands
[nonmul_operand
]->accept(this);
1210 this->result
= get_temp(ir
->type
);
1211 result_dst
= st_dst_reg(this->result
);
1212 result_dst
.writemask
= (1 << ir
->type
->vector_elements
) - 1;
1213 emit(ir
, TGSI_OPCODE_MAD
, result_dst
, a
, b
, c
);
1219 * Emit MAD(a, -b, a) instead of AND(a, NOT(b))
1221 * The logic values are 1.0 for true and 0.0 for false. Logical-and is
1222 * implemented using multiplication, and logical-or is implemented using
1223 * addition. Logical-not can be implemented as (true - x), or (1.0 - x).
1224 * As result, the logical expression (a & !b) can be rewritten as:
1228 * - (a * 1) - (a * b)
1232 * This final expression can be implemented as a single MAD(a, -b, a)
1236 glsl_to_tgsi_visitor::try_emit_mad_for_and_not(ir_expression
*ir
, int try_operand
)
1238 const int other_operand
= 1 - try_operand
;
1241 ir_expression
*expr
= ir
->operands
[try_operand
]->as_expression();
1242 if (!expr
|| expr
->operation
!= ir_unop_logic_not
)
1245 ir
->operands
[other_operand
]->accept(this);
1247 expr
->operands
[0]->accept(this);
1250 b
.negate
= ~b
.negate
;
1252 this->result
= get_temp(ir
->type
);
1253 emit(ir
, TGSI_OPCODE_MAD
, st_dst_reg(this->result
), a
, b
, a
);
1259 glsl_to_tgsi_visitor::try_emit_sat(ir_expression
*ir
)
1261 /* Saturates were only introduced to vertex programs in
1262 * NV_vertex_program3, so don't give them to drivers in the VP.
1264 if (this->prog
->Target
== GL_VERTEX_PROGRAM_ARB
)
1267 ir_rvalue
*sat_src
= ir
->as_rvalue_to_saturate();
1271 sat_src
->accept(this);
1272 st_src_reg src
= this->result
;
1274 /* If we generated an expression instruction into a temporary in
1275 * processing the saturate's operand, apply the saturate to that
1276 * instruction. Otherwise, generate a MOV to do the saturate.
1278 * Note that we have to be careful to only do this optimization if
1279 * the instruction in question was what generated src->result. For
1280 * example, ir_dereference_array might generate a MUL instruction
1281 * to create the reladdr, and return us a src reg using that
1282 * reladdr. That MUL result is not the value we're trying to
1285 ir_expression
*sat_src_expr
= sat_src
->as_expression();
1286 if (sat_src_expr
&& (sat_src_expr
->operation
== ir_binop_mul
||
1287 sat_src_expr
->operation
== ir_binop_add
||
1288 sat_src_expr
->operation
== ir_binop_dot
)) {
1289 glsl_to_tgsi_instruction
*new_inst
;
1290 new_inst
= (glsl_to_tgsi_instruction
*)this->instructions
.get_tail();
1291 new_inst
->saturate
= true;
1293 this->result
= get_temp(ir
->type
);
1294 st_dst_reg result_dst
= st_dst_reg(this->result
);
1295 result_dst
.writemask
= (1 << ir
->type
->vector_elements
) - 1;
1296 glsl_to_tgsi_instruction
*inst
;
1297 inst
= emit(ir
, TGSI_OPCODE_MOV
, result_dst
, src
);
1298 inst
->saturate
= true;
1305 glsl_to_tgsi_visitor::reladdr_to_temp(ir_instruction
*ir
,
1306 st_src_reg
*reg
, int *num_reladdr
)
1311 emit_arl(ir
, address_reg
, *reg
->reladdr
);
1313 if (*num_reladdr
!= 1) {
1314 st_src_reg temp
= get_temp(glsl_type::vec4_type
);
1316 emit(ir
, TGSI_OPCODE_MOV
, st_dst_reg(temp
), *reg
);
1324 glsl_to_tgsi_visitor::visit(ir_expression
*ir
)
1326 unsigned int operand
;
1327 st_src_reg op
[Elements(ir
->operands
)];
1328 st_src_reg result_src
;
1329 st_dst_reg result_dst
;
1331 /* Quick peephole: Emit MAD(a, b, c) instead of ADD(MUL(a, b), c)
1333 if (ir
->operation
== ir_binop_add
) {
1334 if (try_emit_mad(ir
, 1))
1336 if (try_emit_mad(ir
, 0))
1340 /* Quick peephole: Emit OPCODE_MAD(-a, -b, a) instead of AND(a, NOT(b))
1342 if (ir
->operation
== ir_binop_logic_and
) {
1343 if (try_emit_mad_for_and_not(ir
, 1))
1345 if (try_emit_mad_for_and_not(ir
, 0))
1349 if (try_emit_sat(ir
))
1352 if (ir
->operation
== ir_quadop_vector
)
1353 assert(!"ir_quadop_vector should have been lowered");
1355 for (operand
= 0; operand
< ir
->get_num_operands(); operand
++) {
1356 this->result
.file
= PROGRAM_UNDEFINED
;
1357 ir
->operands
[operand
]->accept(this);
1358 if (this->result
.file
== PROGRAM_UNDEFINED
) {
1360 printf("Failed to get tree for expression operand:\n");
1361 ir
->operands
[operand
]->accept(&v
);
1364 op
[operand
] = this->result
;
1366 /* Matrix expression operands should have been broken down to vector
1367 * operations already.
1369 assert(!ir
->operands
[operand
]->type
->is_matrix());
1372 int vector_elements
= ir
->operands
[0]->type
->vector_elements
;
1373 if (ir
->operands
[1]) {
1374 vector_elements
= MAX2(vector_elements
,
1375 ir
->operands
[1]->type
->vector_elements
);
1378 this->result
.file
= PROGRAM_UNDEFINED
;
1380 /* Storage for our result. Ideally for an assignment we'd be using
1381 * the actual storage for the result here, instead.
1383 result_src
= get_temp(ir
->type
);
1384 /* convenience for the emit functions below. */
1385 result_dst
= st_dst_reg(result_src
);
1386 /* Limit writes to the channels that will be used by result_src later.
1387 * This does limit this temp's use as a temporary for multi-instruction
1390 result_dst
.writemask
= (1 << ir
->type
->vector_elements
) - 1;
1392 switch (ir
->operation
) {
1393 case ir_unop_logic_not
:
1394 if (result_dst
.type
!= GLSL_TYPE_FLOAT
)
1395 emit(ir
, TGSI_OPCODE_NOT
, result_dst
, op
[0]);
1397 /* Previously 'SEQ dst, src, 0.0' was used for this. However, many
1398 * older GPUs implement SEQ using multiple instructions (i915 uses two
1399 * SGE instructions and a MUL instruction). Since our logic values are
1400 * 0.0 and 1.0, 1-x also implements !x.
1402 op
[0].negate
= ~op
[0].negate
;
1403 emit(ir
, TGSI_OPCODE_ADD
, result_dst
, op
[0], st_src_reg_for_float(1.0));
1407 if (result_dst
.type
== GLSL_TYPE_INT
|| result_dst
.type
== GLSL_TYPE_UINT
)
1408 emit(ir
, TGSI_OPCODE_INEG
, result_dst
, op
[0]);
1410 op
[0].negate
= ~op
[0].negate
;
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]);
1452 /* The X component contains 1 or -1 depending on whether the framebuffer
1453 * is a FBO or the window system buffer, respectively.
1454 * It is then multiplied with the source operand of DDY.
1456 static const gl_state_index transform_y_state
[STATE_LENGTH
]
1457 = { STATE_INTERNAL
, STATE_FB_WPOS_Y_TRANSFORM
};
1459 unsigned transform_y_index
=
1460 _mesa_add_state_reference(this->prog
->Parameters
,
1463 st_src_reg transform_y
= st_src_reg(PROGRAM_STATE_VAR
,
1465 glsl_type::vec4_type
);
1466 transform_y
.swizzle
= SWIZZLE_XXXX
;
1468 st_src_reg temp
= get_temp(glsl_type::vec4_type
);
1470 emit(ir
, TGSI_OPCODE_MUL
, st_dst_reg(temp
), transform_y
, op
[0]);
1471 emit(ir
, TGSI_OPCODE_DDY
, result_dst
, temp
);
1475 case ir_unop_noise
: {
1476 /* At some point, a motivated person could add a better
1477 * implementation of noise. Currently not even the nvidia
1478 * binary drivers do anything more than this. In any case, the
1479 * place to do this is in the GL state tracker, not the poor
1482 emit(ir
, TGSI_OPCODE_MOV
, result_dst
, st_src_reg_for_float(0.5));
1487 emit(ir
, TGSI_OPCODE_ADD
, result_dst
, op
[0], op
[1]);
1490 emit(ir
, TGSI_OPCODE_SUB
, result_dst
, op
[0], op
[1]);
1494 emit(ir
, TGSI_OPCODE_MUL
, result_dst
, op
[0], op
[1]);
1497 if (result_dst
.type
== GLSL_TYPE_FLOAT
)
1498 assert(!"not reached: should be handled by ir_div_to_mul_rcp");
1500 emit(ir
, TGSI_OPCODE_DIV
, result_dst
, op
[0], op
[1]);
1503 if (result_dst
.type
== GLSL_TYPE_FLOAT
)
1504 assert(!"ir_binop_mod should have been converted to b * fract(a/b)");
1506 emit(ir
, TGSI_OPCODE_MOD
, result_dst
, op
[0], op
[1]);
1510 emit(ir
, TGSI_OPCODE_SLT
, result_dst
, op
[0], op
[1]);
1512 case ir_binop_greater
:
1513 emit(ir
, TGSI_OPCODE_SLT
, result_dst
, op
[1], op
[0]);
1515 case ir_binop_lequal
:
1516 emit(ir
, TGSI_OPCODE_SGE
, result_dst
, op
[1], op
[0]);
1518 case ir_binop_gequal
:
1519 emit(ir
, TGSI_OPCODE_SGE
, result_dst
, op
[0], op
[1]);
1521 case ir_binop_equal
:
1522 emit(ir
, TGSI_OPCODE_SEQ
, result_dst
, op
[0], op
[1]);
1524 case ir_binop_nequal
:
1525 emit(ir
, TGSI_OPCODE_SNE
, result_dst
, op
[0], op
[1]);
1527 case ir_binop_all_equal
:
1528 /* "==" operator producing a scalar boolean. */
1529 if (ir
->operands
[0]->type
->is_vector() ||
1530 ir
->operands
[1]->type
->is_vector()) {
1531 st_src_reg temp
= get_temp(native_integers
?
1532 glsl_type::get_instance(ir
->operands
[0]->type
->base_type
, 4, 1) :
1533 glsl_type::vec4_type
);
1535 if (native_integers
) {
1536 st_dst_reg temp_dst
= st_dst_reg(temp
);
1537 st_src_reg temp1
= st_src_reg(temp
), temp2
= st_src_reg(temp
);
1539 emit(ir
, TGSI_OPCODE_SEQ
, st_dst_reg(temp
), op
[0], op
[1]);
1541 /* Emit 1-3 AND operations to combine the SEQ results. */
1542 switch (ir
->operands
[0]->type
->vector_elements
) {
1546 temp_dst
.writemask
= WRITEMASK_Y
;
1547 temp1
.swizzle
= SWIZZLE_YYYY
;
1548 temp2
.swizzle
= SWIZZLE_ZZZZ
;
1549 emit(ir
, TGSI_OPCODE_AND
, temp_dst
, temp1
, temp2
);
1552 temp_dst
.writemask
= WRITEMASK_X
;
1553 temp1
.swizzle
= SWIZZLE_XXXX
;
1554 temp2
.swizzle
= SWIZZLE_YYYY
;
1555 emit(ir
, TGSI_OPCODE_AND
, temp_dst
, temp1
, temp2
);
1556 temp_dst
.writemask
= WRITEMASK_Y
;
1557 temp1
.swizzle
= SWIZZLE_ZZZZ
;
1558 temp2
.swizzle
= SWIZZLE_WWWW
;
1559 emit(ir
, TGSI_OPCODE_AND
, temp_dst
, temp1
, temp2
);
1562 temp1
.swizzle
= SWIZZLE_XXXX
;
1563 temp2
.swizzle
= SWIZZLE_YYYY
;
1564 emit(ir
, TGSI_OPCODE_AND
, result_dst
, temp1
, temp2
);
1566 emit(ir
, TGSI_OPCODE_SNE
, st_dst_reg(temp
), op
[0], op
[1]);
1568 /* After the dot-product, the value will be an integer on the
1569 * range [0,4]. Zero becomes 1.0, and positive values become zero.
1571 emit_dp(ir
, result_dst
, temp
, temp
, vector_elements
);
1573 /* Negating the result of the dot-product gives values on the range
1574 * [-4, 0]. Zero becomes 1.0, and negative values become zero.
1575 * This is achieved using SGE.
1577 st_src_reg sge_src
= result_src
;
1578 sge_src
.negate
= ~sge_src
.negate
;
1579 emit(ir
, TGSI_OPCODE_SGE
, result_dst
, sge_src
, st_src_reg_for_float(0.0));
1582 emit(ir
, TGSI_OPCODE_SEQ
, result_dst
, op
[0], op
[1]);
1585 case ir_binop_any_nequal
:
1586 /* "!=" operator producing a scalar boolean. */
1587 if (ir
->operands
[0]->type
->is_vector() ||
1588 ir
->operands
[1]->type
->is_vector()) {
1589 st_src_reg temp
= get_temp(native_integers
?
1590 glsl_type::get_instance(ir
->operands
[0]->type
->base_type
, 4, 1) :
1591 glsl_type::vec4_type
);
1592 emit(ir
, TGSI_OPCODE_SNE
, st_dst_reg(temp
), op
[0], op
[1]);
1594 if (native_integers
) {
1595 st_dst_reg temp_dst
= st_dst_reg(temp
);
1596 st_src_reg temp1
= st_src_reg(temp
), temp2
= st_src_reg(temp
);
1598 /* Emit 1-3 OR operations to combine the SNE results. */
1599 switch (ir
->operands
[0]->type
->vector_elements
) {
1603 temp_dst
.writemask
= WRITEMASK_Y
;
1604 temp1
.swizzle
= SWIZZLE_YYYY
;
1605 temp2
.swizzle
= SWIZZLE_ZZZZ
;
1606 emit(ir
, TGSI_OPCODE_OR
, temp_dst
, temp1
, temp2
);
1609 temp_dst
.writemask
= WRITEMASK_X
;
1610 temp1
.swizzle
= SWIZZLE_XXXX
;
1611 temp2
.swizzle
= SWIZZLE_YYYY
;
1612 emit(ir
, TGSI_OPCODE_OR
, temp_dst
, temp1
, temp2
);
1613 temp_dst
.writemask
= WRITEMASK_Y
;
1614 temp1
.swizzle
= SWIZZLE_ZZZZ
;
1615 temp2
.swizzle
= SWIZZLE_WWWW
;
1616 emit(ir
, TGSI_OPCODE_OR
, temp_dst
, temp1
, temp2
);
1619 temp1
.swizzle
= SWIZZLE_XXXX
;
1620 temp2
.swizzle
= SWIZZLE_YYYY
;
1621 emit(ir
, TGSI_OPCODE_OR
, result_dst
, temp1
, temp2
);
1623 /* After the dot-product, the value will be an integer on the
1624 * range [0,4]. Zero stays zero, and positive values become 1.0.
1626 glsl_to_tgsi_instruction
*const dp
=
1627 emit_dp(ir
, result_dst
, temp
, temp
, vector_elements
);
1628 if (this->prog
->Target
== GL_FRAGMENT_PROGRAM_ARB
) {
1629 /* The clamping to [0,1] can be done for free in the fragment
1630 * shader with a saturate.
1632 dp
->saturate
= true;
1634 /* Negating the result of the dot-product gives values on the range
1635 * [-4, 0]. Zero stays zero, and negative values become 1.0. This
1636 * achieved using SLT.
1638 st_src_reg slt_src
= result_src
;
1639 slt_src
.negate
= ~slt_src
.negate
;
1640 emit(ir
, TGSI_OPCODE_SLT
, result_dst
, slt_src
, st_src_reg_for_float(0.0));
1644 emit(ir
, TGSI_OPCODE_SNE
, result_dst
, op
[0], op
[1]);
1649 assert(ir
->operands
[0]->type
->is_vector());
1651 /* After the dot-product, the value will be an integer on the
1652 * range [0,4]. Zero stays zero, and positive values become 1.0.
1654 glsl_to_tgsi_instruction
*const dp
=
1655 emit_dp(ir
, result_dst
, op
[0], op
[0],
1656 ir
->operands
[0]->type
->vector_elements
);
1657 if (this->prog
->Target
== GL_FRAGMENT_PROGRAM_ARB
&&
1658 result_dst
.type
== GLSL_TYPE_FLOAT
) {
1659 /* The clamping to [0,1] can be done for free in the fragment
1660 * shader with a saturate.
1662 dp
->saturate
= true;
1663 } else if (result_dst
.type
== GLSL_TYPE_FLOAT
) {
1664 /* Negating the result of the dot-product gives values on the range
1665 * [-4, 0]. Zero stays zero, and negative values become 1.0. This
1666 * is achieved using SLT.
1668 st_src_reg slt_src
= result_src
;
1669 slt_src
.negate
= ~slt_src
.negate
;
1670 emit(ir
, TGSI_OPCODE_SLT
, result_dst
, slt_src
, st_src_reg_for_float(0.0));
1673 /* Use SNE 0 if integers are being used as boolean values. */
1674 emit(ir
, TGSI_OPCODE_SNE
, result_dst
, result_src
, st_src_reg_for_int(0));
1679 case ir_binop_logic_xor
:
1680 if (native_integers
)
1681 emit(ir
, TGSI_OPCODE_XOR
, result_dst
, op
[0], op
[1]);
1683 emit(ir
, TGSI_OPCODE_SNE
, result_dst
, op
[0], op
[1]);
1686 case ir_binop_logic_or
: {
1687 if (native_integers
) {
1688 /* If integers are used as booleans, we can use an actual "or"
1691 assert(native_integers
);
1692 emit(ir
, TGSI_OPCODE_OR
, result_dst
, op
[0], op
[1]);
1694 /* After the addition, the value will be an integer on the
1695 * range [0,2]. Zero stays zero, and positive values become 1.0.
