2 * Copyright (C) 2005-2007 Brian Paul All Rights Reserved.
3 * Copyright (C) 2008 VMware, Inc. All Rights Reserved.
4 * Copyright © 2010 Intel Corporation
5 * Copyright © 2011 Bryan Cain
7 * Permission is hereby granted, free of charge, to any person obtaining a
8 * copy of this software and associated documentation files (the "Software"),
9 * to deal in the Software without restriction, including without limitation
10 * the rights to use, copy, modify, merge, publish, distribute, sublicense,
11 * and/or sell copies of the Software, and to permit persons to whom the
12 * Software is furnished to do so, subject to the following conditions:
14 * The above copyright notice and this permission notice (including the next
15 * paragraph) shall be included in all copies or substantial portions of the
18 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
19 * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
20 * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL
21 * THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
22 * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING
23 * FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER
24 * DEALINGS IN THE SOFTWARE.
28 * \file glsl_to_tgsi.cpp
30 * Translate GLSL IR to TGSI.
34 #include "main/compiler.h"
36 #include "ir_visitor.h"
37 #include "ir_print_visitor.h"
38 #include "ir_expression_flattening.h"
39 #include "glsl_types.h"
40 #include "glsl_parser_extras.h"
41 #include "../glsl/program.h"
42 #include "ir_optimization.h"
45 #include "main/mtypes.h"
46 #include "main/shaderobj.h"
47 #include "program/hash_table.h"
50 #include "main/shaderapi.h"
51 #include "main/uniforms.h"
52 #include "program/prog_instruction.h"
53 #include "program/prog_optimize.h"
54 #include "program/prog_print.h"
55 #include "program/program.h"
56 #include "program/prog_parameter.h"
57 #include "program/sampler.h"
59 #include "pipe/p_compiler.h"
60 #include "pipe/p_context.h"
61 #include "pipe/p_screen.h"
62 #include "pipe/p_shader_tokens.h"
63 #include "pipe/p_state.h"
64 #include "util/u_math.h"
65 #include "tgsi/tgsi_ureg.h"
66 #include "tgsi/tgsi_info.h"
67 #include "st_context.h"
68 #include "st_program.h"
69 #include "st_glsl_to_tgsi.h"
70 #include "st_mesa_to_tgsi.h"
73 #define PROGRAM_IMMEDIATE PROGRAM_FILE_MAX
74 #define PROGRAM_ANY_CONST ((1 << PROGRAM_LOCAL_PARAM) | \
75 (1 << PROGRAM_ENV_PARAM) | \
76 (1 << PROGRAM_STATE_VAR) | \
77 (1 << PROGRAM_CONSTANT) | \
78 (1 << PROGRAM_UNIFORM))
81 * Maximum number of temporary registers.
83 * It is too big for stack allocated arrays -- it will cause stack overflow on
84 * Windows and likely Mac OS X.
86 #define MAX_TEMPS 4096
88 /* will be 4 for GLSL 4.00 */
89 #define MAX_GLSL_TEXTURE_OFFSET 1
94 static int swizzle_for_size(int size
);
97 * This struct is a corresponding struct to TGSI ureg_src.