1697 glsl_to_tgsi_instruction
*add
=
1698 emit(ir
, TGSI_OPCODE_ADD
, result_dst
, op
[0], op
[1]);
1699 if (this->prog
->Target
== GL_FRAGMENT_PROGRAM_ARB
) {
1700 /* The clamping to [0,1] can be done for free in the fragment
1701 * shader with a saturate if floats are being used as boolean values.
1703 add
->saturate
= true;
1705 /* Negating the result of the addition gives values on the range
1706 * [-2, 0]. Zero stays zero, and negative values become 1.0. This
1707 * is achieved using SLT.
1709 st_src_reg slt_src
= result_src
;
1710 slt_src
.negate
= ~slt_src
.negate
;
1711 emit(ir
, TGSI_OPCODE_SLT
, result_dst
, slt_src
, st_src_reg_for_float(0.0));
1717 case ir_binop_logic_and
:
1718 /* If native integers are disabled, the bool args are stored as float 0.0
1719 * or 1.0, so "mul" gives us "and". If they're enabled, just use the
1720 * actual AND opcode.
1722 if (native_integers
)
1723 emit(ir
, TGSI_OPCODE_AND
, result_dst
, op
[0], op
[1]);
1725 emit(ir
, TGSI_OPCODE_MUL
, result_dst
, op
[0], op
[1]);
1729 assert(ir
->operands
[0]->type
->is_vector());
1730 assert(ir
->operands
[0]->type
== ir
->operands
[1]->type
);
1731 emit_dp(ir
, result_dst
, op
[0], op
[1],
1732 ir
->operands
[0]->type
->vector_elements
);
1736 /* sqrt(x) = x * rsq(x). */
1737 emit_scalar(ir
, TGSI_OPCODE_RSQ
, result_dst
, op
[0]);
1738 emit(ir
, TGSI_OPCODE_MUL
, result_dst
, result_src
, op
[0]);
1739 /* For incoming channels <= 0, set the result to 0. */
1740 op
[0].negate
= ~op
[0].negate
;
1741 emit(ir
, TGSI_OPCODE_CMP
, result_dst
,
1742 op
[0], result_src
, st_src_reg_for_float(0.0));
1745 emit_scalar(ir
, TGSI_OPCODE_RSQ
, result_dst
, op
[0]);
1748 if (native_integers
) {
1749 emit(ir
, TGSI_OPCODE_I2F
, result_dst
, op
[0]);
1752 /* fallthrough to next case otherwise */
1754 if (native_integers
) {
1755 emit(ir
, TGSI_OPCODE_AND
, result_dst
, op
[0], st_src_reg_for_float(1.0));
1758 /* fallthrough to next case otherwise */
1761 /* Converting between signed and unsigned integers is a no-op. */
1765 if (native_integers
) {
1766 /* Booleans are stored as integers using ~0 for true and 0 for false.
1767 * GLSL requires that int(bool) return 1 for true and 0 for false.
1768 * This conversion is done with AND, but it could be done with NEG.
1770 emit(ir
, TGSI_OPCODE_AND
, result_dst
, op
[0], st_src_reg_for_int(1));
1772 /* Booleans and integers are both stored as floats when native
1773 * integers are disabled.
1779 if (native_integers
)
1780 emit(ir
, TGSI_OPCODE_F2I
, result_dst
, op
[0]);
1782 emit(ir
, TGSI_OPCODE_TRUNC
, result_dst
, op
[0]);
1785 if (native_integers
)
1786 emit(ir
, TGSI_OPCODE_F2U
, result_dst
, op
[0]);
1788 emit(ir
, TGSI_OPCODE_TRUNC
, result_dst
, op
[0]);
1790 case ir_unop_bitcast_f2i
:
1791 case ir_unop_bitcast_f2u
:
1792 case ir_unop_bitcast_i2f
:
1793 case ir_unop_bitcast_u2f
:
1797 emit(ir
, TGSI_OPCODE_SNE
, result_dst
, op
[0], st_src_reg_for_float(0.0));
1800 if (native_integers
)
1801 emit(ir
, TGSI_OPCODE_INEG
, result_dst
, op
[0]);
1803 emit(ir
, TGSI_OPCODE_SNE
, result_dst
, op
[0], st_src_reg_for_float(0.0));
1806 emit(ir
, TGSI_OPCODE_TRUNC
, result_dst
, op
[0]);
1809 emit(ir
, TGSI_OPCODE_CEIL
, result_dst
, op
[0]);
1812 emit(ir
, TGSI_OPCODE_FLR
, result_dst
, op
[0]);
1814 case ir_unop_round_even
:
1815 emit(ir
, TGSI_OPCODE_ROUND
, result_dst
, op
[0]);
1818 emit(ir
, TGSI_OPCODE_FRC
, result_dst
, op
[0]);
1822 emit(ir
, TGSI_OPCODE_MIN
, result_dst
, op
[0], op
[1]);
1825 emit(ir
, TGSI_OPCODE_MAX
, result_dst
, op
[0], op
[1]);
1828 emit_scalar(ir
, TGSI_OPCODE_POW
, result_dst
, op
[0], op
[1]);
1831 case ir_unop_bit_not
:
1832 if (native_integers
) {
1833 emit(ir
, TGSI_OPCODE_NOT
, result_dst
, op
[0]);
1837 if (native_integers
) {
1838 emit(ir
, TGSI_OPCODE_U2F
, result_dst
, op
[0]);
1841 case ir_binop_lshift
:
1842 if (native_integers
) {
1843 emit(ir
, TGSI_OPCODE_SHL
, result_dst
, op
[0], op
[1]);
1846 case ir_binop_rshift
:
1847 if (native_integers
) {
1848 emit(ir
, TGSI_OPCODE_ISHR
, result_dst
, op
[0], op
[1]);
1851 case ir_binop_bit_and
:
1852 if (native_integers
) {
1853 emit(ir
, TGSI_OPCODE_AND
, result_dst
, op
[0], op
[1]);
1856 case ir_binop_bit_xor
:
1857 if (native_integers
) {
1858 emit(ir
, TGSI_OPCODE_XOR
, result_dst
, op
[0], op
[1]);
1861 case ir_binop_bit_or
:
1862 if (native_integers
) {
1863 emit(ir
, TGSI_OPCODE_OR
, result_dst
, op
[0], op
[1]);
1867 assert(!"GLSL 1.30 features unsupported");
1870 case ir_quadop_vector
:
1871 /* This operation should have already been handled.
1873 assert(!"Should not get here.");
1877 this->result
= result_src
;
1882 glsl_to_tgsi_visitor::visit(ir_swizzle
*ir
)
1888 /* Note that this is only swizzles in expressions, not those on the left
1889 * hand side of an assignment, which do write masking. See ir_assignment
1893 ir
->val
->accept(this);
1895 assert(src
.file
!= PROGRAM_UNDEFINED
);
1897 for (i
= 0; i
< 4; i
++) {
1898 if (i
< ir
->type
->vector_elements
) {
1901 swizzle
[i
] = GET_SWZ(src
.swizzle
, ir
->mask
.x
);
1904 swizzle
[i
] = GET_SWZ(src
.swizzle
, ir
->mask
.y
);
1907 swizzle
[i
] = GET_SWZ(src
.swizzle
, ir
->mask
.z
);
1910 swizzle
[i
] = GET_SWZ(src
.swizzle
, ir
->mask
.w
);
1914 /* If the type is smaller than a vec4, replicate the last
1917 swizzle
[i
] = swizzle
[ir
->type
->vector_elements
- 1];
1921 src
.swizzle
= MAKE_SWIZZLE4(swizzle
[0], swizzle
[1], swizzle
[2], swizzle
[3]);
1927 glsl_to_tgsi_visitor::visit(ir_dereference_variable
*ir
)
1929 variable_storage
*entry
= find_variable_storage(ir
->var
);
1930 ir_variable
*var
= ir
->var
;
1933 switch (var
->mode
) {
1934 case ir_var_uniform
:
1935 entry
= new(mem_ctx
) variable_storage(var
, PROGRAM_UNIFORM
,
1937 this->variables
.push_tail(entry
);
1941 /* The linker assigns locations for varyings and attributes,
1942 * including deprecated builtins (like gl_Color), user-assign
1943 * generic attributes (glBindVertexLocation), and
1944 * user-defined varyings.
1946 * FINISHME: We would hit this path for function arguments. Fix!
1948 assert(var
->location
!= -1);
1949 entry
= new(mem_ctx
) variable_storage(var
,
1954 assert(var
->location
!= -1);
1955 entry
= new(mem_ctx
) variable_storage(var
,
1957 var
->location
+ var
->index
);
1959 case ir_var_system_value
:
1960 entry
= new(mem_ctx
) variable_storage(var
,
1961 PROGRAM_SYSTEM_VALUE
,
1965 case ir_var_temporary
:
1966 entry
= new(mem_ctx
) variable_storage(var
, PROGRAM_TEMPORARY
,
1968 this->variables
.push_tail(entry
);
1970 next_temp
+= type_size(var
->type
);
1975 printf("Failed to make storage for %s\n", var
->name
);
1980 this->result
= st_src_reg(entry
->file
, entry
->index
, var
->type
);
1981 if (!native_integers
)
1982 this->result
.type
= GLSL_TYPE_FLOAT
;
1986 glsl_to_tgsi_visitor::visit(ir_dereference_array
*ir
)
1990 int element_size
= type_size(ir
->type
);
1992 index
= ir
->array_index
->constant_expression_value();
1994 ir
->array
->accept(this);
1998 src
.index
+= index
->value
.i
[0] * element_size
;
2000 /* Variable index array dereference. It eats the "vec4" of the
2001 * base of the array and an index that offsets the TGSI register
2004 ir
->array_index
->accept(this);
2006 st_src_reg index_reg
;
2008 if (element_size
== 1) {
2009 index_reg
= this->result
;
2011 index_reg
= get_temp(native_integers
?
2012 glsl_type::int_type
: glsl_type::float_type
);
2014 emit(ir
, TGSI_OPCODE_MUL
, st_dst_reg(index_reg
),
2015 this->result
, st_src_reg_for_type(index_reg
.type
, element_size
));
2018 /* If there was already a relative address register involved, add the
2019 * new and the old together to get the new offset.
2021 if (src
.reladdr
!= NULL
) {
2022 st_src_reg accum_reg
= get_temp(native_integers
?
2023 glsl_type::int_type
: glsl_type::float_type
);
2025 emit(ir
, TGSI_OPCODE_ADD
, st_dst_reg(accum_reg
),
2026 index_reg
, *src
.reladdr
);
2028 index_reg
= accum_reg
;
2031 src
.reladdr
= ralloc(mem_ctx
, st_src_reg
);
2032 memcpy(src
.reladdr
, &index_reg
, sizeof(index_reg
));
2035 /* If the type is smaller than a vec4, replicate the last channel out. */
2036 if (ir
->type
->is_scalar() || ir
->type
->is_vector())
2037 src
.swizzle
= swizzle_for_size(ir
->type
->vector_elements
);
2039 src
.swizzle
= SWIZZLE_NOOP
;
2045 glsl_to_tgsi_visitor::visit(ir_dereference_record
*ir
)
2048 const glsl_type
*struct_type
= ir
->record
->type
;
2051 ir
->record
->accept(this);
2053 for (i
= 0; i
< struct_type
->length
; i
++) {
2054 if (strcmp(struct_type
->fields
.structure
[i
].name
, ir
->field
) == 0)
2056 offset
+= type_size(struct_type
->fields
.structure
[i
].type
);
2059 /* If the type is smaller than a vec4, replicate the last channel out. */
2060 if (ir
->type
->is_scalar() || ir
->type
->is_vector())
2061 this->result
.swizzle
= swizzle_for_size(ir
->type
->vector_elements
);
2063 this->result
.swizzle
= SWIZZLE_NOOP
;
2065 this->result
.index
+= offset
;
2069 * We want to be careful in assignment setup to hit the actual storage
2070 * instead of potentially using a temporary like we might with the
2071 * ir_dereference handler.
2074 get_assignment_lhs(ir_dereference
*ir
, glsl_to_tgsi_visitor
*v
)
2076 /* The LHS must be a dereference. If the LHS is a variable indexed array
2077 * access of a vector, it must be separated into a series conditional moves
2078 * before reaching this point (see ir_vec_index_to_cond_assign).
2080 assert(ir
->as_dereference());
2081 ir_dereference_array
*deref_array
= ir
->as_dereference_array();
2083 assert(!deref_array
->array
->type
->is_vector());
2086 /* Use the rvalue deref handler for the most part. We'll ignore
2087 * swizzles in it and write swizzles using writemask, though.
2090 return st_dst_reg(v
->result
);
2094 * Process the condition of a conditional assignment
2096 * Examines the condition of a conditional assignment to generate the optimal
2097 * first operand of a \c CMP instruction. If the condition is a relational
2098 * operator with 0 (e.g., \c ir_binop_less), the value being compared will be
2099 * used as the source for the \c CMP instruction. Otherwise the comparison
2100 * is processed to a boolean result, and the boolean result is used as the
2101 * operand to the CMP instruction.
2104 glsl_to_tgsi_visitor::process_move_condition(ir_rvalue
*ir
)
2106 ir_rvalue
*src_ir
= ir
;
2108 bool switch_order
= false;
2110 ir_expression
*const expr
= ir
->as_expression();
2111 if ((expr
!= NULL
) && (expr
->get_num_operands() == 2)) {
2112 bool zero_on_left
= false;
2114 if (expr
->operands
[0]->is_zero()) {
2115 src_ir
= expr
->operands
[1];
2116 zero_on_left
= true;
2117 } else if (expr
->operands
[1]->is_zero()) {
2118 src_ir
= expr
->operands
[0];
2119 zero_on_left
= false;
2123 * (a < 0) T F F ( a < 0) T F F
2124 * (0 < a) F F T (-a < 0) F F T
2125 * (a <= 0) T T F (-a < 0) F F T (swap order of other operands)
2126 * (0 <= a) F T T ( a < 0) T F F (swap order of other operands)
2127 * (a > 0) F F T (-a < 0) F F T
2128 * (0 > a) T F F ( a < 0) T F F
2129 * (a >= 0) F T T ( a < 0) T F F (swap order of other operands)
2130 * (0 >= a) T T F (-a < 0) F F T (swap order of other operands)
2132 * Note that exchanging the order of 0 and 'a' in the comparison simply
2133 * means that the value of 'a' should be negated.
2136 switch (expr
->operation
) {
2138 switch_order
= false;
2139 negate
= zero_on_left
;
2142 case ir_binop_greater
:
2143 switch_order
= false;
2144 negate
= !zero_on_left
;
2147 case ir_binop_lequal
:
2148 switch_order
= true;
2149 negate
= !zero_on_left
;
2152 case ir_binop_gequal
:
2153 switch_order
= true;
2154 negate
= zero_on_left
;
2158 /* This isn't the right kind of comparison afterall, so make sure
2159 * the whole condition is visited.
2167 src_ir
->accept(this);
2169 /* We use the TGSI_OPCODE_CMP (a < 0 ? b : c) for conditional moves, and the
2170 * condition we produced is 0.0 or 1.0. By flipping the sign, we can
2171 * choose which value TGSI_OPCODE_CMP produces without an extra instruction
2172 * computing the condition.
2175 this->result
.negate
= ~this->result
.negate
;
2177 return switch_order
;
2181 glsl_to_tgsi_visitor::visit(ir_assignment
*ir
)
2187 ir
->rhs
->accept(this);
2190 l
= get_assignment_lhs(ir
->lhs
, this);
2192 /* FINISHME: This should really set to the correct maximal writemask for each
2193 * FINISHME: component written (in the loops below). This case can only
2194 * FINISHME: occur for matrices, arrays, and structures.
2196 if (ir
->write_mask
== 0) {
2197 assert(!ir
->lhs
->type
->is_scalar() && !ir
->lhs
->type
->is_vector());
2198 l
.writemask
= WRITEMASK_XYZW
;
2199 } else if (ir
->lhs
->type
->is_scalar() &&
2200 ir
->lhs
->variable_referenced()->mode
== ir_var_out
) {
2201 /* FINISHME: This hack makes writing to gl_FragDepth, which lives in the
2202 * FINISHME: W component of fragment shader output zero, work correctly.
2204 l
.writemask
= WRITEMASK_XYZW
;
2207 int first_enabled_chan
= 0;
2210 l
.writemask
= ir
->write_mask
;
2212 for (int i
= 0; i
< 4; i
++) {
2213 if (l
.writemask
& (1 << i
)) {
2214 first_enabled_chan
= GET_SWZ(r
.swizzle
, i
);
2219 /* Swizzle a small RHS vector into the channels being written.
2221 * glsl ir treats write_mask as dictating how many channels are
2222 * present on the RHS while TGSI treats write_mask as just
2223 * showing which channels of the vec4 RHS get written.
2225 for (int i
= 0; i
< 4; i
++) {
2226 if (l
.writemask
& (1 << i
))
2227 swizzles
[i
] = GET_SWZ(r
.swizzle
, rhs_chan
++);
2229 swizzles
[i
] = first_enabled_chan
;
2231 r
.swizzle
= MAKE_SWIZZLE4(swizzles
[0], swizzles
[1],
2232 swizzles
[2], swizzles
[3]);
2235 assert(l
.file
!= PROGRAM_UNDEFINED
);
2236 assert(r
.file
!= PROGRAM_UNDEFINED
);
2238 if (ir
->condition
) {
2239 const bool switch_order
= this->process_move_condition(ir
->condition
);
2240 st_src_reg condition
= this->result
;
2242 for (i
= 0; i
< type_size(ir
->lhs
->type
); i
++) {
2243 st_src_reg l_src
= st_src_reg(l
);
2244 st_src_reg condition_temp
= condition
;
2245 l_src
.swizzle
= swizzle_for_size(ir
->lhs
->type
->vector_elements
);
2247 if (native_integers
) {
2248 /* This is necessary because TGSI's CMP instruction expects the
2249 * condition to be a float, and we store booleans as integers.
2250 * If TGSI had a UCMP instruction or similar, this extra
2251 * instruction would not be necessary.
2253 condition_temp
= get_temp(glsl_type::vec4_type
);
2254 condition
.negate
= 0;
2255 emit(ir
, TGSI_OPCODE_I2F
, st_dst_reg(condition_temp
), condition
);
2256 condition_temp
.swizzle
= condition
.swizzle
;
2260 emit(ir
, TGSI_OPCODE_CMP
, l
, condition_temp
, l_src
, r
);
2262 emit(ir
, TGSI_OPCODE_CMP
, l
, condition_temp
, r
, l_src
);
2268 } else if (ir
->rhs
->as_expression() &&
2269 this->instructions
.get_tail() &&
2270 ir
->rhs
== ((glsl_to_tgsi_instruction
*)this->instructions
.get_tail())->ir
&&
2271 type_size(ir
->lhs
->type
) == 1 &&
2272 l
.writemask
== ((glsl_to_tgsi_instruction
*)this->instructions
.get_tail())->dst
.writemask
) {
2273 /* To avoid emitting an extra MOV when assigning an expression to a
2274 * variable, emit the last instruction of the expression again, but
2275 * replace the destination register with the target of the assignment.
2276 * Dead code elimination will remove the original instruction.
2278 glsl_to_tgsi_instruction
*inst
, *new_inst
;
2279 inst
= (glsl_to_tgsi_instruction
*)this->instructions
.get_tail();
2280 new_inst
= emit(ir
, inst
->op
, l
, inst
->src
[0], inst
->src
[1], inst
->src
[2]);
2281 new_inst
->saturate
= inst
->saturate
;
2282 inst
->dead_mask
= inst
->dst
.writemask
;
2284 for (i
= 0; i
< type_size(ir
->lhs
->type
); i
++) {
2285 emit(ir
, TGSI_OPCODE_MOV
, l
, r
);
2294 glsl_to_tgsi_visitor::visit(ir_constant
*ir
)
2297 GLfloat stack_vals
[4] = { 0 };
2298 gl_constant_value
*values
= (gl_constant_value
*) stack_vals
;
2299 GLenum gl_type
= GL_NONE
;
2301 static int in_array
= 0;
2302 gl_register_file file
= in_array
? PROGRAM_CONSTANT
: PROGRAM_IMMEDIATE
;
2304 /* Unfortunately, 4 floats is all we can get into
2305 * _mesa_add_typed_unnamed_constant. So, make a temp to store an
2306 * aggregate constant and move each constant value into it. If we
2307 * get lucky, copy propagation will eliminate the extra moves.
2309 if (ir
->type
->base_type
== GLSL_TYPE_STRUCT
) {
2310 st_src_reg temp_base
= get_temp(ir
->type
);
2311 st_dst_reg temp
= st_dst_reg(temp_base
);
2313 foreach_iter(exec_list_iterator
, iter
, ir
->components
) {
2314 ir_constant
*field_value
= (ir_constant
*)iter
.get();
2315 int size
= type_size(field_value
->type
);
2319 field_value
->accept(this);
2322 for (i
= 0; i
< (unsigned int)size
; i
++) {
2323 emit(ir
, TGSI_OPCODE_MOV
, temp
, src
);
2329 this->result
= temp_base
;
2333 if (ir
->type
->is_array()) {
2334 st_src_reg temp_base
= get_temp(ir
->type
);
2335 st_dst_reg temp
= st_dst_reg(temp_base
);
2336 int size
= type_size(ir
->type
->fields
.array
);
2341 for (i
= 0; i
< ir
->type
->length
; i
++) {
2342 ir
->array_elements
[i
]->accept(this);
2344 for (int j
= 0; j
< size
; j
++) {
2345 emit(ir
, TGSI_OPCODE_MOV
, temp
, src
);
2351 this->result
= temp_base
;
2356 if (ir
->type
->is_matrix()) {
2357 st_src_reg mat
= get_temp(ir
->type
);
2358 st_dst_reg mat_column
= st_dst_reg(mat
);
2360 for (i
= 0; i
< ir
->type
->matrix_columns
; i
++) {
2361 assert(ir
->type
->base_type
== GLSL_TYPE_FLOAT
);
2362 values
= (gl_constant_value
*) &ir
->value
.f
[i
* ir
->type
->vector_elements
];
2364 src
= st_src_reg(file
, -1, ir
->type
->base_type
);
2365 src
.index
= add_constant(file
,
2367 ir
->type
->vector_elements
,
2370 emit(ir
, TGSI_OPCODE_MOV
, mat_column
, src
);
2379 switch (ir
->type
->base_type
) {
2380 case GLSL_TYPE_FLOAT
:
2382 for (i
= 0; i
< ir
->type
->vector_elements
; i
++) {
2383 values
[i
].f
= ir
->value
.f
[i
];
2386 case GLSL_TYPE_UINT
:
2387 gl_type
= native_integers
? GL_UNSIGNED_INT
: GL_FLOAT
;
2388 for (i
= 0; i
< ir
->type
->vector_elements
; i
++) {
2389 if (native_integers
)
2390 values
[i
].u
= ir
->value
.u
[i
];
2392 values
[i
].f
= ir
->value
.u
[i
];
2396 gl_type
= native_integers
? GL_INT
: GL_FLOAT
;
2397 for (i
= 0; i
< ir
->type
->vector_elements
; i
++) {
2398 if (native_integers
)
2399 values
[i
].i
= ir
->value
.i
[i
];
2401 values
[i
].f
= ir
->value
.i
[i
];
2404 case GLSL_TYPE_BOOL
:
2405 gl_type
= native_integers
? GL_BOOL
: GL_FLOAT
;
2406 for (i
= 0; i
< ir
->type
->vector_elements
; i
++) {
2407 if (native_integers
)
2408 values
[i
].u
= ir
->value
.b
[i
] ? ~0 : 0;
2410 values
[i
].f
= ir
->value
.b
[i
];
2414 assert(!"Non-float/uint/int/bool constant");
2417 this->result
= st_src_reg(file
, -1, ir
->type
);
2418 this->result
.index
= add_constant(file
,
2420 ir
->type
->vector_elements
,
2422 &this->result
.swizzle
);
2426 glsl_to_tgsi_visitor::get_function_signature(ir_function_signature
*sig
)
2428 function_entry
*entry
;
2430 foreach_iter(exec_list_iterator
, iter
, this->function_signatures
) {
2431 entry
= (function_entry
*)iter
.get();
2433 if (entry
->sig
== sig
)
2437 entry
= ralloc(mem_ctx
, function_entry
);
2439 entry
->sig_id
= this->next_signature_id
++;
2440 entry
->bgn_inst
= NULL
;
2442 /* Allocate storage for all the parameters. */
2443 foreach_iter(exec_list_iterator
, iter
, sig
->parameters
) {
2444 ir_variable
*param
= (ir_variable
*)iter
.get();
2445 variable_storage
*storage
;
2447 storage
= find_variable_storage(param
);
2450 storage
= new(mem_ctx
) variable_storage(param
, PROGRAM_TEMPORARY
,
2452 this->variables
.push_tail(storage
);
2454 this->next_temp
+= type_size(param
->type
);
2457 if (!sig
->return_type
->is_void()) {
2458 entry
->return_reg
= get_temp(sig
->return_type
);
2460 entry
->return_reg
= undef_src
;
2463 this->function_signatures
.push_tail(entry
);
2468 glsl_to_tgsi_visitor::visit(ir_call
*ir
)
2470 glsl_to_tgsi_instruction
*call_inst
;
2471 ir_function_signature
*sig
= ir
->callee
;
2472 function_entry
*entry
= get_function_signature(sig
);
2475 /* Process in parameters. */
2476 exec_list_iterator sig_iter
= sig
->parameters
.iterator();
2477 foreach_iter(exec_list_iterator
, iter
, *ir
) {
2478 ir_rvalue
*param_rval
= (ir_rvalue
*)iter
.get();
2479 ir_variable
*param
= (ir_variable
*)sig_iter
.get();
2481 if (param
->mode
== ir_var_in
||
2482 param
->mode
== ir_var_inout
) {
2483 variable_storage
*storage
= find_variable_storage(param
);
2486 param_rval
->accept(this);
2487 st_src_reg r
= this->result
;
2490 l
.file
= storage
->file
;
2491 l
.index
= storage
->index
;
2493 l
.writemask
= WRITEMASK_XYZW
;
2494 l
.cond_mask
= COND_TR
;
2496 for (i
= 0; i
< type_size(param
->type
); i
++) {
2497 emit(ir
, TGSI_OPCODE_MOV
, l
, r
);
2505 assert(!sig_iter
.has_next());
2507 /* Emit call instruction */
2508 call_inst
= emit(ir
, TGSI_OPCODE_CAL
);
2509 call_inst
->function
= entry
;
2511 /* Process out parameters. */
2512 sig_iter
= sig
->parameters
.iterator();
2513 foreach_iter(exec_list_iterator
, iter
, *ir
) {
2514 ir_rvalue
*param_rval
= (ir_rvalue
*)iter
.get();
2515 ir_variable
*param
= (ir_variable
*)sig_iter
.get();
2517 if (param
->mode
== ir_var_out
||
2518 param
->mode
== ir_var_inout
) {
2519 variable_storage
*storage
= find_variable_storage(param
);
2523 r
.file
= storage
->file
;
2524 r
.index
= storage
->index
;
2526 r
.swizzle
= SWIZZLE_NOOP
;
2529 param_rval
->accept(this);
2530 st_dst_reg l
= st_dst_reg(this->result
);
2532 for (i
= 0; i
< type_size(param
->type
); i
++) {
2533 emit(ir
, TGSI_OPCODE_MOV
, l
, r
);
2541 assert(!sig_iter
.has_next());
2543 /* Process return value. */
2544 this->result
= entry
->return_reg
;
2548 glsl_to_tgsi_visitor::visit(ir_texture
*ir
)
2550 st_src_reg result_src
, coord
, lod_info
, projector
, dx
, dy
, offset
;
2551 st_dst_reg result_dst
, coord_dst
;
2552 glsl_to_tgsi_instruction
*inst
= NULL
;
2553 unsigned opcode
= TGSI_OPCODE_NOP
;
2555 if (ir
->coordinate
) {
2556 ir
->coordinate
->accept(this);
2558 /* Put our coords in a temp. We'll need to modify them for shadow,
2559 * projection, or LOD, so the only case we'd use it as is is if
2560 * we're doing plain old texturing. The optimization passes on
2561 * glsl_to_tgsi_visitor should handle cleaning up our mess in that case.
2563 coord
= get_temp(glsl_type::vec4_type
);
2564 coord_dst
= st_dst_reg(coord
);
2565 emit(ir
, TGSI_OPCODE_MOV
, coord_dst
, this->result
);
2568 if (ir
->projector
) {
2569 ir
->projector
->accept(this);
2570 projector
= this->result
;
2573 /* Storage for our result. Ideally for an assignment we'd be using
2574 * the actual storage for the result here, instead.
2576 result_src
= get_temp(glsl_type::vec4_type
);
2577 result_dst
= st_dst_reg(result_src
);
2581 opcode
= TGSI_OPCODE_TEX
;
2584 opcode
= TGSI_OPCODE_TXB
;
2585 ir
->lod_info
.bias
->accept(this);
2586 lod_info
= this->result
;
2589 opcode
= TGSI_OPCODE_TXL
;
2590 ir
->lod_info
.lod
->accept(this);
2591 lod_info
= this->result
;
2594 opcode
= TGSI_OPCODE_TXD
;
2595 ir
->lod_info
.grad
.dPdx
->accept(this);
2597 ir
->lod_info
.grad
.dPdy
->accept(this);
2601 opcode
= TGSI_OPCODE_TXQ
;
2602 ir
->lod_info
.lod
->accept(this);
2603 lod_info
= this->result
;
2606 opcode
= TGSI_OPCODE_TXF
;
2607 ir
->lod_info
.lod
->accept(this);
2608 lod_info
= this->result
;
2610 ir
->offset
->accept(this);
2611 offset
= this->result
;
2616 const glsl_type
*sampler_type
= ir
->sampler
->type
;
2618 if (ir
->projector
) {
2619 if (opcode
== TGSI_OPCODE_TEX
) {
2620 /* Slot the projector in as the last component of the coord. */
2621 coord_dst
.writemask
= WRITEMASK_W
;
2622 emit(ir
, TGSI_OPCODE_MOV
, coord_dst
, projector
);
2623 coord_dst
.writemask
= WRITEMASK_XYZW
;
2624 opcode
= TGSI_OPCODE_TXP
;
2626 st_src_reg coord_w
= coord
;
2627 coord_w
.swizzle
= SWIZZLE_WWWW
;
2629 /* For the other TEX opcodes there's no projective version
2630 * since the last slot is taken up by LOD info. Do the
2631 * projective divide now.
2633 coord_dst
.writemask
= WRITEMASK_W
;
2634 emit(ir
, TGSI_OPCODE_RCP
, coord_dst
, projector
);
2636 /* In the case where we have to project the coordinates "by hand,"
2637 * the shadow comparator value must also be projected.
2639 st_src_reg tmp_src
= coord
;
2640 if (ir
->shadow_comparitor
) {
2641 /* Slot the shadow value in as the second to last component of the
2644 ir
->shadow_comparitor
->accept(this);
2646 tmp_src
= get_temp(glsl_type::vec4_type
);
2647 st_dst_reg tmp_dst
= st_dst_reg(tmp_src
);
2649 /* Projective division not allowed for array samplers. */
2650 assert(!sampler_type
->sampler_array
);
2652 tmp_dst
.writemask
= WRITEMASK_Z
;
2653 emit(ir
, TGSI_OPCODE_MOV
, tmp_dst
, this->result
);
2655 tmp_dst
.writemask
= WRITEMASK_XY
;
2656 emit(ir
, TGSI_OPCODE_MOV
, tmp_dst
, coord
);
2659 coord_dst
.writemask
= WRITEMASK_XYZ
;
2660 emit(ir
, TGSI_OPCODE_MUL
, coord_dst
, tmp_src
, coord_w
);
2662 coord_dst
.writemask
= WRITEMASK_XYZW
;
2663 coord
.swizzle
= SWIZZLE_XYZW
;
2667 /* If projection is done and the opcode is not TGSI_OPCODE_TXP, then the shadow
2668 * comparator was put in the correct place (and projected) by the code,
2669 * above, that handles by-hand projection.
2671 if (ir
->shadow_comparitor
&& (!ir
->projector
|| opcode
== TGSI_OPCODE_TXP
)) {
2672 /* Slot the shadow value in as the second to last component of the
2675 ir
->shadow_comparitor
->accept(this);
2677 /* XXX This will need to be updated for cubemap array samplers. */
2678 if ((sampler_type
->sampler_dimensionality
== GLSL_SAMPLER_DIM_2D
&&
2679 sampler_type
->sampler_array
) ||
2680 sampler_type
->sampler_dimensionality
== GLSL_SAMPLER_DIM_CUBE
) {
2681 coord_dst
.writemask
= WRITEMASK_W
;
2683 coord_dst
.writemask
= WRITEMASK_Z
;
2686 emit(ir
, TGSI_OPCODE_MOV
, coord_dst
, this->result
);
2687 coord_dst
.writemask
= WRITEMASK_XYZW
;
2690 if (opcode
== TGSI_OPCODE_TXL
|| opcode
== TGSI_OPCODE_TXB
||
2691 opcode
== TGSI_OPCODE_TXF
) {
2692 /* TGSI stores LOD or LOD bias in the last channel of the coords. */
2693 coord_dst
.writemask
= WRITEMASK_W
;
2694 emit(ir
, TGSI_OPCODE_MOV
, coord_dst
, lod_info
);
2695 coord_dst
.writemask
= WRITEMASK_XYZW
;
2698 if (opcode
== TGSI_OPCODE_TXD
)
2699 inst
= emit(ir
, opcode
, result_dst
, coord
, dx
, dy
);
2700 else if (opcode
== TGSI_OPCODE_TXQ
)
2701 inst
= emit(ir
, opcode
, result_dst
, lod_info
);
2702 else if (opcode
== TGSI_OPCODE_TXF
) {
2703 inst
= emit(ir
, opcode
, result_dst
, coord
);
2705 inst
= emit(ir
, opcode
, result_dst
, coord
);
2707 if (ir
->shadow_comparitor
)
2708 inst
->tex_shadow
= GL_TRUE
;
2710 inst
->sampler
= _mesa_get_sampler_uniform_value(ir
->sampler
,
2711 this->shader_program
,
2715 inst
->tex_offset_num_offset
= 1;
2716 inst
->tex_offsets
[0].Index
= offset
.index
;
2717 inst
->tex_offsets
[0].File
= offset
.file
;
2718 inst
->tex_offsets
[0].SwizzleX
= GET_SWZ(offset
.swizzle
, 0);
2719 inst
->tex_offsets
[0].SwizzleY
= GET_SWZ(offset
.swizzle
, 1);
2720 inst
->tex_offsets
[0].SwizzleZ
= GET_SWZ(offset
.swizzle
, 2);
2723 switch (sampler_type
->sampler_dimensionality
) {
2724 case GLSL_SAMPLER_DIM_1D
:
2725 inst
->tex_target
= (sampler_type
->sampler_array
)
2726 ? TEXTURE_1D_ARRAY_INDEX
: TEXTURE_1D_INDEX
;
2728 case GLSL_SAMPLER_DIM_2D
:
2729 inst
->tex_target
= (sampler_type
->sampler_array
)
2730 ? TEXTURE_2D_ARRAY_INDEX
: TEXTURE_2D_INDEX
;
2732 case GLSL_SAMPLER_DIM_3D
:
2733 inst
->tex_target
= TEXTURE_3D_INDEX
;
2735 case GLSL_SAMPLER_DIM_CUBE
:
2736 inst
->tex_target
= TEXTURE_CUBE_INDEX
;
2738 case GLSL_SAMPLER_DIM_RECT
:
2739 inst
->tex_target
= TEXTURE_RECT_INDEX
;
2741 case GLSL_SAMPLER_DIM_BUF
:
2742 assert(!"FINISHME: Implement ARB_texture_buffer_object");
2744 case GLSL_SAMPLER_DIM_EXTERNAL
:
2745 inst
->tex_target
= TEXTURE_EXTERNAL_INDEX
;
2748 assert(!"Should not get here.");
2751 this->result
= result_src
;
2755 glsl_to_tgsi_visitor::visit(ir_return
*ir
)
2757 if (ir
->get_value()) {
2761 assert(current_function
);
2763 ir
->get_value()->accept(this);
2764 st_src_reg r
= this->result
;
2766 l
= st_dst_reg(current_function
->return_reg
);
2768 for (i
= 0; i
< type_size(current_function
->sig
->return_type
); i
++) {
2769 emit(ir
, TGSI_OPCODE_MOV
, l
, r
);
2775 emit(ir
, TGSI_OPCODE_RET
);
2779 glsl_to_tgsi_visitor::visit(ir_discard
*ir
)
2781 if (ir
->condition
) {
2782 ir
->condition
->accept(this);
2783 this->result
.negate
= ~this->result
.negate
;
2784 emit(ir
, TGSI_OPCODE_KIL
, undef_dst
, this->result
);
2786 emit(ir
, TGSI_OPCODE_KILP
);
2791 glsl_to_tgsi_visitor::visit(ir_if
*ir
)
2793 glsl_to_tgsi_instruction
*cond_inst
, *if_inst
;
2794 glsl_to_tgsi_instruction
*prev_inst
;
2796 prev_inst
= (glsl_to_tgsi_instruction
*)this->instructions
.get_tail();
2798 ir
->condition
->accept(this);
2799 assert(this->result
.file
!= PROGRAM_UNDEFINED
);
2801 if (this->options
->EmitCondCodes
) {
2802 cond_inst
= (glsl_to_tgsi_instruction
*)this->instructions
.get_tail();
2804 /* See if we actually generated any instruction for generating
2805 * the condition. If not, then cook up a move to a temp so we
2806 * have something to set cond_update on.