101 st_src_reg(gl_register_file file
, int index
, const glsl_type
*type
)
105 if (type
&& (type
->is_scalar() || type
->is_vector() || type
->is_matrix()))
106 this->swizzle
= swizzle_for_size(type
->vector_elements
);
108 this->swizzle
= SWIZZLE_XYZW
;
111 this->type
= type
? type
->base_type
: GLSL_TYPE_ERROR
;
112 this->reladdr
= NULL
;
115 st_src_reg(gl_register_file file
, int index
, int type
)
121 this->swizzle
= SWIZZLE_XYZW
;
123 this->reladdr
= NULL
;
126 st_src_reg(gl_register_file file
, int index
, int type
, int index2D
)
131 this->index2D
= index2D
;
132 this->swizzle
= SWIZZLE_XYZW
;
134 this->reladdr
= NULL
;
139 this->type
= GLSL_TYPE_ERROR
;
140 this->file
= PROGRAM_UNDEFINED
;
145 this->reladdr
= NULL
;
148 explicit st_src_reg(st_dst_reg reg
);
150 gl_register_file file
; /**< PROGRAM_* from Mesa */
151 int index
; /**< temporary index, VERT_ATTRIB_*, FRAG_ATTRIB_*, etc. */
153 GLuint swizzle
; /**< SWIZZLE_XYZWONEZERO swizzles from Mesa. */
154 int negate
; /**< NEGATE_XYZW mask from mesa */
155 int type
; /** GLSL_TYPE_* from GLSL IR (enum glsl_base_type) */
156 /** Register index should be offset by the integer in this reg. */
162 st_dst_reg(gl_register_file file
, int writemask
, int type
)
166 this->writemask
= writemask
;
167 this->cond_mask
= COND_TR
;
168 this->reladdr
= NULL
;
174 this->type
= GLSL_TYPE_ERROR
;
175 this->file
= PROGRAM_UNDEFINED
;
178 this->cond_mask
= COND_TR
;
179 this->reladdr
= NULL
;
182 explicit st_dst_reg(st_src_reg reg
);
184 gl_register_file file
; /**< PROGRAM_* from Mesa */
185 int index
; /**< temporary index, VERT_ATTRIB_*, FRAG_ATTRIB_*, etc. */
186 int writemask
; /**< Bitfield of WRITEMASK_[XYZW] */
188 int type
; /** GLSL_TYPE_* from GLSL IR (enum glsl_base_type) */
189 /** Register index should be offset by the integer in this reg. */
193 st_src_reg::st_src_reg(st_dst_reg reg
)
195 this->type
= reg
.type
;
196 this->file
= reg
.file
;
197 this->index
= reg
.index
;
198 this->swizzle
= SWIZZLE_XYZW
;
200 this->reladdr
= reg
.reladdr
;
204 st_dst_reg::st_dst_reg(st_src_reg reg
)
206 this->type
= reg
.type
;
207 this->file
= reg
.file
;
208 this->index
= reg
.index
;
209 this->writemask
= WRITEMASK_XYZW
;
210 this->cond_mask
= COND_TR
;
211 this->reladdr
= reg
.reladdr
;
214 class glsl_to_tgsi_instruction
: public exec_node
{
216 /* Callers of this ralloc-based new need not call delete. It's
217 * easier to just ralloc_free 'ctx' (or any of its ancestors). */
218 static void* operator new(size_t size
, void *ctx
)
222 node
= rzalloc_size(ctx
, size
);
223 assert(node
!= NULL
);
231 /** Pointer to the ir source this tree came from for debugging */
233 GLboolean cond_update
;
235 int sampler
; /**< sampler index */
236 int tex_target
; /**< One of TEXTURE_*_INDEX */
237 GLboolean tex_shadow
;
238 struct tgsi_texture_offset tex_offsets
[MAX_GLSL_TEXTURE_OFFSET
];
239 unsigned tex_offset_num_offset
;
240 int dead_mask
; /**< Used in dead code elimination */
242 class function_entry
*function
; /* Set on TGSI_OPCODE_CAL or TGSI_OPCODE_BGNSUB */
245 class variable_storage
: public exec_node
{
247 variable_storage(ir_variable
*var
, gl_register_file file
, int index
)
248 : file(file
), index(index
), var(var
)
253 gl_register_file file
;
255 ir_variable
*var
; /* variable that maps to this, if any */
258 class immediate_storage
: public exec_node
{
260 immediate_storage(gl_constant_value
*values
, int size
, int type
)
262 memcpy(this->values
, values
, size
* sizeof(gl_constant_value
));
267 gl_constant_value values
[4];
268 int size
; /**< Number of components (1-4) */
269 int type
; /**< GL_FLOAT, GL_INT, GL_BOOL, or GL_UNSIGNED_INT */
272 class function_entry
: public exec_node
{
274 ir_function_signature
*sig
;
277 * identifier of this function signature used by the program.