2808 if (cond_inst
== prev_inst
) {
2809 st_src_reg temp
= get_temp(glsl_type::bool_type
);
2810 cond_inst
= emit(ir
->condition
, TGSI_OPCODE_MOV
, st_dst_reg(temp
), result
);
2812 cond_inst
->cond_update
= GL_TRUE
;
2814 if_inst
= emit(ir
->condition
, TGSI_OPCODE_IF
);
2815 if_inst
->dst
.cond_mask
= COND_NE
;
2817 if_inst
= emit(ir
->condition
, TGSI_OPCODE_IF
, undef_dst
, this->result
);
2820 this->instructions
.push_tail(if_inst
);
2822 visit_exec_list(&ir
->then_instructions
, this);
2824 if (!ir
->else_instructions
.is_empty()) {
2825 emit(ir
->condition
, TGSI_OPCODE_ELSE
);
2826 visit_exec_list(&ir
->else_instructions
, this);
2829 if_inst
= emit(ir
->condition
, TGSI_OPCODE_ENDIF
);
2832 glsl_to_tgsi_visitor::glsl_to_tgsi_visitor()
2834 result
.file
= PROGRAM_UNDEFINED
;
2836 next_signature_id
= 1;
2838 current_function
= NULL
;
2839 num_address_regs
= 0;
2841 indirect_addr_temps
= false;
2842 indirect_addr_consts
= false;
2844 native_integers
= false;
2845 mem_ctx
= ralloc_context(NULL
);
2848 shader_program
= NULL
;
2852 glsl_to_tgsi_visitor::~glsl_to_tgsi_visitor()
2854 ralloc_free(mem_ctx
);
2857 extern "C" void free_glsl_to_tgsi_visitor(glsl_to_tgsi_visitor
*v
)
2864 * Count resources used by the given gpu program (number of texture
2868 count_resources(glsl_to_tgsi_visitor
*v
, gl_program
*prog
)
2870 v
->samplers_used
= 0;
2872 foreach_iter(exec_list_iterator
, iter
, v
->instructions
) {
2873 glsl_to_tgsi_instruction
*inst
= (glsl_to_tgsi_instruction
*)iter
.get();
2875 if (is_tex_instruction(inst
->op
)) {
2876 v
->samplers_used
|= 1 << inst
->sampler
;
2878 if (inst
->tex_shadow
) {
2879 prog
->ShadowSamplers
|= 1 << inst
->sampler
;
2884 prog
->SamplersUsed
= v
->samplers_used
;
2886 if (v
->shader_program
!= NULL
)
2887 _mesa_update_shader_textures_used(v
->shader_program
, prog
);
2891 set_uniform_initializer(struct gl_context
*ctx
, void *mem_ctx
,
2892 struct gl_shader_program
*shader_program
,
2893 const char *name
, const glsl_type
*type
,
2896 if (type
->is_record()) {
2897 ir_constant
*field_constant
;
2899 field_constant
= (ir_constant
*)val
->components
.get_head();
2901 for (unsigned int i
= 0; i
< type
->length
; i
++) {
2902 const glsl_type
*field_type
= type
->fields
.structure
[i
].type
;
2903 const char *field_name
= ralloc_asprintf(mem_ctx
, "%s.%s", name
,
2904 type
->fields
.structure
[i
].name
);
2905 set_uniform_initializer(ctx
, mem_ctx
, shader_program
, field_name
,
2906 field_type
, field_constant
);
2907 field_constant
= (ir_constant
*)field_constant
->next
;
2913 unsigned index
= _mesa_get_uniform_location(ctx
, shader_program
, name
,
2915 if (offset
== GL_INVALID_INDEX
) {
2916 fail_link(shader_program
,
2917 "Couldn't find uniform for initializer %s\n", name
);
2920 int loc
= _mesa_uniform_merge_location_offset(index
, offset
);
2922 for (unsigned int i
= 0; i
< (type
->is_array() ? type
->length
: 1); i
++) {
2923 ir_constant
*element
;
2924 const glsl_type
*element_type
;
2925 if (type
->is_array()) {
2926 element
= val
->array_elements
[i
];
2927 element_type
= type
->fields
.array
;
2930 element_type
= type
;
2935 if (element_type
->base_type
== GLSL_TYPE_BOOL
) {
2936 int *conv
= ralloc_array(mem_ctx
, int, element_type
->components());
2937 for (unsigned int j
= 0; j
< element_type
->components(); j
++) {
2938 conv
[j
] = element
->value
.b
[j
];
2940 values
= (void *)conv
;
2941 element_type
= glsl_type::get_instance(GLSL_TYPE_INT
,
2942 element_type
->vector_elements
,
2945 values
= &element
->value
;
2948 if (element_type
->is_matrix()) {
2949 _mesa_uniform_matrix(ctx
, shader_program
,
2950 element_type
->matrix_columns
,
2951 element_type
->vector_elements
,
2952 loc
, 1, GL_FALSE
, (GLfloat
*)values
);
2954 _mesa_uniform(ctx
, shader_program
, loc
, element_type
->matrix_columns
,
2955 values
, element_type
->gl_type
);
2963 * Returns the mask of channels (bitmask of WRITEMASK_X,Y,Z,W) which
2964 * are read from the given src in this instruction
2967 get_src_arg_mask(st_dst_reg dst
, st_src_reg src
)
2969 int read_mask
= 0, comp
;
2971 /* Now, given the src swizzle and the written channels, find which
2972 * components are actually read
2974 for (comp
= 0; comp
< 4; ++comp
) {
2975 const unsigned coord
= GET_SWZ(src
.swizzle
, comp
);
2977 if (dst
.writemask
& (1 << comp
) && coord
<= SWIZZLE_W
)
2978 read_mask
|= 1 << coord
;
2985 * This pass replaces CMP T0, T1 T2 T0 with MOV T0, T2 when the CMP
2986 * instruction is the first instruction to write to register T0. There are
2987 * several lowering passes done in GLSL IR (e.g. branches and
2988 * relative addressing) that create a large number of conditional assignments
2989 * that ir_to_mesa converts to CMP instructions like the one mentioned above.
2991 * Here is why this conversion is safe:
2992 * CMP T0, T1 T2 T0 can be expanded to:
2998 * If (T1 < 0.0) evaluates to true then our replacement MOV T0, T2 is the same
2999 * as the original program. If (T1 < 0.0) evaluates to false, executing
3000 * MOV T0, T0 will store a garbage value in T0 since T0 is uninitialized.
3001 * Therefore, it doesn't matter that we are replacing MOV T0, T0 with MOV T0, T2
3002 * because any instruction that was going to read from T0 after this was going
3003 * to read a garbage value anyway.
3006 glsl_to_tgsi_visitor::simplify_cmp(void)
3008 unsigned *tempWrites
;
3009 unsigned outputWrites
[MAX_PROGRAM_OUTPUTS
];
3011 tempWrites
= new unsigned[MAX_TEMPS
];
3015 memset(tempWrites
, 0, sizeof(unsigned) * MAX_TEMPS
);
3016 memset(outputWrites
, 0, sizeof(outputWrites
));
3018 foreach_iter(exec_list_iterator
, iter
, this->instructions
) {
3019 glsl_to_tgsi_instruction
*inst
= (glsl_to_tgsi_instruction
*)iter
.get();
3020 unsigned prevWriteMask
= 0;
3022 /* Give up if we encounter relative addressing or flow control. */
3023 if (inst
->dst
.reladdr
||
3024 tgsi_get_opcode_info(inst
->op
)->is_branch
||
3025 inst
->op
== TGSI_OPCODE_BGNSUB
||
3026 inst
->op
== TGSI_OPCODE_CONT
||
3027 inst
->op
== TGSI_OPCODE_END
||
3028 inst
->op
== TGSI_OPCODE_ENDSUB
||
3029 inst
->op
== TGSI_OPCODE_RET
) {
3033 if (inst
->dst
.file
== PROGRAM_OUTPUT
) {
3034 assert(inst
->dst
.index
< MAX_PROGRAM_OUTPUTS
);
3035 prevWriteMask
= outputWrites
[inst
->dst
.index
];
3036 outputWrites
[inst
->dst
.index
] |= inst
->dst
.writemask
;
3037 } else if (inst
->dst
.file
== PROGRAM_TEMPORARY
) {
3038 assert(inst
->dst
.index
< MAX_TEMPS
);
3039 prevWriteMask
= tempWrites
[inst
->dst
.index
];
3040 tempWrites
[inst
->dst
.index
] |= inst
->dst
.writemask
;
3043 /* For a CMP to be considered a conditional write, the destination
3044 * register and source register two must be the same. */
3045 if (inst
->op
== TGSI_OPCODE_CMP
3046 && !(inst
->dst
.writemask
& prevWriteMask
)
3047 && inst
->src
[2].file
== inst
->dst
.file
3048 && inst
->src
[2].index
== inst
->dst
.index
3049 && inst
->dst
.writemask
== get_src_arg_mask(inst
->dst
, inst
->src
[2])) {
3051 inst
->op
= TGSI_OPCODE_MOV
;
3052 inst
->src
[0] = inst
->src
[1];
3056 delete [] tempWrites
;
3059 /* Replaces all references to a temporary register index with another index. */
3061 glsl_to_tgsi_visitor::rename_temp_register(int index
, int new_index
)
3063 foreach_iter(exec_list_iterator
, iter
, this->instructions
) {
3064 glsl_to_tgsi_instruction
*inst
= (glsl_to_tgsi_instruction
*)iter
.get();
3067 for (j
=0; j
< num_inst_src_regs(inst
->op
); j
++) {
3068 if (inst
->src
[j
].file
== PROGRAM_TEMPORARY
&&
3069 inst
->src
[j
].index
== index
) {
3070 inst
->src
[j
].index
= new_index
;
3074 if (inst
->dst
.file
== PROGRAM_TEMPORARY
&& inst
->dst
.index
== index
) {
3075 inst
->dst
.index
= new_index
;
3081 glsl_to_tgsi_visitor::get_first_temp_read(int index
)
3083 int depth
= 0; /* loop depth */
3084 int loop_start
= -1; /* index of the first active BGNLOOP (if any) */
3087 foreach_iter(exec_list_iterator
, iter
, this->instructions
) {
3088 glsl_to_tgsi_instruction
*inst
= (glsl_to_tgsi_instruction
*)iter
.get();
3090 for (j
=0; j
< num_inst_src_regs(inst
->op
); j
++) {
3091 if (inst
->src
[j
].file
== PROGRAM_TEMPORARY
&&
3092 inst
->src
[j
].index
== index
) {
3093 return (depth
== 0) ? i
: loop_start
;
3097 if (inst
->op
== TGSI_OPCODE_BGNLOOP
) {
3100 } else if (inst
->op
== TGSI_OPCODE_ENDLOOP
) {
3113 glsl_to_tgsi_visitor::get_first_temp_write(int index
)
3115 int depth
= 0; /* loop depth */
3116 int loop_start
= -1; /* index of the first active BGNLOOP (if any) */
3119 foreach_iter(exec_list_iterator
, iter
, this->instructions
) {
3120 glsl_to_tgsi_instruction
*inst
= (glsl_to_tgsi_instruction
*)iter
.get();
3122 if (inst
->dst
.file
== PROGRAM_TEMPORARY
&& inst
->dst
.index
== index
) {
3123 return (depth
== 0) ? i
: loop_start
;
3126 if (inst
->op
== TGSI_OPCODE_BGNLOOP
) {
3129 } else if (inst
->op
== TGSI_OPCODE_ENDLOOP
) {
3142 glsl_to_tgsi_visitor::get_last_temp_read(int index
)
3144 int depth
= 0; /* loop depth */
3145 int last
= -1; /* index of last instruction that reads the temporary */
3148 foreach_iter(exec_list_iterator
, iter
, this->instructions
) {
3149 glsl_to_tgsi_instruction
*inst
= (glsl_to_tgsi_instruction
*)iter
.get();
3151 for (j
=0; j
< num_inst_src_regs(inst
->op
); j
++) {
3152 if (inst
->src
[j
].file
== PROGRAM_TEMPORARY
&&
3153 inst
->src
[j
].index
== index
) {
3154 last
= (depth
== 0) ? i
: -2;
3158 if (inst
->op
== TGSI_OPCODE_BGNLOOP
)
3160 else if (inst
->op
== TGSI_OPCODE_ENDLOOP
)
3161 if (--depth
== 0 && last
== -2)
3173 glsl_to_tgsi_visitor::get_last_temp_write(int index
)
3175 int depth
= 0; /* loop depth */
3176 int last
= -1; /* index of last instruction that writes to the temporary */
3179 foreach_iter(exec_list_iterator
, iter
, this->instructions
) {
3180 glsl_to_tgsi_instruction
*inst
= (glsl_to_tgsi_instruction
*)iter
.get();
3182 if (inst
->dst
.file
== PROGRAM_TEMPORARY
&& inst
->dst
.index
== index
)
3183 last
= (depth
== 0) ? i
: -2;
3185 if (inst
->op
== TGSI_OPCODE_BGNLOOP
)
3187 else if (inst
->op
== TGSI_OPCODE_ENDLOOP
)
3188 if (--depth
== 0 && last
== -2)
3200 * On a basic block basis, tracks available PROGRAM_TEMPORARY register
3201 * channels for copy propagation and updates following instructions to
3202 * use the original versions.
3204 * The glsl_to_tgsi_visitor lazily produces code assuming that this pass
3205 * will occur. As an example, a TXP production before this pass:
3207 * 0: MOV TEMP[1], INPUT[4].xyyy;
3208 * 1: MOV TEMP[1].w, INPUT[4].wwww;
3209 * 2: TXP TEMP[2], TEMP[1], texture[0], 2D;
3213 * 0: MOV TEMP[1], INPUT[4].xyyy;
3214 * 1: MOV TEMP[1].w, INPUT[4].wwww;
3215 * 2: TXP TEMP[2], INPUT[4].xyyw, texture[0], 2D;
3217 * which allows for dead code elimination on TEMP[1]'s writes.
3220 glsl_to_tgsi_visitor::copy_propagate(void)
3222 glsl_to_tgsi_instruction
**acp
= rzalloc_array(mem_ctx
,
3223 glsl_to_tgsi_instruction
*,
3224 this->next_temp
* 4);
3225 int *acp_level
= rzalloc_array(mem_ctx
, int, this->next_temp
* 4);
3228 foreach_iter(exec_list_iterator
, iter
, this->instructions
) {
3229 glsl_to_tgsi_instruction
*inst
= (glsl_to_tgsi_instruction
*)iter
.get();
3231 assert(inst
->dst
.file
!= PROGRAM_TEMPORARY
3232 || inst
->dst
.index
< this->next_temp
);
3234 /* First, do any copy propagation possible into the src regs. */
3235 for (int r
= 0; r
< 3; r
++) {
3236 glsl_to_tgsi_instruction
*first
= NULL
;
3238 int acp_base
= inst
->src
[r
].index
* 4;
3240 if (inst
->src
[r
].file
!= PROGRAM_TEMPORARY
||
3241 inst
->src
[r
].reladdr
)
3244 /* See if we can find entries in the ACP consisting of MOVs
3245 * from the same src register for all the swizzled channels
3246 * of this src register reference.
3248 for (int i
= 0; i
< 4; i
++) {
3249 int src_chan
= GET_SWZ(inst
->src
[r
].swizzle
, i
);
3250 glsl_to_tgsi_instruction
*copy_chan
= acp
[acp_base
+ src_chan
];
3257 assert(acp_level
[acp_base
+ src_chan
] <= level
);
3262 if (first
->src
[0].file
!= copy_chan
->src
[0].file
||
3263 first
->src
[0].index
!= copy_chan
->src
[0].index
) {
3271 /* We've now validated that we can copy-propagate to
3272 * replace this src register reference. Do it.
3274 inst
->src
[r
].file
= first
->src
[0].file
;
3275 inst
->src
[r
].index
= first
->src
[0].index
;
3278 for (int i
= 0; i
< 4; i
++) {
3279 int src_chan
= GET_SWZ(inst
->src
[r
].swizzle
, i
);
3280 glsl_to_tgsi_instruction
*copy_inst
= acp
[acp_base
+ src_chan
];
3281 swizzle
|= (GET_SWZ(copy_inst
->src
[0].swizzle
, src_chan
) <<
3284 inst
->src
[r
].swizzle
= swizzle
;
3289 case TGSI_OPCODE_BGNLOOP
:
3290 case TGSI_OPCODE_ENDLOOP
:
3291 /* End of a basic block, clear the ACP entirely. */
3292 memset(acp
, 0, sizeof(*acp
) * this->next_temp
* 4);
3295 case TGSI_OPCODE_IF
:
3299 case TGSI_OPCODE_ENDIF
:
3300 case TGSI_OPCODE_ELSE
:
3301 /* Clear all channels written inside the block from the ACP, but
3302 * leaving those that were not touched.
3304 for (int r
= 0; r
< this->next_temp
; r
++) {
3305 for (int c
= 0; c
< 4; c
++) {
3306 if (!acp
[4 * r
+ c
])
3309 if (acp_level
[4 * r
+ c
] >= level
)
3310 acp
[4 * r
+ c
] = NULL
;
3313 if (inst
->op
== TGSI_OPCODE_ENDIF
)
3318 /* Continuing the block, clear any written channels from
3321 if (inst
->dst
.file
== PROGRAM_TEMPORARY
&& inst
->dst
.reladdr
) {
3322 /* Any temporary might be written, so no copy propagation
3323 * across this instruction.
3325 memset(acp
, 0, sizeof(*acp
) * this->next_temp
* 4);
3326 } else if (inst
->dst
.file
== PROGRAM_OUTPUT
&&
3327 inst
->dst
.reladdr
) {
3328 /* Any output might be written, so no copy propagation
3329 * from outputs across this instruction.
3331 for (int r
= 0; r
< this->next_temp
; r
++) {
3332 for (int c
= 0; c
< 4; c
++) {
3333 if (!acp
[4 * r
+ c
])
3336 if (acp
[4 * r
+ c
]->src
[0].file
== PROGRAM_OUTPUT
)
3337 acp
[4 * r
+ c
] = NULL
;
3340 } else if (inst
->dst
.file
== PROGRAM_TEMPORARY
||
3341 inst
->dst
.file
== PROGRAM_OUTPUT
) {
3342 /* Clear where it's used as dst. */
3343 if (inst
->dst
.file
== PROGRAM_TEMPORARY
) {
3344 for (int c
= 0; c
< 4; c
++) {
3345 if (inst
->dst
.writemask
& (1 << c
)) {
3346 acp
[4 * inst
->dst
.index
+ c
] = NULL
;
3351 /* Clear where it's used as src. */
3352 for (int r
= 0; r
< this->next_temp
; r
++) {
3353 for (int c
= 0; c
< 4; c
++) {
3354 if (!acp
[4 * r
+ c
])
3357 int src_chan
= GET_SWZ(acp
[4 * r
+ c
]->src
[0].swizzle
, c
);
3359 if (acp
[4 * r
+ c
]->src
[0].file
== inst
->dst
.file
&&
3360 acp
[4 * r
+ c
]->src
[0].index
== inst
->dst
.index
&&
3361 inst
->dst
.writemask
& (1 << src_chan
))
3363 acp
[4 * r
+ c
] = NULL
;
3371 /* If this is a copy, add it to the ACP. */
3372 if (inst
->op
== TGSI_OPCODE_MOV
&&
3373 inst
->dst
.file
== PROGRAM_TEMPORARY
&&
3374 !inst
->dst
.reladdr
&&
3376 !inst
->src
[0].reladdr
&&
3377 !inst
->src
[0].negate
) {
3378 for (int i
= 0; i
< 4; i
++) {
3379 if (inst
->dst
.writemask
& (1 << i
)) {
3380 acp
[4 * inst
->dst
.index
+ i
] = inst
;
3381 acp_level
[4 * inst
->dst
.index
+ i
] = level
;
3387 ralloc_free(acp_level
);
3392 * Tracks available PROGRAM_TEMPORARY registers for dead code elimination.
3394 * The glsl_to_tgsi_visitor lazily produces code assuming that this pass
3395 * will occur. As an example, a TXP production after copy propagation but
3398 * 0: MOV TEMP[1], INPUT[4].xyyy;
3399 * 1: MOV TEMP[1].w, INPUT[4].wwww;
3400 * 2: TXP TEMP[2], INPUT[4].xyyw, texture[0], 2D;
3402 * and after this pass:
3404 * 0: TXP TEMP[2], INPUT[4].xyyw, texture[0], 2D;
3406 * FIXME: assumes that all functions are inlined (no support for BGNSUB/ENDSUB)
3407 * FIXME: doesn't eliminate all dead code inside of loops; it steps around them
3410 glsl_to_tgsi_visitor::eliminate_dead_code(void)
3414 for (i
=0; i
< this->next_temp
; i
++) {
3415 int last_read
= get_last_temp_read(i
);
3418 foreach_iter(exec_list_iterator
, iter
, this->instructions
) {
3419 glsl_to_tgsi_instruction
*inst
= (glsl_to_tgsi_instruction
*)iter
.get();
3421 if (inst
->dst
.file
== PROGRAM_TEMPORARY
&& inst
->dst
.index
== i
&&
3434 * On a basic block basis, tracks available PROGRAM_TEMPORARY registers for dead
3435 * code elimination. This is less primitive than eliminate_dead_code(), as it
3436 * is per-channel and can detect consecutive writes without a read between them
3437 * as dead code. However, there is some dead code that can be eliminated by
3438 * eliminate_dead_code() but not this function - for example, this function
3439 * cannot eliminate an instruction writing to a register that is never read and
3440 * is the only instruction writing to that register.
3442 * The glsl_to_tgsi_visitor lazily produces code assuming that this pass
3446 glsl_to_tgsi_visitor::eliminate_dead_code_advanced(void)
3448 glsl_to_tgsi_instruction
**writes
= rzalloc_array(mem_ctx
,
3449 glsl_to_tgsi_instruction
*,
3450 this->next_temp
* 4);
3451 int *write_level
= rzalloc_array(mem_ctx
, int, this->next_temp
* 4);
3455 foreach_iter(exec_list_iterator
, iter
, this->instructions
) {
3456 glsl_to_tgsi_instruction
*inst
= (glsl_to_tgsi_instruction
*)iter
.get();
3458 assert(inst
->dst
.file
!= PROGRAM_TEMPORARY
3459 || inst
->dst
.index
< this->next_temp
);
3462 case TGSI_OPCODE_BGNLOOP
:
3463 case TGSI_OPCODE_ENDLOOP
:
3464 case TGSI_OPCODE_CONT
:
3465 case TGSI_OPCODE_BRK
:
3466 /* End of a basic block, clear the write array entirely.
3468 * This keeps us from killing dead code when the writes are
3469 * on either side of a loop, even when the register isn't touched
3470 * inside the loop. However, glsl_to_tgsi_visitor doesn't seem to emit
3471 * dead code of this type, so it shouldn't make a difference as long as
3472 * the dead code elimination pass in the GLSL compiler does its job.
3474 memset(writes
, 0, sizeof(*writes
) * this->next_temp
* 4);
3477 case TGSI_OPCODE_ENDIF
:
3478 case TGSI_OPCODE_ELSE
:
3479 /* Promote the recorded level of all channels written inside the
3480 * preceding if or else block to the level above the if/else block.
3482 for (int r
= 0; r
< this->next_temp
; r
++) {
3483 for (int c
= 0; c
< 4; c
++) {
3484 if (!writes
[4 * r
+ c
])
3487 if (write_level
[4 * r
+ c
] == level
)
3488 write_level
[4 * r
+ c
] = level
-1;
3492 if(inst
->op
== TGSI_OPCODE_ENDIF
)
3497 case TGSI_OPCODE_IF
:
3499 /* fallthrough to default case to mark the condition as read */
3502 /* Continuing the block, clear any channels from the write array that
3503 * are read by this instruction.
3505 for (unsigned i
= 0; i
< Elements(inst
->src
); i
++) {
3506 if (inst
->src
[i
].file
== PROGRAM_TEMPORARY
&& inst
->src
[i
].reladdr
){
3507 /* Any temporary might be read, so no dead code elimination
3508 * across this instruction.
3510 memset(writes
, 0, sizeof(*writes
) * this->next_temp
* 4);
3511 } else if (inst
->src
[i
].file
== PROGRAM_TEMPORARY
) {
3512 /* Clear where it's used as src. */
3513 int src_chans
= 1 << GET_SWZ(inst
->src
[i
].swizzle
, 0);
3514 src_chans
|= 1 << GET_SWZ(inst
->src
[i
].swizzle
, 1);
3515 src_chans
|= 1 << GET_SWZ(inst
->src
[i
].swizzle
, 2);
3516 src_chans
|= 1 << GET_SWZ(inst
->src
[i
].swizzle
, 3);
3518 for (int c
= 0; c
< 4; c
++) {
3519 if (src_chans
& (1 << c
)) {
3520 writes
[4 * inst
->src
[i
].index
+ c
] = NULL
;
3528 /* If this instruction writes to a temporary, add it to the write array.
3529 * If there is already an instruction in the write array for one or more
3530 * of the channels, flag that channel write as dead.
3532 if (inst
->dst
.file
== PROGRAM_TEMPORARY
&&
3533 !inst
->dst
.reladdr
&&
3535 for (int c
= 0; c
< 4; c
++) {
3536 if (inst
->dst
.writemask
& (1 << c
)) {
3537 if (writes
[4 * inst
->dst
.index
+ c
]) {
3538 if (write_level
[4 * inst
->dst
.index
+ c
] < level
)
3541 writes
[4 * inst
->dst
.index
+ c
]->dead_mask
|= (1 << c
);
3543 writes
[4 * inst
->dst
.index
+ c
] = inst
;
3544 write_level
[4 * inst
->dst
.index
+ c
] = level
;
3550 /* Anything still in the write array at this point is dead code. */
3551 for (int r
= 0; r
< this->next_temp
; r
++) {
3552 for (int c
= 0; c
< 4; c
++) {
3553 glsl_to_tgsi_instruction
*inst
= writes
[4 * r
+ c
];
3555 inst
->dead_mask
|= (1 << c
);
3559 /* Now actually remove the instructions that are completely dead and update
3560 * the writemask of other instructions with dead channels.
3562 foreach_iter(exec_list_iterator
, iter
, this->instructions
) {
3563 glsl_to_tgsi_instruction
*inst
= (glsl_to_tgsi_instruction
*)iter
.get();
3565 if (!inst
->dead_mask
|| !inst
->dst
.writemask
)
3567 else if ((inst
->dst
.writemask
& ~inst
->dead_mask
) == 0) {
3572 inst
->dst
.writemask
&= ~(inst
->dead_mask
);
3575 ralloc_free(write_level
);
3576 ralloc_free(writes
);
3581 /* Merges temporary registers together where possible to reduce the number of
3582 * registers needed to run a program.
3584 * Produces optimal code only after copy propagation and dead code elimination
3587 glsl_to_tgsi_visitor::merge_registers(void)
3589 int *last_reads
= rzalloc_array(mem_ctx
, int, this->next_temp
);
3590 int *first_writes
= rzalloc_array(mem_ctx
, int, this->next_temp
);
3593 /* Read the indices of the last read and first write to each temp register
3594 * into an array so that we don't have to traverse the instruction list as
3596 for (i
=0; i
< this->next_temp
; i
++) {
3597 last_reads
[i
] = get_last_temp_read(i
);
3598 first_writes
[i
] = get_first_temp_write(i
);
3601 /* Start looking for registers with non-overlapping usages that can be
3602 * merged together. */
3603 for (i
=0; i
< this->next_temp
; i
++) {
3604 /* Don't touch unused registers. */
3605 if (last_reads
[i
] < 0 || first_writes
[i
] < 0) continue;
3607 for (j
=0; j
< this->next_temp
; j
++) {
3608 /* Don't touch unused registers. */
3609 if (last_reads
[j
] < 0 || first_writes
[j
] < 0) continue;
3611 /* We can merge the two registers if the first write to j is after or
3612 * in the same instruction as the last read from i. Note that the
3613 * register at index i will always be used earlier or at the same time
3614 * as the register at index j. */
3615 if (first_writes
[i
] <= first_writes
[j
] &&
3616 last_reads
[i
] <= first_writes
[j
])
3618 rename_temp_register(j
, i
); /* Replace all references to j with i.*/
3620 /* Update the first_writes and last_reads arrays with the new
3621 * values for the merged register index, and mark the newly unused
3622 * register index as such. */
3623 last_reads
[i
] = last_reads
[j
];
3624 first_writes
[j
] = -1;
3630 ralloc_free(last_reads
);
3631 ralloc_free(first_writes
);
3634 /* Reassign indices to temporary registers by reusing unused indices created
3635 * by optimization passes. */
3637 glsl_to_tgsi_visitor::renumber_registers(void)
3642 for (i
=0; i
< this->next_temp
; i
++) {
3643 if (get_first_temp_read(i
) < 0) continue;
3645 rename_temp_register(i
, new_index
);
3649 this->next_temp
= new_index
;
3653 * Returns a fragment program which implements the current pixel transfer ops.
3654 * Based on get_pixel_transfer_program in st_atom_pixeltransfer.c.
3657 get_pixel_transfer_visitor(struct st_fragment_program
*fp
,
3658 glsl_to_tgsi_visitor
*original
,
3659 int scale_and_bias
, int pixel_maps
)
3661 glsl_to_tgsi_visitor
*v
= new glsl_to_tgsi_visitor();
3662 struct st_context
*st
= st_context(original
->ctx
);
3663 struct gl_program
*prog
= &fp
->Base
.Base
;
3664 struct gl_program_parameter_list
*params
= _mesa_new_parameter_list();
3665 st_src_reg coord
, src0
;
3667 glsl_to_tgsi_instruction
*inst
;
3669 /* Copy attributes of the glsl_to_tgsi_visitor in the original shader. */
3670 v
->ctx
= original
->ctx
;
3672 v
->shader_program
= NULL
;
3673 v
->glsl_version
= original
->glsl_version
;
3674 v
->native_integers
= original
->native_integers
;
3675 v
->options
= original
->options
;
3676 v
->next_temp
= original
->next_temp
;
3677 v
->num_address_regs
= original
->num_address_regs
;
3678 v
->samplers_used
= prog
->SamplersUsed
= original
->samplers_used
;
3679 v
->indirect_addr_temps
= original
->indirect_addr_temps
;
3680 v
->indirect_addr_consts
= original
->indirect_addr_consts
;
3681 memcpy(&v
->immediates
, &original
->immediates
, sizeof(v
->immediates
));
3682 v
->num_immediates
= original
->num_immediates
;
3685 * Get initial pixel color from the texture.