279 * At the point that TGSI instructions for function calls are
280 * generated, we don't know the address of the first instruction of
281 * the function body. So we make the BranchTarget that is called a
282 * small integer and rewrite them during set_branchtargets().
287 * Pointer to first instruction of the function body.
289 * Set during function body emits after main() is processed.
291 glsl_to_tgsi_instruction
*bgn_inst
;
294 * Index of the first instruction of the function body in actual TGSI.
296 * Set after conversion from glsl_to_tgsi_instruction to TGSI.
300 /** Storage for the return value. */
301 st_src_reg return_reg
;
304 struct glsl_to_tgsi_visitor
: public ir_visitor
{
306 glsl_to_tgsi_visitor();
307 ~glsl_to_tgsi_visitor();
309 function_entry
*current_function
;
311 struct gl_context
*ctx
;
312 struct gl_program
*prog
;
313 struct gl_shader_program
*shader_program
;
314 struct gl_shader_compiler_options
*options
;
318 int num_address_regs
;
320 bool indirect_addr_temps
;
321 bool indirect_addr_consts
;
324 bool native_integers
;
326 variable_storage
*find_variable_storage(ir_variable
*var
);
328 int add_constant(gl_register_file file
, gl_constant_value values
[4],
329 int size
, int datatype
, GLuint
*swizzle_out
);
331 function_entry
*get_function_signature(ir_function_signature
*sig
);
333 st_src_reg
get_temp(const glsl_type
*type
);
334 void reladdr_to_temp(ir_instruction
*ir
, st_src_reg
*reg
, int *num_reladdr
);
336 st_src_reg
st_src_reg_for_float(float val
);
337 st_src_reg
st_src_reg_for_int(int val
);
338 st_src_reg
st_src_reg_for_type(int type
, int val
);
341 * \name Visit methods
343 * As typical for the visitor pattern, there must be one \c visit method for
344 * each concrete subclass of \c ir_instruction. Virtual base classes within
345 * the hierarchy should not have \c visit methods.
348 virtual void visit(ir_variable
*);
349 virtual void visit(ir_loop
*);
350 virtual void visit(ir_loop_jump
*);
351 virtual void visit(ir_function_signature
*);
352 virtual void visit(ir_function
*);
353 virtual void visit(ir_expression
*);
354 virtual void visit(ir_swizzle
*);
355 virtual void visit(ir_dereference_variable
*);
356 virtual void visit(ir_dereference_array
*);
357 virtual void visit(ir_dereference_record
*);
358 virtual void visit(ir_assignment
*);
359 virtual void visit(ir_constant
*);
360 virtual void visit(ir_call
*);
361 virtual void visit(ir_return
*);
362 virtual void visit(ir_discard
*);
363 virtual void visit(ir_texture
*);
364 virtual void visit(ir_if
*);
369 /** List of variable_storage */
372 /** List of immediate_storage */
373 exec_list immediates
;
374 unsigned num_immediates
;
376 /** List of function_entry */
377 exec_list function_signatures
;
378 int next_signature_id
;
380 /** List of glsl_to_tgsi_instruction */
381 exec_list instructions
;
383 glsl_to_tgsi_instruction
*emit(ir_instruction
*ir
, unsigned op
);
385 glsl_to_tgsi_instruction
*emit(ir_instruction
*ir
, unsigned op
,
386 st_dst_reg dst
, st_src_reg src0
);
388 glsl_to_tgsi_instruction
*emit(ir_instruction
*ir
, unsigned op
,
389 st_dst_reg dst
, st_src_reg src0
, st_src_reg src1
);
391 glsl_to_tgsi_instruction
*emit(ir_instruction
*ir
, unsigned op
,
393 st_src_reg src0
, st_src_reg src1
, st_src_reg src2
);