3686 * TEX colorTemp, fragment.texcoord[0], texture[0], 2D;
3688 coord
= st_src_reg(PROGRAM_INPUT
, FRAG_ATTRIB_TEX0
, glsl_type::vec2_type
);
3689 src0
= v
->get_temp(glsl_type::vec4_type
);
3690 dst0
= st_dst_reg(src0
);
3691 inst
= v
->emit(NULL
, TGSI_OPCODE_TEX
, dst0
, coord
);
3693 inst
->tex_target
= TEXTURE_2D_INDEX
;
3695 prog
->InputsRead
|= FRAG_BIT_TEX0
;
3696 prog
->SamplersUsed
|= (1 << 0); /* mark sampler 0 as used */
3697 v
->samplers_used
|= (1 << 0);
3699 if (scale_and_bias
) {
3700 static const gl_state_index scale_state
[STATE_LENGTH
] =
3701 { STATE_INTERNAL
, STATE_PT_SCALE
,
3702 (gl_state_index
) 0, (gl_state_index
) 0, (gl_state_index
) 0 };
3703 static const gl_state_index bias_state
[STATE_LENGTH
] =
3704 { STATE_INTERNAL
, STATE_PT_BIAS
,
3705 (gl_state_index
) 0, (gl_state_index
) 0, (gl_state_index
) 0 };
3706 GLint scale_p
, bias_p
;
3707 st_src_reg scale
, bias
;
3709 scale_p
= _mesa_add_state_reference(params
, scale_state
);
3710 bias_p
= _mesa_add_state_reference(params
, bias_state
);
3712 /* MAD colorTemp, colorTemp, scale, bias; */
3713 scale
= st_src_reg(PROGRAM_STATE_VAR
, scale_p
, GLSL_TYPE_FLOAT
);
3714 bias
= st_src_reg(PROGRAM_STATE_VAR
, bias_p
, GLSL_TYPE_FLOAT
);
3715 inst
= v
->emit(NULL
, TGSI_OPCODE_MAD
, dst0
, src0
, scale
, bias
);
3719 st_src_reg temp
= v
->get_temp(glsl_type::vec4_type
);
3720 st_dst_reg temp_dst
= st_dst_reg(temp
);
3722 assert(st
->pixel_xfer
.pixelmap_texture
);
3724 /* With a little effort, we can do four pixel map look-ups with
3725 * two TEX instructions:
3728 /* TEX temp.rg, colorTemp.rgba, texture[1], 2D; */
3729 temp_dst
.writemask
= WRITEMASK_XY
; /* write R,G */
3730 inst
= v
->emit(NULL
, TGSI_OPCODE_TEX
, temp_dst
, src0
);
3732 inst
->tex_target
= TEXTURE_2D_INDEX
;
3734 /* TEX temp.ba, colorTemp.baba, texture[1], 2D; */
3735 src0
.swizzle
= MAKE_SWIZZLE4(SWIZZLE_Z
, SWIZZLE_W
, SWIZZLE_Z
, SWIZZLE_W
);
3736 temp_dst
.writemask
= WRITEMASK_ZW
; /* write B,A */
3737 inst
= v
->emit(NULL
, TGSI_OPCODE_TEX
, temp_dst
, src0
);
3739 inst
->tex_target
= TEXTURE_2D_INDEX
;
3741 prog
->SamplersUsed
|= (1 << 1); /* mark sampler 1 as used */
3742 v
->samplers_used
|= (1 << 1);
3744 /* MOV colorTemp, temp; */
3745 inst
= v
->emit(NULL
, TGSI_OPCODE_MOV
, dst0
, temp
);
3748 /* Now copy the instructions from the original glsl_to_tgsi_visitor into the
3750 foreach_iter(exec_list_iterator
, iter
, original
->instructions
) {
3751 glsl_to_tgsi_instruction
*inst
= (glsl_to_tgsi_instruction
*)iter
.get();
3752 glsl_to_tgsi_instruction
*newinst
;
3753 st_src_reg src_regs
[3];
3755 if (inst
->dst
.file
== PROGRAM_OUTPUT
)
3756 prog
->OutputsWritten
|= BITFIELD64_BIT(inst
->dst
.index
);
3758 for (int i
=0; i
<3; i
++) {
3759 src_regs
[i
] = inst
->src
[i
];
3760 if (src_regs
[i
].file
== PROGRAM_INPUT
&&
3761 src_regs
[i
].index
== FRAG_ATTRIB_COL0
)
3763 src_regs
[i
].file
= PROGRAM_TEMPORARY
;
3764 src_regs
[i
].index
= src0
.index
;
3766 else if (src_regs
[i
].file
== PROGRAM_INPUT
)
3767 prog
->InputsRead
|= BITFIELD64_BIT(src_regs
[i
].index
);
3770 newinst
= v
->emit(NULL
, inst
->op
, inst
->dst
, src_regs
[0], src_regs
[1], src_regs
[2]);
3771 newinst
->tex_target
= inst
->tex_target
;
3774 /* Make modifications to fragment program info. */
3775 prog
->Parameters
= _mesa_combine_parameter_lists(params
,
3776 original
->prog
->Parameters
);
3777 _mesa_free_parameter_list(params
);
3778 count_resources(v
, prog
);
3779 fp
->glsl_to_tgsi
= v
;
3783 * Make fragment program for glBitmap:
3784 * Sample the texture and kill the fragment if the bit is 0.
3785 * This program will be combined with the user's fragment program.
3787 * Based on make_bitmap_fragment_program in st_cb_bitmap.c.
3790 get_bitmap_visitor(struct st_fragment_program
*fp
,
3791 glsl_to_tgsi_visitor
*original
, int samplerIndex
)
3793 glsl_to_tgsi_visitor
*v
= new glsl_to_tgsi_visitor();
3794 struct st_context
*st
= st_context(original
->ctx
);
3795 struct gl_program
*prog
= &fp
->Base
.Base
;
3796 st_src_reg coord
, src0
;
3798 glsl_to_tgsi_instruction
*inst
;
3800 /* Copy attributes of the glsl_to_tgsi_visitor in the original shader. */
3801 v
->ctx
= original
->ctx
;
3803 v
->shader_program
= NULL
;
3804 v
->glsl_version
= original
->glsl_version
;
3805 v
->native_integers
= original
->native_integers
;
3806 v
->options
= original
->options
;
3807 v
->next_temp
= original
->next_temp
;
3808 v
->num_address_regs
= original
->num_address_regs
;
3809 v
->samplers_used
= prog
->SamplersUsed
= original
->samplers_used
;
3810 v
->indirect_addr_temps
= original
->indirect_addr_temps
;
3811 v
->indirect_addr_consts
= original
->indirect_addr_consts
;
3812 memcpy(&v
->immediates
, &original
->immediates
, sizeof(v
->immediates
));
3813 v
->num_immediates
= original
->num_immediates
;
3815 /* TEX tmp0, fragment.texcoord[0], texture[0], 2D; */
3816 coord
= st_src_reg(PROGRAM_INPUT
, FRAG_ATTRIB_TEX0
, glsl_type::vec2_type
);
3817 src0
= v
->get_temp(glsl_type::vec4_type
);
3818 dst0
= st_dst_reg(src0
);
3819 inst
= v
->emit(NULL
, TGSI_OPCODE_TEX
, dst0
, coord
);
3820 inst
->sampler
= samplerIndex
;
3821 inst
->tex_target
= TEXTURE_2D_INDEX
;
3823 prog
->InputsRead
|= FRAG_BIT_TEX0
;
3824 prog
->SamplersUsed
|= (1 << samplerIndex
); /* mark sampler as used */
3825 v
->samplers_used
|= (1 << samplerIndex
);
3827 /* KIL if -tmp0 < 0 # texel=0 -> keep / texel=0 -> discard */
3828 src0
.negate
= NEGATE_XYZW
;
3829 if (st
->bitmap
.tex_format
== PIPE_FORMAT_L8_UNORM
)
3830 src0
.swizzle
= SWIZZLE_XXXX
;
3831 inst
= v
->emit(NULL
, TGSI_OPCODE_KIL
, undef_dst
, src0
);
3833 /* Now copy the instructions from the original glsl_to_tgsi_visitor into the
3835 foreach_iter(exec_list_iterator
, iter
, original
->instructions
) {
3836 glsl_to_tgsi_instruction
*inst
= (glsl_to_tgsi_instruction
*)iter
.get();
3837 glsl_to_tgsi_instruction
*newinst
;
3838 st_src_reg src_regs
[3];
3840 if (inst
->dst
.file
== PROGRAM_OUTPUT
)
3841 prog
->OutputsWritten
|= BITFIELD64_BIT(inst
->dst
.index
);
3843 for (int i
=0; i
<3; i
++) {
3844 src_regs
[i
] = inst
->src
[i
];
3845 if (src_regs
[i
].file
== PROGRAM_INPUT
)
3846 prog
->InputsRead
|= BITFIELD64_BIT(src_regs
[i
].index
);
3849 newinst
= v
->emit(NULL
, inst
->op
, inst
->dst
, src_regs
[0], src_regs
[1], src_regs
[2]);
3850 newinst
->tex_target
= inst
->tex_target
;
3853 /* Make modifications to fragment program info. */
3854 prog
->Parameters
= _mesa_clone_parameter_list(original
->prog
->Parameters
);
3855 count_resources(v
, prog
);
3856 fp
->glsl_to_tgsi
= v
;
3859 /* ------------------------- TGSI conversion stuff -------------------------- */
3861 unsigned branch_target
;
3866 * Intermediate state used during shader translation.
3868 struct st_translate
{
3869 struct ureg_program
*ureg
;
3871 struct ureg_dst temps
[MAX_TEMPS
];
3872 struct ureg_src
*constants
;
3873 struct ureg_src
*immediates
;
3874 struct ureg_dst outputs
[PIPE_MAX_SHADER_OUTPUTS
];
3875 struct ureg_src inputs
[PIPE_MAX_SHADER_INPUTS
];
3876 struct ureg_dst address
[1];
3877 struct ureg_src samplers
[PIPE_MAX_SAMPLERS
];
3878 struct ureg_src systemValues
[SYSTEM_VALUE_MAX
];
3880 const GLuint
*inputMapping
;
3881 const GLuint
*outputMapping
;
3883 /* For every instruction that contains a label (eg CALL), keep
3884 * details so that we can go back afterwards and emit the correct
3885 * tgsi instruction number for each label.
3887 struct label
*labels
;
3888 unsigned labels_size
;
3889 unsigned labels_count
;
3891 /* Keep a record of the tgsi instruction number that each mesa
3892 * instruction starts at, will be used to fix up labels after
3897 unsigned insn_count
;
3899 unsigned procType
; /**< TGSI_PROCESSOR_VERTEX/FRAGMENT */
3904 /** Map Mesa's SYSTEM_VALUE_x to TGSI_SEMANTIC_x */
3905 static unsigned mesa_sysval_to_semantic
[SYSTEM_VALUE_MAX
] = {
3907 TGSI_SEMANTIC_VERTEXID
,
3908 TGSI_SEMANTIC_INSTANCEID
3912 * Make note of a branch to a label in the TGSI code.
3913 * After we've emitted all instructions, we'll go over the list
3914 * of labels built here and patch the TGSI code with the actual
3915 * location of each label.
3917 static unsigned *get_label(struct st_translate
*t
, unsigned branch_target
)
3921 if (t
->labels_count
+ 1 >= t
->labels_size
) {
3922 t
->labels_size
= 1 << (util_logbase2(t
->labels_size
) + 1);
3923 t
->labels
= (struct label
*)realloc(t
->labels
,
3924 t
->labels_size
* sizeof(struct label
));
3925 if (t
->labels
== NULL
) {
3926 static unsigned dummy
;
3932 i
= t
->labels_count
++;
3933 t
->labels
[i
].branch_target
= branch_target
;
3934 return &t
->labels
[i
].token
;
3938 * Called prior to emitting the TGSI code for each instruction.
3939 * Allocate additional space for instructions if needed.
3940 * Update the insn[] array so the next glsl_to_tgsi_instruction points to
3941 * the next TGSI instruction.
3943 static void set_insn_start(struct st_translate
*t
, unsigned start
)
3945 if (t
->insn_count
+ 1 >= t
->insn_size
) {
3946 t
->insn_size
= 1 << (util_logbase2(t
->insn_size
) + 1);
3947 t
->insn
= (unsigned *)realloc(t
->insn
, t
->insn_size
* sizeof(t
->insn
[0]));
3948 if (t
->insn
== NULL
) {
3954 t
->insn
[t
->insn_count
++] = start
;
3958 * Map a glsl_to_tgsi constant/immediate to a TGSI immediate.
3960 static struct ureg_src
3961 emit_immediate(struct st_translate
*t
,
3962 gl_constant_value values
[4],
3965 struct ureg_program
*ureg
= t
->ureg
;
3970 return ureg_DECL_immediate(ureg
, &values
[0].f
, size
);
3972 return ureg_DECL_immediate_int(ureg
, &values
[0].i
, size
);
3973 case GL_UNSIGNED_INT
:
3975 return ureg_DECL_immediate_uint(ureg
, &values
[0].u
, size
);
3977 assert(!"should not get here - type must be float, int, uint, or bool");
3978 return ureg_src_undef();
3983 * Map a glsl_to_tgsi dst register to a TGSI ureg_dst register.
3985 static struct ureg_dst
3986 dst_register(struct st_translate
*t
,
3987 gl_register_file file
,
3991 case PROGRAM_UNDEFINED
:
3992 return ureg_dst_undef();
3994 case PROGRAM_TEMPORARY
:
3995 if (ureg_dst_is_undef(t
->temps
[index
]))
3996 t
->temps
[index
] = ureg_DECL_local_temporary(t
->ureg
);
3998 return t
->temps
[index
];
4000 case PROGRAM_OUTPUT
:
4001 if (t
->procType
== TGSI_PROCESSOR_VERTEX
)
4002 assert(index
< VERT_RESULT_MAX
);
4003 else if (t
->procType
== TGSI_PROCESSOR_FRAGMENT
)
4004 assert(index
< FRAG_RESULT_MAX
);
4006 assert(index
< GEOM_RESULT_MAX
);
4008 assert(t
->outputMapping
[index
] < Elements(t
->outputs
));
4010 return t
->outputs
[t
->outputMapping
[index
]];
4012 case PROGRAM_ADDRESS
:
4013 return t
->address
[index
];
4016 assert(!"unknown dst register file");
4017 return ureg_dst_undef();
4022 * Map a glsl_to_tgsi src register to a TGSI ureg_src register.
4024 static struct ureg_src
4025 src_register(struct st_translate
*t
,
4026 gl_register_file file
,
4030 case PROGRAM_UNDEFINED
:
4031 return ureg_src_undef();
4033 case PROGRAM_TEMPORARY
:
4035 assert(index
< Elements(t
->temps
));
4036 if (ureg_dst_is_undef(t
->temps
[index
]))
4037 t
->temps
[index
] = ureg_DECL_local_temporary(t
->ureg
);
4038 return ureg_src(t
->temps
[index
]);
4040 case PROGRAM_NAMED_PARAM
:
4041 case PROGRAM_ENV_PARAM
:
4042 case PROGRAM_LOCAL_PARAM
:
4043 case PROGRAM_UNIFORM
:
4045 return t
->constants
[index
];
4046 case PROGRAM_STATE_VAR
:
4047 case PROGRAM_CONSTANT
: /* ie, immediate */
4049 return ureg_DECL_constant(t
->ureg
, 0);
4051 return t
->constants
[index
];
4053 case PROGRAM_IMMEDIATE
:
4054 return t
->immediates
[index
];
4057 assert(t
->inputMapping
[index
] < Elements(t
->inputs
));
4058 return t
->inputs
[t
->inputMapping
[index
]];
4060 case PROGRAM_OUTPUT
:
4061 assert(t
->outputMapping
[index
] < Elements(t
->outputs
));
4062 return ureg_src(t
->outputs
[t
->outputMapping
[index
]]); /* not needed? */
4064 case PROGRAM_ADDRESS
:
4065 return ureg_src(t
->address
[index
]);
4067 case PROGRAM_SYSTEM_VALUE
:
4068 assert(index
< Elements(t
->systemValues
));
4069 return t
->systemValues
[index
];
4072 assert(!"unknown src register file");
4073 return ureg_src_undef();
4078 * Create a TGSI ureg_dst register from an st_dst_reg.
4080 static struct ureg_dst
4081 translate_dst(struct st_translate
*t
,
4082 const st_dst_reg
*dst_reg
,
4083 bool saturate
, bool clamp_color
)
4085 struct ureg_dst dst
= dst_register(t
,
4089 dst
= ureg_writemask(dst
, dst_reg
->writemask
);
4092 dst
= ureg_saturate(dst
);
4093 else if (clamp_color
&& dst_reg
->file
== PROGRAM_OUTPUT
) {
4094 /* Clamp colors for ARB_color_buffer_float. */
4095 switch (t
->procType
) {
4096 case TGSI_PROCESSOR_VERTEX
:
4097 /* XXX if the geometry shader is present, this must be done there
4098 * instead of here. */
4099 if (dst_reg
->index
== VERT_RESULT_COL0
||
4100 dst_reg
->index
== VERT_RESULT_COL1
||
4101 dst_reg
->index
== VERT_RESULT_BFC0
||
4102 dst_reg
->index
== VERT_RESULT_BFC1
) {
4103 dst
= ureg_saturate(dst
);
4107 case TGSI_PROCESSOR_FRAGMENT
:
4108 if (dst_reg
->index
>= FRAG_RESULT_COLOR
) {
4109 dst
= ureg_saturate(dst
);
4115 if (dst_reg
->reladdr
!= NULL
)
4116 dst
= ureg_dst_indirect(dst
, ureg_src(t
->address
[0]));
4122 * Create a TGSI ureg_src register from an st_src_reg.
4124 static struct ureg_src
4125 translate_src(struct st_translate
*t
, const st_src_reg
*src_reg
)
4127 struct ureg_src src
= src_register(t
, src_reg
->file
, src_reg
->index
);
4129 src
= ureg_swizzle(src
,
4130 GET_SWZ(src_reg
->swizzle
, 0) & 0x3,
4131 GET_SWZ(src_reg
->swizzle
, 1) & 0x3,
4132 GET_SWZ(src_reg
->swizzle
, 2) & 0x3,
4133 GET_SWZ(src_reg
->swizzle
, 3) & 0x3);
4135 if ((src_reg
->negate
& 0xf) == NEGATE_XYZW
)
4136 src
= ureg_negate(src
);
4138 if (src_reg
->reladdr
!= NULL
) {
4139 /* Normally ureg_src_indirect() would be used here, but a stupid compiler
4140 * bug in g++ makes ureg_src_indirect (an inline C function) erroneously
4141 * set the bit for src.Negate. So we have to do the operation manually
4142 * here to work around the compiler's problems. */
4143 /*src = ureg_src_indirect(src, ureg_src(t->address[0]));*/
4144 struct ureg_src addr
= ureg_src(t
->address
[0]);
4146 src
.IndirectFile
= addr
.File
;
4147 src
.IndirectIndex
= addr
.Index
;
4148 src
.IndirectSwizzle
= addr
.SwizzleX
;
4150 if (src_reg
->file
!= PROGRAM_INPUT
&&
4151 src_reg
->file
!= PROGRAM_OUTPUT
) {
4152 /* If src_reg->index was negative, it was set to zero in
4153 * src_register(). Reassign it now. But don't do this
4154 * for input/output regs since they get remapped while
4155 * const buffers don't.