395 unsigned get_opcode(ir_instruction
*ir
, unsigned op
,
397 st_src_reg src0
, st_src_reg src1
);
400 * Emit the correct dot-product instruction for the type of arguments
402 glsl_to_tgsi_instruction
*emit_dp(ir_instruction
*ir
,
408 void emit_scalar(ir_instruction
*ir
, unsigned op
,
409 st_dst_reg dst
, st_src_reg src0
);
411 void emit_scalar(ir_instruction
*ir
, unsigned op
,
412 st_dst_reg dst
, st_src_reg src0
, st_src_reg src1
);
414 void try_emit_float_set(ir_instruction
*ir
, unsigned op
, st_dst_reg dst
);
416 void emit_arl(ir_instruction
*ir
, st_dst_reg dst
, st_src_reg src0
);
418 void emit_scs(ir_instruction
*ir
, unsigned op
,
419 st_dst_reg dst
, const st_src_reg
&src
);
421 bool try_emit_mad(ir_expression
*ir
,
423 bool try_emit_mad_for_and_not(ir_expression
*ir
,
425 bool try_emit_sat(ir_expression
*ir
);
427 void emit_swz(ir_expression
*ir
);
429 bool process_move_condition(ir_rvalue
*ir
);
431 void simplify_cmp(void);
433 void rename_temp_register(int index
, int new_index
);
434 int get_first_temp_read(int index
);
435 int get_first_temp_write(int index
);
436 int get_last_temp_read(int index
);
437 int get_last_temp_write(int index
);
439 void copy_propagate(void);
440 void eliminate_dead_code(void);
441 int eliminate_dead_code_advanced(void);
442 void merge_registers(void);
443 void renumber_registers(void);
445 void emit_block_mov(ir_assignment
*ir
, const struct glsl_type
*type
,
446 st_dst_reg
*l
, st_src_reg
*r
);
451 static st_src_reg undef_src
= st_src_reg(PROGRAM_UNDEFINED
, 0, GLSL_TYPE_ERROR
);
453 static st_dst_reg undef_dst
= st_dst_reg(PROGRAM_UNDEFINED
, SWIZZLE_NOOP
, GLSL_TYPE_ERROR
);
455 static st_dst_reg address_reg
= st_dst_reg(PROGRAM_ADDRESS
, WRITEMASK_X
, GLSL_TYPE_FLOAT
);
458 fail_link(struct gl_shader_program
*prog
, const char *fmt
, ...) PRINTFLIKE(2, 3);
461 fail_link(struct gl_shader_program
*prog
, const char *fmt
, ...)
465 ralloc_vasprintf_append(&prog
->InfoLog
, fmt
, args
);
468 prog
->LinkStatus
= GL_FALSE
;
472 swizzle_for_size(int size
)
474 int size_swizzles
[4] = {
475 MAKE_SWIZZLE4(SWIZZLE_X
, SWIZZLE_X
, SWIZZLE_X
, SWIZZLE_X
),
476 MAKE_SWIZZLE4(SWIZZLE_X
, SWIZZLE_Y
, SWIZZLE_Y
, SWIZZLE_Y
),
477 MAKE_SWIZZLE4(SWIZZLE_X
, SWIZZLE_Y
, SWIZZLE_Z
, SWIZZLE_Z
),
478 MAKE_SWIZZLE4(SWIZZLE_X
, SWIZZLE_Y
, SWIZZLE_Z
, SWIZZLE_W
),
481 assert((size
>= 1) && (size
<= 4));
482 return size_swizzles
[size
- 1];
486 is_tex_instruction(unsigned opcode
)
488 const tgsi_opcode_info
* info
= tgsi_get_opcode_info(opcode
);
493 num_inst_dst_regs(unsigned opcode
)
495 const tgsi_opcode_info
* info
= tgsi_get_opcode_info(opcode
);
496 return info
->num_dst
;
500 num_inst_src_regs(unsigned opcode
)
502 const tgsi_opcode_info
* info
= tgsi_get_opcode_info(opcode
);
503 return info
->is_tex
? info
->num_src
- 1 : info
->num_src
;
506 glsl_to_tgsi_instruction
*
507 glsl_to_tgsi_visitor::emit(ir_instruction
*ir
, unsigned op
,
509 st_src_reg src0
, st_src_reg src1
, st_src_reg src2
)
511 glsl_to_tgsi_instruction
*inst
= new(mem_ctx
) glsl_to_tgsi_instruction();
512 int num_reladdr
= 0, i
;
514 op
= get_opcode(ir
, op
, dst
, src0
, src1
);
516 /* If we have to do relative addressing, we want to load the ARL
517 * reg directly for one of the regs, and preload the other reladdr
518 * sources into temps.