4157 src
.Index
= src_reg
->index
;
4164 static struct tgsi_texture_offset
4165 translate_tex_offset(struct st_translate
*t
,
4166 const struct tgsi_texture_offset
*in_offset
)
4168 struct tgsi_texture_offset offset
;
4170 assert(in_offset
->File
== PROGRAM_IMMEDIATE
);
4172 offset
.File
= TGSI_FILE_IMMEDIATE
;
4173 offset
.Index
= in_offset
->Index
;
4174 offset
.SwizzleX
= in_offset
->SwizzleX
;
4175 offset
.SwizzleY
= in_offset
->SwizzleY
;
4176 offset
.SwizzleZ
= in_offset
->SwizzleZ
;
4182 compile_tgsi_instruction(struct st_translate
*t
,
4183 const glsl_to_tgsi_instruction
*inst
,
4184 bool clamp_dst_color_output
)
4186 struct ureg_program
*ureg
= t
->ureg
;
4188 struct ureg_dst dst
[1];
4189 struct ureg_src src
[4];
4190 struct tgsi_texture_offset texoffsets
[MAX_GLSL_TEXTURE_OFFSET
];
4195 num_dst
= num_inst_dst_regs(inst
->op
);
4196 num_src
= num_inst_src_regs(inst
->op
);
4199 dst
[0] = translate_dst(t
,
4202 clamp_dst_color_output
);
4204 for (i
= 0; i
< num_src
; i
++)
4205 src
[i
] = translate_src(t
, &inst
->src
[i
]);
4208 case TGSI_OPCODE_BGNLOOP
:
4209 case TGSI_OPCODE_CAL
:
4210 case TGSI_OPCODE_ELSE
:
4211 case TGSI_OPCODE_ENDLOOP
:
4212 case TGSI_OPCODE_IF
:
4213 assert(num_dst
== 0);
4214 ureg_label_insn(ureg
,
4218 inst
->op
== TGSI_OPCODE_CAL
? inst
->function
->sig_id
: 0));
4221 case TGSI_OPCODE_TEX
:
4222 case TGSI_OPCODE_TXB
:
4223 case TGSI_OPCODE_TXD
:
4224 case TGSI_OPCODE_TXL
:
4225 case TGSI_OPCODE_TXP
:
4226 case TGSI_OPCODE_TXQ
:
4227 case TGSI_OPCODE_TXF
:
4228 src
[num_src
++] = t
->samplers
[inst
->sampler
];
4229 for (i
= 0; i
< inst
->tex_offset_num_offset
; i
++) {
4230 texoffsets
[i
] = translate_tex_offset(t
, &inst
->tex_offsets
[i
]);
4235 st_translate_texture_target(inst
->tex_target
, inst
->tex_shadow
),
4236 texoffsets
, inst
->tex_offset_num_offset
,
4240 case TGSI_OPCODE_SCS
:
4241 dst
[0] = ureg_writemask(dst
[0], TGSI_WRITEMASK_XY
);
4242 ureg_insn(ureg
, inst
->op
, dst
, num_dst
, src
, num_src
);
4255 * Emit the TGSI instructions for inverting and adjusting WPOS.
4256 * This code is unavoidable because it also depends on whether
4257 * a FBO is bound (STATE_FB_WPOS_Y_TRANSFORM).
4260 emit_wpos_adjustment( struct st_translate
*t
,
4261 const struct gl_program
*program
,
4263 GLfloat adjX
, GLfloat adjY
[2])
4265 struct ureg_program
*ureg
= t
->ureg
;
4267 /* Fragment program uses fragment position input.
4268 * Need to replace instances of INPUT[WPOS] with temp T
4269 * where T = INPUT[WPOS] by y is inverted.
4271 static const gl_state_index wposTransformState
[STATE_LENGTH
]
4272 = { STATE_INTERNAL
, STATE_FB_WPOS_Y_TRANSFORM
,
4273 (gl_state_index
)0, (gl_state_index
)0, (gl_state_index
)0 };
4275 /* XXX: note we are modifying the incoming shader here! Need to
4276 * do this before emitting the constant decls below, or this
4279 unsigned wposTransConst
= _mesa_add_state_reference(program
->Parameters
,
4280 wposTransformState
);
4282 struct ureg_src wpostrans
= ureg_DECL_constant( ureg
, wposTransConst
);
4283 struct ureg_dst wpos_temp
= ureg_DECL_temporary( ureg
);
4284 struct ureg_src wpos_input
= t
->inputs
[t
->inputMapping
[FRAG_ATTRIB_WPOS
]];
4286 /* First, apply the coordinate shift: */
4287 if (adjX
|| adjY
[0] || adjY
[1]) {
4288 if (adjY
[0] != adjY
[1]) {
4289 /* Adjust the y coordinate by adjY[1] or adjY[0] respectively
4290 * depending on whether inversion is actually going to be applied
4291 * or not, which is determined by testing against the inversion
4292 * state variable used below, which will be either +1 or -1.
4294 struct ureg_dst adj_temp
= ureg_DECL_local_temporary(ureg
);
4296 ureg_CMP(ureg
, adj_temp
,
4297 ureg_scalar(wpostrans
, invert
? 2 : 0),
4298 ureg_imm4f(ureg
, adjX
, adjY
[0], 0.0f
, 0.0f
),
4299 ureg_imm4f(ureg
, adjX
, adjY
[1], 0.0f
, 0.0f
));
4300 ureg_ADD(ureg
, wpos_temp
, wpos_input
, ureg_src(adj_temp
));
4302 ureg_ADD(ureg
, wpos_temp
, wpos_input
,
4303 ureg_imm4f(ureg
, adjX
, adjY
[0], 0.0f
, 0.0f
));
4305 wpos_input
= ureg_src(wpos_temp
);
4307 /* MOV wpos_temp, input[wpos]
4309 ureg_MOV( ureg
, wpos_temp
, wpos_input
);
4312 /* Now the conditional y flip: STATE_FB_WPOS_Y_TRANSFORM.xy/zw will be
4313 * inversion/identity, or the other way around if we're drawing to an FBO.
4316 /* MAD wpos_temp.y, wpos_input, wpostrans.xxxx, wpostrans.yyyy
4319 ureg_writemask(wpos_temp
, TGSI_WRITEMASK_Y
),
4321 ureg_scalar(wpostrans
, 0),
4322 ureg_scalar(wpostrans
, 1));
4324 /* MAD wpos_temp.y, wpos_input, wpostrans.zzzz, wpostrans.wwww
4327 ureg_writemask(wpos_temp
, TGSI_WRITEMASK_Y
),
4329 ureg_scalar(wpostrans
, 2),
4330 ureg_scalar(wpostrans
, 3));
4333 /* Use wpos_temp as position input from here on:
4335 t
->inputs
[t
->inputMapping
[FRAG_ATTRIB_WPOS
]] = ureg_src(wpos_temp
);
4340 * Emit fragment position/ooordinate code.
4343 emit_wpos(struct st_context
*st
,
4344 struct st_translate
*t
,
4345 const struct gl_program
*program
,
4346 struct ureg_program
*ureg
)
4348 const struct gl_fragment_program
*fp
=
4349 (const struct gl_fragment_program
*) program
;
4350 struct pipe_screen
*pscreen
= st
->pipe
->screen
;
4351 GLfloat adjX
= 0.0f
;
4352 GLfloat adjY
[2] = { 0.0f
, 0.0f
};
4353 boolean invert
= FALSE
;
4355 /* Query the pixel center conventions supported by the pipe driver and set
4356 * adjX, adjY to help out if it cannot handle the requested one internally.
4358 * The bias of the y-coordinate depends on whether y-inversion takes place
4359 * (adjY[1]) or not (adjY[0]), which is in turn dependent on whether we are
4360 * drawing to an FBO (causes additional inversion), and whether the the pipe
4361 * driver origin and the requested origin differ (the latter condition is
4362 * stored in the 'invert' variable).
4364 * For height = 100 (i = integer, h = half-integer, l = lower, u = upper):
4366 * center shift only:
4371 * l,i -> u,i: ( 0.0 + 1.0) * -1 + 100 = 99
4372 * l,h -> u,h: ( 0.5 + 0.0) * -1 + 100 = 99.5
4373 * u,i -> l,i: (99.0 + 1.0) * -1 + 100 = 0
4374 * u,h -> l,h: (99.5 + 0.0) * -1 + 100 = 0.5
4376 * inversion and center shift:
4377 * l,i -> u,h: ( 0.0 + 0.5) * -1 + 100 = 99.5
4378 * l,h -> u,i: ( 0.5 + 0.5) * -1 + 100 = 99
4379 * u,i -> l,h: (99.0 + 0.5) * -1 + 100 = 0.5
4380 * u,h -> l,i: (99.5 + 0.5) * -1 + 100 = 0
4382 if (fp
->OriginUpperLeft
) {
4383 /* Fragment shader wants origin in upper-left */
4384 if (pscreen
->get_param(pscreen
, PIPE_CAP_TGSI_FS_COORD_ORIGIN_UPPER_LEFT
)) {
4385 /* the driver supports upper-left origin */
4387 else if (pscreen
->get_param(pscreen
, PIPE_CAP_TGSI_FS_COORD_ORIGIN_LOWER_LEFT
)) {
4388 /* the driver supports lower-left origin, need to invert Y */
4389 ureg_property_fs_coord_origin(ureg
, TGSI_FS_COORD_ORIGIN_LOWER_LEFT
);
4396 /* Fragment shader wants origin in lower-left */
4397 if (pscreen
->get_param(pscreen
, PIPE_CAP_TGSI_FS_COORD_ORIGIN_LOWER_LEFT
))
4398 /* the driver supports lower-left origin */
4399 ureg_property_fs_coord_origin(ureg
, TGSI_FS_COORD_ORIGIN_LOWER_LEFT
);
4400 else if (pscreen
->get_param(pscreen
, PIPE_CAP_TGSI_FS_COORD_ORIGIN_UPPER_LEFT
))
4401 /* the driver supports upper-left origin, need to invert Y */
4407 if (fp
->PixelCenterInteger
) {
4408 /* Fragment shader wants pixel center integer */
4409 if (pscreen
->get_param(pscreen
, PIPE_CAP_TGSI_FS_COORD_PIXEL_CENTER_INTEGER
)) {
4410 /* the driver supports pixel center integer */
4412 ureg_property_fs_coord_pixel_center(ureg
, TGSI_FS_COORD_PIXEL_CENTER_INTEGER
);
4414 else if (pscreen
->get_param(pscreen
, PIPE_CAP_TGSI_FS_COORD_PIXEL_CENTER_HALF_INTEGER
)) {
4415 /* the driver supports pixel center half integer, need to bias X,Y */
4424 /* Fragment shader wants pixel center half integer */
4425 if (pscreen
->get_param(pscreen
, PIPE_CAP_TGSI_FS_COORD_PIXEL_CENTER_HALF_INTEGER
)) {
4426 /* the driver supports pixel center half integer */
4428 else if (pscreen
->get_param(pscreen
, PIPE_CAP_TGSI_FS_COORD_PIXEL_CENTER_INTEGER
)) {
4429 /* the driver supports pixel center integer, need to bias X,Y */
4430 adjX
= adjY
[0] = adjY
[1] = 0.5f
;
4431 ureg_property_fs_coord_pixel_center(ureg
, TGSI_FS_COORD_PIXEL_CENTER_INTEGER
);
4437 /* we invert after adjustment so that we avoid the MOV to temporary,
4438 * and reuse the adjustment ADD instead */
4439 emit_wpos_adjustment(t
, program
, invert
, adjX
, adjY
);
4443 * OpenGL's fragment gl_FrontFace input is 1 for front-facing, 0 for back.
4444 * TGSI uses +1 for front, -1 for back.
4445 * This function converts the TGSI value to the GL value. Simply clamping/
4446 * saturating the value to [0,1] does the job.
4449 emit_face_var(struct st_translate
*t
)
4451 struct ureg_program
*ureg
= t
->ureg
;
4452 struct ureg_dst face_temp
= ureg_DECL_temporary(ureg
);
4453 struct ureg_src face_input
= t
->inputs
[t
->inputMapping
[FRAG_ATTRIB_FACE
]];
4455 /* MOV_SAT face_temp, input[face] */
4456 face_temp
= ureg_saturate(face_temp
);
4457 ureg_MOV(ureg
, face_temp
, face_input
);
4459 /* Use face_temp as face input from here on: */
4460 t
->inputs
[t
->inputMapping
[FRAG_ATTRIB_FACE
]] = ureg_src(face_temp
);
4464 emit_edgeflags(struct st_translate
*t
)
4466 struct ureg_program
*ureg
= t
->ureg
;
4467 struct ureg_dst edge_dst
= t
->outputs
[t
->outputMapping
[VERT_RESULT_EDGE
]];
4468 struct ureg_src edge_src
= t
->inputs
[t
->inputMapping
[VERT_ATTRIB_EDGEFLAG
]];
4470 ureg_MOV(ureg
, edge_dst
, edge_src
);
4474 * Translate intermediate IR (glsl_to_tgsi_instruction) to TGSI format.
4475 * \param program the program to translate
4476 * \param numInputs number of input registers used
4477 * \param inputMapping maps Mesa fragment program inputs to TGSI generic
4479 * \param inputSemanticName the TGSI_SEMANTIC flag for each input
4480 * \param inputSemanticIndex the semantic index (ex: which texcoord) for
4482 * \param interpMode the TGSI_INTERPOLATE_LINEAR/PERSP mode for each input
4483 * \param numOutputs number of output registers used
4484 * \param outputMapping maps Mesa fragment program outputs to TGSI
4486 * \param outputSemanticName the TGSI_SEMANTIC flag for each output
4487 * \param outputSemanticIndex the semantic index (ex: which texcoord) for
4490 * \return PIPE_OK or PIPE_ERROR_OUT_OF_MEMORY
4492 extern "C" enum pipe_error
4493 st_translate_program(
4494 struct gl_context
*ctx
,
4496 struct ureg_program
*ureg
,
4497 glsl_to_tgsi_visitor
*program
,
4498 const struct gl_program
*proginfo
,
4500 const GLuint inputMapping
[],
4501 const ubyte inputSemanticName
[],
4502 const ubyte inputSemanticIndex
[],
4503 const GLuint interpMode
[],
4504 const GLboolean is_centroid
[],
4506 const GLuint outputMapping
[],
4507 const ubyte outputSemanticName
[],
4508 const ubyte outputSemanticIndex
[],
4509 boolean passthrough_edgeflags
,
4510 boolean clamp_color
)
4512 struct st_translate
*t
;
4514 enum pipe_error ret
= PIPE_OK
;
4516 assert(numInputs
<= Elements(t
->inputs
));
4517 assert(numOutputs
<= Elements(t
->outputs
));
4519 t
= CALLOC_STRUCT(st_translate
);
4521 ret
= PIPE_ERROR_OUT_OF_MEMORY
;
4525 memset(t
, 0, sizeof *t
);
4527 t
->procType
= procType
;
4528 t
->inputMapping
= inputMapping
;
4529 t
->outputMapping
= outputMapping
;
4532 if (program
->shader_program
) {
4533 for (i
= 0; i
< program
->shader_program
->NumUserUniformStorage
; i
++) {
4534 struct gl_uniform_storage
*const storage
=
4535 &program
->shader_program
->UniformStorage
[i
];
4537 _mesa_uniform_detach_all_driver_storage(storage
);
4542 * Declare input attributes.
4544 if (procType
== TGSI_PROCESSOR_FRAGMENT
) {
4545 for (i
= 0; i
< numInputs
; i
++) {
4546 t
->inputs
[i
] = ureg_DECL_fs_input_cyl_centroid(ureg
,
4547 inputSemanticName
[i
],
4548 inputSemanticIndex
[i
],
4553 if (proginfo
->InputsRead
& FRAG_BIT_WPOS
) {
4554 /* Must do this after setting up t->inputs, and before
4555 * emitting constant references, below:
4557 emit_wpos(st_context(ctx
), t
, proginfo
, ureg
);
4560 if (proginfo
->InputsRead
& FRAG_BIT_FACE
)
4564 * Declare output attributes.
4566 for (i
= 0; i
< numOutputs
; i
++) {
4567 switch (outputSemanticName
[i
]) {
4568 case TGSI_SEMANTIC_POSITION
:
4569 t
->outputs
[i
] = ureg_DECL_output(ureg
,
4570 TGSI_SEMANTIC_POSITION
, /* Z/Depth */
4571 outputSemanticIndex
[i
]);
4572 t
->outputs
[i
] = ureg_writemask(t
->outputs
[i
], TGSI_WRITEMASK_Z
);
4574 case TGSI_SEMANTIC_STENCIL
:
4575 t
->outputs
[i
] = ureg_DECL_output(ureg
,
4576 TGSI_SEMANTIC_STENCIL
, /* Stencil */
4577 outputSemanticIndex
[i
]);
4578 t
->outputs
[i
] = ureg_writemask(t
->outputs
[i
], TGSI_WRITEMASK_Y
);
4580 case TGSI_SEMANTIC_COLOR
:
4581 t
->outputs
[i
] = ureg_DECL_output(ureg
,
4582 TGSI_SEMANTIC_COLOR
,
4583 outputSemanticIndex
[i
]);
4586 assert(!"fragment shader outputs must be POSITION/STENCIL/COLOR");
4587 ret
= PIPE_ERROR_BAD_INPUT
;
4592 else if (procType
== TGSI_PROCESSOR_GEOMETRY
) {
4593 for (i
= 0; i
< numInputs
; i
++) {
4594 t
->inputs
[i
] = ureg_DECL_gs_input(ureg
,
4596 inputSemanticName
[i
],
4597 inputSemanticIndex
[i
]);
4600 for (i
= 0; i
< numOutputs
; i
++) {
4601 t
->outputs
[i
] = ureg_DECL_output(ureg
,
4602 outputSemanticName
[i
],
4603 outputSemanticIndex
[i
]);
4607 assert(procType
== TGSI_PROCESSOR_VERTEX
);
4609 for (i
= 0; i
< numInputs
; i
++) {
4610 t
->inputs
[i
] = ureg_DECL_vs_input(ureg
, i
);
4613 for (i
= 0; i
< numOutputs
; i
++) {
4614 t
->outputs
[i
] = ureg_DECL_output(ureg
,
4615 outputSemanticName
[i
],
4616 outputSemanticIndex
[i
]);
4618 if (passthrough_edgeflags
)
4622 /* Declare address register.