520 num_reladdr
+= dst
.reladdr
!= NULL
;
521 num_reladdr
+= src0
.reladdr
!= NULL
;
522 num_reladdr
+= src1
.reladdr
!= NULL
;
523 num_reladdr
+= src2
.reladdr
!= NULL
;
525 reladdr_to_temp(ir
, &src2
, &num_reladdr
);
526 reladdr_to_temp(ir
, &src1
, &num_reladdr
);
527 reladdr_to_temp(ir
, &src0
, &num_reladdr
);
530 emit_arl(ir
, address_reg
, *dst
.reladdr
);
533 assert(num_reladdr
== 0);
543 inst
->function
= NULL
;
545 if (op
== TGSI_OPCODE_ARL
|| op
== TGSI_OPCODE_UARL
)
546 this->num_address_regs
= 1;
548 /* Update indirect addressing status used by TGSI */
551 case PROGRAM_TEMPORARY
:
552 this->indirect_addr_temps
= true;
554 case PROGRAM_LOCAL_PARAM
:
555 case PROGRAM_ENV_PARAM
:
556 case PROGRAM_STATE_VAR
:
557 case PROGRAM_CONSTANT
:
558 case PROGRAM_UNIFORM
:
559 this->indirect_addr_consts
= true;
561 case PROGRAM_IMMEDIATE
:
562 assert(!"immediates should not have indirect addressing");
569 for (i
=0; i
<3; i
++) {
570 if(inst
->src
[i
].reladdr
) {
571 switch(inst
->src
[i
].file
) {
572 case PROGRAM_TEMPORARY
:
573 this->indirect_addr_temps
= true;
575 case PROGRAM_LOCAL_PARAM
:
576 case PROGRAM_ENV_PARAM
:
577 case PROGRAM_STATE_VAR
:
578 case PROGRAM_CONSTANT
:
579 case PROGRAM_UNIFORM
:
580 this->indirect_addr_consts
= true;
582 case PROGRAM_IMMEDIATE
:
583 assert(!"immediates should not have indirect addressing");
592 this->instructions
.push_tail(inst
);
595 try_emit_float_set(ir
, op
, dst
);
601 glsl_to_tgsi_instruction
*
602 glsl_to_tgsi_visitor::emit(ir_instruction
*ir
, unsigned op
,
603 st_dst_reg dst
, st_src_reg src0
, st_src_reg src1
)
605 return emit(ir
, op
, dst
, src0
, src1
, undef_src
);
608 glsl_to_tgsi_instruction
*
609 glsl_to_tgsi_visitor::emit(ir_instruction
*ir
, unsigned op
,
610 st_dst_reg dst
, st_src_reg src0
)
612 assert(dst
.writemask
!= 0);
613 return emit(ir
, op
, dst
, src0
, undef_src
, undef_src
);
616 glsl_to_tgsi_instruction
*
617 glsl_to_tgsi_visitor::emit(ir_instruction
*ir
, unsigned op
)
619 return emit(ir
, op
, undef_dst
, undef_src
, undef_src
, undef_src
);
623 * Emits the code to convert the result of float SET instructions to integers.
626 glsl_to_tgsi_visitor::try_emit_float_set(ir_instruction
*ir
, unsigned op
,
629 if ((op
== TGSI_OPCODE_SEQ
||
630 op
== TGSI_OPCODE_SNE
||
631 op
== TGSI_OPCODE_SGE
||
632 op
== TGSI_OPCODE_SLT
))
634 st_src_reg src
= st_src_reg(dst
);
635 src
.negate
= ~src
.negate
;
636 dst
.type
= GLSL_TYPE_FLOAT
;
637 emit(ir
, TGSI_OPCODE_F2I
, dst
, src
);
642 * Determines whether to use an integer, unsigned integer, or float opcode
643 * based on the operands and input opcode, then emits the result.
646 glsl_to_tgsi_visitor::get_opcode(ir_instruction
*ir
, unsigned op
,
648 st_src_reg src0
, st_src_reg src1
)
650 int type
= GLSL_TYPE_FLOAT
;
652 assert(src0
.type
!= GLSL_TYPE_ARRAY
);
653 assert(src0
.type
!= GLSL_TYPE_STRUCT
);
654 assert(src1
.type
!= GLSL_TYPE_ARRAY
);
655 assert(src1
.type
!= GLSL_TYPE_STRUCT
);
657 if (src0
.type
== GLSL_TYPE_FLOAT
|| src1
.type
== GLSL_TYPE_FLOAT
)
658 type
= GLSL_TYPE_FLOAT
;
659 else if (native_integers
)
660 type
= src0
.type
== GLSL_TYPE_BOOL
? GLSL_TYPE_INT
: src0
.type
;
662 #define case4(c, f, i, u) \
663 case TGSI_OPCODE_##c: \
664 if (type == GLSL_TYPE_INT) op = TGSI_OPCODE_##i; \
665 else if (type == GLSL_TYPE_UINT) op = TGSI_OPCODE_##u; \
666 else op = TGSI_OPCODE_##f; \
668 #define case3(f, i, u) case4(f, f, i, u)
669 #define case2fi(f, i) case4(f, f, i, i)
670 #define case2iu(i, u) case4(i, LAST, i, u)
676 case3(DIV
, IDIV
, UDIV
);
677 case3(MAX
, IMAX
, UMAX
);
678 case3(MIN
, IMIN
, UMIN
);
683 case3(SGE
, ISGE
, USGE
);
684 case3(SLT
, ISLT
, USLT
);
689 case3(ABS
, IABS
, IABS
);
694 assert(op
!= TGSI_OPCODE_LAST
);
698 glsl_to_tgsi_instruction
*
699 glsl_to_tgsi_visitor::emit_dp(ir_instruction
*ir
,
700 st_dst_reg dst
, st_src_reg src0
, st_src_reg src1
,
703 static const unsigned dot_opcodes
[] = {
704 TGSI_OPCODE_DP2
, TGSI_OPCODE_DP3
, TGSI_OPCODE_DP4
707 return emit(ir
, dot_opcodes
[elements
- 2], dst
, src0
, src1
);
711 * Emits TGSI scalar opcodes to produce unique answers across channels.
713 * Some TGSI opcodes are scalar-only, like ARB_fp/vp. The src X
714 * channel determines the result across all channels. So to do a vec4
715 * of this operation, we want to emit a scalar per source channel used
716 * to produce dest channels.