4624 if (program
->num_address_regs
> 0) {
4625 assert(program
->num_address_regs
== 1);
4626 t
->address
[0] = ureg_DECL_address(ureg
);
4629 /* Declare misc input registers
4632 GLbitfield sysInputs
= proginfo
->SystemValuesRead
;
4633 unsigned numSys
= 0;
4634 for (i
= 0; sysInputs
; i
++) {
4635 if (sysInputs
& (1 << i
)) {
4636 unsigned semName
= mesa_sysval_to_semantic
[i
];
4637 t
->systemValues
[i
] = ureg_DECL_system_value(ureg
, numSys
, semName
, 0);
4638 if (semName
== TGSI_SEMANTIC_INSTANCEID
||
4639 semName
== TGSI_SEMANTIC_VERTEXID
) {
4640 /* From Gallium perspective, these system values are always
4641 * integer, and require native integer support. However, if
4642 * native integer is supported on the vertex stage but not the
4643 * pixel stage (e.g, i915g + draw), Mesa will generate IR that
4644 * assumes these system values are floats. To resolve the
4645 * inconsistency, we insert a U2F.
4647 struct st_context
*st
= st_context(ctx
);
4648 struct pipe_screen
*pscreen
= st
->pipe
->screen
;
4649 assert(procType
== TGSI_PROCESSOR_VERTEX
);
4650 assert(pscreen
->get_shader_param(pscreen
, PIPE_SHADER_VERTEX
, PIPE_SHADER_CAP_INTEGERS
));
4651 if (!ctx
->Const
.NativeIntegers
) {
4652 struct ureg_dst temp
= ureg_DECL_local_temporary(t
->ureg
);
4653 ureg_U2F( t
->ureg
, ureg_writemask(temp
, TGSI_WRITEMASK_X
), t
->systemValues
[i
]);
4654 t
->systemValues
[i
] = ureg_scalar(ureg_src(temp
), 0);
4658 sysInputs
&= ~(1 << i
);
4663 if (program
->indirect_addr_temps
) {
4664 /* If temps are accessed with indirect addressing, declare temporaries
4665 * in sequential order. Else, we declare them on demand elsewhere.
4666 * (Note: the number of temporaries is equal to program->next_temp)
4668 for (i
= 0; i
< (unsigned)program
->next_temp
; i
++) {
4669 /* XXX use TGSI_FILE_TEMPORARY_ARRAY when it's supported by ureg */
4670 t
->temps
[i
] = ureg_DECL_local_temporary(t
->ureg
);
4674 /* Emit constants and uniforms. TGSI uses a single index space for these,
4675 * so we put all the translated regs in t->constants.
4677 if (proginfo
->Parameters
) {
4678 t
->constants
= (struct ureg_src
*)CALLOC(proginfo
->Parameters
->NumParameters
* sizeof(t
->constants
[0]));
4679 if (t
->constants
== NULL
) {
4680 ret
= PIPE_ERROR_OUT_OF_MEMORY
;
4684 for (i
= 0; i
< proginfo
->Parameters
->NumParameters
; i
++) {
4685 switch (proginfo
->Parameters
->Parameters
[i
].Type
) {
4686 case PROGRAM_ENV_PARAM
:
4687 case PROGRAM_LOCAL_PARAM
:
4688 case PROGRAM_STATE_VAR
:
4689 case PROGRAM_NAMED_PARAM
:
4690 case PROGRAM_UNIFORM
:
4691 t
->constants
[i
] = ureg_DECL_constant(ureg
, i
);
4694 /* Emit immediates for PROGRAM_CONSTANT only when there's no indirect
4695 * addressing of the const buffer.
4696 * FIXME: Be smarter and recognize param arrays:
4697 * indirect addressing is only valid within the referenced
4700 case PROGRAM_CONSTANT
:
4701 if (program
->indirect_addr_consts
)
4702 t
->constants
[i
] = ureg_DECL_constant(ureg
, i
);
4704 t
->constants
[i
] = emit_immediate(t
,
4705 proginfo
->Parameters
->ParameterValues
[i
],
4706 proginfo
->Parameters
->Parameters
[i
].DataType
,
4715 /* Emit immediate values.
4717 t
->immediates
= (struct ureg_src
*)CALLOC(program
->num_immediates
* sizeof(struct ureg_src
));
4718 if (t
->immediates
== NULL
) {
4719 ret
= PIPE_ERROR_OUT_OF_MEMORY
;
4723 foreach_iter(exec_list_iterator
, iter
, program
->immediates
) {
4724 immediate_storage
*imm
= (immediate_storage
*)iter
.get();
4725 assert(i
< program
->num_immediates
);
4726 t
->immediates
[i
++] = emit_immediate(t
, imm
->values
, imm
->type
, imm
->size
);
4728 assert(i
== program
->num_immediates
);
4730 /* texture samplers */
4731 for (i
= 0; i
< ctx
->Const
.MaxTextureImageUnits
; i
++) {
4732 if (program
->samplers_used
& (1 << i
)) {
4733 t
->samplers
[i
] = ureg_DECL_sampler(ureg
, i
);
4737 /* Emit each instruction in turn:
4739 foreach_iter(exec_list_iterator
, iter
, program
->instructions
) {
4740 set_insn_start(t
, ureg_get_instruction_number(ureg
));
4741 compile_tgsi_instruction(t
, (glsl_to_tgsi_instruction
*)iter
.get(),
4745 /* Fix up all emitted labels:
4747 for (i
= 0; i
< t
->labels_count
; i
++) {
4748 ureg_fixup_label(ureg
, t
->labels
[i
].token
,
4749 t
->insn
[t
->labels
[i
].branch_target
]);
4752 if (program
->shader_program
) {
4753 /* This has to be done last. Any operation the can cause
4754 * prog->ParameterValues to get reallocated (e.g., anything that adds a
4755 * program constant) has to happen before creating this linkage.
4757 for (unsigned i
= 0; i
< MESA_SHADER_TYPES
; i
++) {
4758 if (program
->shader_program
->_LinkedShaders
[i
] == NULL
)
4761 _mesa_associate_uniform_storage(ctx
, program
->shader_program
,
4762 program
->shader_program
->_LinkedShaders
[i
]->Program
->Parameters
);
4771 FREE(t
->immediates
);
4774 debug_printf("%s: translate error flag set\n", __FUNCTION__
);
4782 /* ----------------------------- End TGSI code ------------------------------ */
4785 * Convert a shader's GLSL IR into a Mesa gl_program, although without
4786 * generating Mesa IR.
4788 static struct gl_program
*
4789 get_mesa_program(struct gl_context
*ctx
,
4790 struct gl_shader_program
*shader_program
,
4791 struct gl_shader
*shader
)
4793 glsl_to_tgsi_visitor
* v
= new glsl_to_tgsi_visitor();
4794 struct gl_program
*prog
;
4796 const char *target_string
;
4798 struct gl_shader_compiler_options
*options
=
4799 &ctx
->ShaderCompilerOptions
[_mesa_shader_type_to_index(shader
->Type
)];
4801 switch (shader
->Type
) {
4802 case GL_VERTEX_SHADER
:
4803 target
= GL_VERTEX_PROGRAM_ARB
;
4804 target_string
= "vertex";
4806 case GL_FRAGMENT_SHADER
:
4807 target
= GL_FRAGMENT_PROGRAM_ARB
;
4808 target_string
= "fragment";
4810 case GL_GEOMETRY_SHADER
:
4811 target
= GL_GEOMETRY_PROGRAM_NV
;
4812 target_string
= "geometry";
4815 assert(!"should not be reached");
4819 validate_ir_tree(shader
->ir
);
4821 prog
= ctx
->Driver
.NewProgram(ctx
, target
, shader_program
->Name
);
4824 prog
->Parameters
= _mesa_new_parameter_list();
4827 v
->shader_program
= shader_program
;
4828 v
->options
= options
;
4829 v
->glsl_version
= ctx
->Const
.GLSLVersion
;
4830 v
->native_integers
= ctx
->Const
.NativeIntegers
;
4832 _mesa_generate_parameters_list_for_uniforms(shader_program
, shader
,
4835 /* Remove reads from output registers. */
4836 lower_output_reads(shader
->ir
);
4838 /* Emit intermediate IR for main(). */
4839 visit_exec_list(shader
->ir
, v
);
4841 /* Now emit bodies for any functions that were used. */
4843 progress
= GL_FALSE
;
4845 foreach_iter(exec_list_iterator
, iter
, v
->function_signatures
) {
4846 function_entry
*entry
= (function_entry
*)iter
.get();
4848 if (!entry
->bgn_inst
) {
4849 v
->current_function
= entry
;
4851 entry
->bgn_inst
= v
->emit(NULL
, TGSI_OPCODE_BGNSUB
);
4852 entry
->bgn_inst
->function
= entry
;
4854 visit_exec_list(&entry
->sig
->body
, v
);
4856 glsl_to_tgsi_instruction
*last
;
4857 last
= (glsl_to_tgsi_instruction
*)v
->instructions
.get_tail();
4858 if (last
->op
!= TGSI_OPCODE_RET
)
4859 v
->emit(NULL
, TGSI_OPCODE_RET
);
4861 glsl_to_tgsi_instruction
*end
;
4862 end
= v
->emit(NULL
, TGSI_OPCODE_ENDSUB
);
4863 end
->function
= entry
;
4871 /* Print out some information (for debugging purposes) used by the
4872 * optimization passes. */
4873 for (i
=0; i
< v
->next_temp
; i
++) {
4874 int fr
= v
->get_first_temp_read(i
);
4875 int fw
= v
->get_first_temp_write(i
);
4876 int lr
= v
->get_last_temp_read(i
);
4877 int lw
= v
->get_last_temp_write(i
);
4879 printf("Temp %d: FR=%3d FW=%3d LR=%3d LW=%3d\n", i
, fr
, fw
, lr
, lw
);
4884 /* Perform optimizations on the instructions in the glsl_to_tgsi_visitor. */
4886 v
->copy_propagate();
4887 while (v
->eliminate_dead_code_advanced());
4889 /* FIXME: These passes to optimize temporary registers don't work when there
4890 * is indirect addressing of the temporary register space. We need proper
4891 * array support so that we don't have to give up these passes in every
4892 * shader that uses arrays.
4894 if (!v
->indirect_addr_temps
) {
4895 v
->eliminate_dead_code();
4896 v
->merge_registers();
4897 v
->renumber_registers();
4900 /* Write the END instruction. */
4901 v
->emit(NULL
, TGSI_OPCODE_END
);
4903 if (ctx
->Shader
.Flags
& GLSL_DUMP
) {
4905 printf("GLSL IR for linked %s program %d:\n", target_string
,
4906 shader_program
->Name
);
4907 _mesa_print_ir(shader
->ir
, NULL
);
4913 prog
->Instructions
= NULL
;
4914 prog
->NumInstructions
= 0;
4916 do_set_program_inouts(shader
->ir
, prog
, shader
->Type
== GL_FRAGMENT_SHADER
);
4917 count_resources(v
, prog
);
4919 _mesa_reference_program(ctx
, &shader
->Program
, prog
);
4921 /* This has to be done last. Any operation the can cause
4922 * prog->ParameterValues to get reallocated (e.g., anything that adds a
4923 * program constant) has to happen before creating this linkage.
4925 _mesa_associate_uniform_storage(ctx
, shader_program
, prog
->Parameters
);
4926 if (!shader_program
->LinkStatus
) {
4930 struct st_vertex_program
*stvp
;
4931 struct st_fragment_program
*stfp
;
4932 struct st_geometry_program
*stgp
;
4934 switch (shader
->Type
) {
4935 case GL_VERTEX_SHADER
:
4936 stvp
= (struct st_vertex_program
*)prog
;
4937 stvp
->glsl_to_tgsi
= v
;
4939 case GL_FRAGMENT_SHADER
:
4940 stfp
= (struct st_fragment_program
*)prog
;
4941 stfp
->glsl_to_tgsi
= v
;
4943 case GL_GEOMETRY_SHADER
:
4944 stgp
= (struct st_geometry_program
*)prog
;
4945 stgp
->glsl_to_tgsi
= v
;
4948 assert(!"should not be reached");
4958 st_new_shader(struct gl_context
*ctx
, GLuint name
, GLuint type
)
4960 struct gl_shader
*shader
;
4961 assert(type
== GL_FRAGMENT_SHADER
|| type
== GL_VERTEX_SHADER
||
4962 type
== GL_GEOMETRY_SHADER_ARB
);
4963 shader
= rzalloc(NULL
, struct gl_shader
);
4965 shader
->Type
= type
;
4966 shader
->Name
= name
;
4967 _mesa_init_shader(ctx
, shader
);
4972 struct gl_shader_program
*
4973 st_new_shader_program(struct gl_context
*ctx
, GLuint name
)
4975 struct gl_shader_program
*shProg
;
4976 shProg
= rzalloc(NULL
, struct gl_shader_program
);
4978 shProg
->Name
= name
;
4979 _mesa_init_shader_program(ctx
, shProg
);
4986 * Called via ctx->Driver.LinkShader()
4987 * This actually involves converting GLSL IR into an intermediate TGSI-like IR
4988 * with code lowering and other optimizations.
4991 st_link_shader(struct gl_context
*ctx
, struct gl_shader_program
*prog
)
4993 assert(prog
->LinkStatus
);
4995 for (unsigned i
= 0; i
< MESA_SHADER_TYPES
; i
++) {
4996 if (prog
->_LinkedShaders
[i
] == NULL
)
5000 exec_list
*ir
= prog
->_LinkedShaders
[i
]->ir
;
5001 const struct gl_shader_compiler_options
*options
=
5002 &ctx
->ShaderCompilerOptions
[_mesa_shader_type_to_index(prog
->_LinkedShaders
[i
]->Type
)];
5005 unsigned what_to_lower
= MOD_TO_FRACT
| DIV_TO_MUL_RCP
|
5006 EXP_TO_EXP2
| LOG_TO_LOG2
;
5007 if (options
->EmitNoPow
)
5008 what_to_lower
|= POW_TO_EXP2
;
5009 if (!ctx
->Const
.NativeIntegers
)
5010 what_to_lower
|= INT_DIV_TO_MUL_RCP
;
5015 do_mat_op_to_vec(ir
);
5016 lower_instructions(ir
, what_to_lower
);
5018 progress
= do_lower_jumps(ir
, true, true, options
->EmitNoMainReturn
, options
->EmitNoCont
, options
->EmitNoLoops
) || progress
;
5020 progress
= do_common_optimization(ir
, true, true,
5021 options
->MaxUnrollIterations
)
5024 progress
= lower_quadop_vector(ir
, false) || progress
;
5026 if (options
->MaxIfDepth
== 0)
5027 progress
= lower_discard(ir
) || progress
;
5029 progress
= lower_if_to_cond_assign(ir
, options
->MaxIfDepth
) || progress
;
5031 if (options
->EmitNoNoise
)
5032 progress
= lower_noise(ir
) || progress
;
5034 /* If there are forms of indirect addressing that the driver
5035 * cannot handle, perform the lowering pass.
5037 if (options
->EmitNoIndirectInput
|| options
->EmitNoIndirectOutput
5038 || options
->EmitNoIndirectTemp
|| options
->EmitNoIndirectUniform
)
5040 lower_variable_index_to_cond_assign(ir
,
5041 options
->EmitNoIndirectInput
,
5042 options
->EmitNoIndirectOutput
,
5043 options
->EmitNoIndirectTemp
,
5044 options
->EmitNoIndirectUniform
)
5047 progress
= do_vec_index_to_cond_assign(ir
) || progress
;
5050 validate_ir_tree(ir
);
5053 for (unsigned i
= 0; i
< MESA_SHADER_TYPES
; i
++) {
5054 struct gl_program
*linked_prog
;
5056 if (prog
->_LinkedShaders
[i
] == NULL
)
5059 linked_prog
= get_mesa_program(ctx
, prog
, prog
->_LinkedShaders
[i
]);
5062 static const GLenum targets
[] = {
5063 GL_VERTEX_PROGRAM_ARB
,
5064 GL_FRAGMENT_PROGRAM_ARB
,
5065 GL_GEOMETRY_PROGRAM_NV
5068 _mesa_reference_program(ctx
, &prog
->_LinkedShaders
[i
]->Program
,
5070 if (!ctx
->Driver
.ProgramStringNotify(ctx
, targets
[i
], linked_prog
)) {
5071 _mesa_reference_program(ctx
, &prog
->_LinkedShaders
[i
]->Program
,
5073 _mesa_reference_program(ctx
, &linked_prog
, NULL
);
5078 _mesa_reference_program(ctx
, &linked_prog
, NULL
);
5085 st_translate_stream_output_info(glsl_to_tgsi_visitor
*glsl_to_tgsi
,
5086 const GLuint outputMapping
[],
5087 struct pipe_stream_output_info
*so
)
5090 struct gl_transform_feedback_info
*info
=
5091 &glsl_to_tgsi
->shader_program
->LinkedTransformFeedback
;
5093 for (i
= 0; i
< info
->NumOutputs
; i
++) {
5094 so
->output
[i
].register_index
=
5095 outputMapping
[info
->Outputs
[i
].OutputRegister
];
5096 so
->output
[i
].start_component
= info
->Outputs
[i
].ComponentOffset
;
5097 so
->output
[i
].num_components
= info
->Outputs
[i
].NumComponents
;
5098 so
->output
[i
].output_buffer
= info
->Outputs
[i
].OutputBuffer
;
5099 so
->output
[i
].dst_offset
= info
->Outputs
[i
].DstOffset
;
5102 for (i
= 0; i
< PIPE_MAX_SO_BUFFERS
; i
++) {
5103 so
->stride
[i
] = info
->BufferStride
[i
];
5105 so
->num_outputs
= info
->NumOutputs
;