719 glsl_to_tgsi_visitor::emit_scalar(ir_instruction
*ir
, unsigned op
,
721 st_src_reg orig_src0
, st_src_reg orig_src1
)
724 int done_mask
= ~dst
.writemask
;
726 /* TGSI RCP is a scalar operation splatting results to all channels,
727 * like ARB_fp/vp. So emit as many RCPs as necessary to cover our
730 for (i
= 0; i
< 4; i
++) {
731 GLuint this_mask
= (1 << i
);
732 glsl_to_tgsi_instruction
*inst
;
733 st_src_reg src0
= orig_src0
;
734 st_src_reg src1
= orig_src1
;
736 if (done_mask
& this_mask
)
739 GLuint src0_swiz
= GET_SWZ(src0
.swizzle
, i
);
740 GLuint src1_swiz
= GET_SWZ(src1
.swizzle
, i
);
741 for (j
= i
+ 1; j
< 4; j
++) {
742 /* If there is another enabled component in the destination that is
743 * derived from the same inputs, generate its value on this pass as
746 if (!(done_mask
& (1 << j
)) &&
747 GET_SWZ(src0
.swizzle
, j
) == src0_swiz
&&
748 GET_SWZ(src1
.swizzle
, j
) == src1_swiz
) {
749 this_mask
|= (1 << j
);
752 src0
.swizzle
= MAKE_SWIZZLE4(src0_swiz
, src0_swiz
,
753 src0_swiz
, src0_swiz
);
754 src1
.swizzle
= MAKE_SWIZZLE4(src1_swiz
, src1_swiz
,
755 src1_swiz
, src1_swiz
);
757 inst
= emit(ir
, op
, dst
, src0
, src1
);
758 inst
->dst
.writemask
= this_mask
;
759 done_mask
|= this_mask
;
764 glsl_to_tgsi_visitor::emit_scalar(ir_instruction
*ir
, unsigned op
,
765 st_dst_reg dst
, st_src_reg src0
)
767 st_src_reg undef
= undef_src
;
769 undef
.swizzle
= SWIZZLE_XXXX
;
771 emit_scalar(ir
, op
, dst
, src0
, undef
);
775 glsl_to_tgsi_visitor::emit_arl(ir_instruction
*ir
,
776 st_dst_reg dst
, st_src_reg src0
)
778 int op
= TGSI_OPCODE_ARL
;
780 if (src0
.type
== GLSL_TYPE_INT
|| src0
.type
== GLSL_TYPE_UINT
)
781 op
= TGSI_OPCODE_UARL
;
783 emit(NULL
, op
, dst
, src0
);
787 * Emit an TGSI_OPCODE_SCS instruction
789 * The \c SCS opcode functions a bit differently than the other TGSI opcodes.
790 * Instead of splatting its result across all four components of the
791 * destination, it writes one value to the \c x component and another value to
792 * the \c y component.
794 * \param ir IR instruction being processed
795 * \param op Either \c TGSI_OPCODE_SIN or \c TGSI_OPCODE_COS depending
796 * on which value is desired.
797 * \param dst Destination register
798 * \param src Source register
801 glsl_to_tgsi_visitor::emit_scs(ir_instruction
*ir
, unsigned op
,
803 const st_src_reg
&src
)
805 /* Vertex programs cannot use the SCS opcode.
807 if (this->prog
->Target
== GL_VERTEX_PROGRAM_ARB
) {
808 emit_scalar(ir
, op
, dst
, src
);
812 const unsigned component
= (op
== TGSI_OPCODE_SIN
) ? 0 : 1;
813 const unsigned scs_mask
= (1U << component
);
814 int done_mask
= ~dst
.writemask
;
817 assert(op
== TGSI_OPCODE_SIN
|| op
== TGSI_OPCODE_COS
);
819 /* If there are compnents in the destination that differ from the component
820 * that will be written by the SCS instrution, we'll need a temporary.
822 if (scs_mask
!= unsigned(dst
.writemask
)) {
823 tmp
= get_temp(glsl_type::vec4_type
);
826 for (unsigned i
= 0; i
< 4; i
++) {
827 unsigned this_mask
= (1U << i
);
828 st_src_reg src0
= src
;
830 if ((done_mask
& this_mask
) != 0)
833 /* The source swizzle specified which component of the source generates
834 * sine / cosine for the current component in the destination. The SCS
835 * instruction requires that this value be swizzle to the X component.
836 * Replace the current swizzle with a swizzle that puts the source in
839 unsigned src0_swiz
= GET_SWZ(src
.swizzle
, i
);
841 src0
.swizzle
= MAKE_SWIZZLE4(src0_swiz
, src0_swiz
,
842 src0_swiz
, src0_swiz
);
843 for (unsigned j
= i
+ 1; j
< 4; j
++) {
844 /* If there is another enabled component in the destination that is
845 * derived from the same inputs, generate its value on this pass as
848 if (!(done_mask
& (1 << j
)) &&
849 GET_SWZ(src0
.swizzle
, j
) == src0_swiz
) {
850 this_mask
|= (1 << j
);
854 if (this_mask
!= scs_mask
) {
855 glsl_to_tgsi_instruction
*inst
;
856 st_dst_reg tmp_dst
= st_dst_reg(tmp
);
858 /* Emit the SCS instruction.
860 inst
= emit(ir
, TGSI_OPCODE_SCS
, tmp_dst
, src0
);
861 inst
->dst
.writemask
= scs_mask
;
863 /* Move the result of the SCS instruction to the desired location in
866 tmp
.swizzle
= MAKE_SWIZZLE4(component
, component
,
867 component
, component
);
868 inst
= emit(ir
, TGSI_OPCODE_SCS
, dst
, tmp
);
869 inst
->dst
.writemask
= this_mask
;
871 /* Emit the SCS instruction to write directly to the destination.
873 glsl_to_tgsi_instruction
*inst
= emit(ir
, TGSI_OPCODE_SCS
, dst
, src0
);
874 inst
->dst
.writemask
= scs_mask
;
877 done_mask
|= this_mask
;
882 glsl_to_tgsi_visitor::add_constant(gl_register_file file
,
883 gl_constant_value values
[4], int size
, int datatype
,
886 if (file
== PROGRAM_CONSTANT
) {
887 return _mesa_add_typed_unnamed_constant(this->prog
->Parameters
, values
,
888 size
, datatype
, swizzle_out
);
891 immediate_storage
*entry
;
892 assert(file
== PROGRAM_IMMEDIATE
);
894 /* Search immediate storage to see if we already have an identical
895 * immediate that we can use instead of adding a duplicate entry.
897 foreach_iter(exec_list_iterator
, iter
, this->immediates
) {
898 entry
= (immediate_storage
*)iter
.get();
900 if (entry
->size
== size
&&
901 entry
->type
== datatype
&&
902 !memcmp(entry
->values
, values
, size
* sizeof(gl_constant_value
))) {
908 /* Add this immediate to the list. */
909 entry
= new(mem_ctx
) immediate_storage(values
, size
, datatype
);
910 this->immediates
.push_tail(entry
);
911 this->num_immediates
++;
917 glsl_to_tgsi_visitor::st_src_reg_for_float(float val
)
919 st_src_reg
src(PROGRAM_IMMEDIATE
, -1, GLSL_TYPE_FLOAT
);
920 union gl_constant_value uval
;
923 src
.index
= add_constant(src
.file
, &uval
, 1, GL_FLOAT
, &src
.swizzle
);
929 glsl_to_tgsi_visitor::st_src_reg_for_int(int val
)
931 st_src_reg
src(PROGRAM_IMMEDIATE
, -1, GLSL_TYPE_INT
);
932 union gl_constant_value uval
;
934 assert(native_integers
);
937 src
.index
= add_constant(src
.file
, &uval
, 1, GL_INT
, &src
.swizzle
);
943 glsl_to_tgsi_visitor::st_src_reg_for_type(int type
, int val
)
946 return type
== GLSL_TYPE_FLOAT
? st_src_reg_for_float(val
) :
947 st_src_reg_for_int(val
);
949 return st_src_reg_for_float(val
);
953 type_size(const struct glsl_type
*type
)
958 switch (type
->base_type
) {
961 case GLSL_TYPE_FLOAT
:
963 if (type
->is_matrix()) {
964 return type
->matrix_columns
;
966 /* Regardless of size of vector, it gets a vec4. This is bad
967 * packing for things like floats, but otherwise arrays become a
968 * mess. Hopefully a later pass over the code can pack scalars
969 * down if appropriate.
973 case GLSL_TYPE_ARRAY
:
974 assert(type
->length
> 0);
975 return type_size(type
->fields
.array
) * type
->length
;
976 case GLSL_TYPE_STRUCT
:
978 for (i
= 0; i
< type
->length
; i
++) {
979 size
+= type_size(type
->fields
.structure
[i
].type
);
982 case GLSL_TYPE_SAMPLER
:
983 /* Samplers take up one slot in UNIFORMS[], but they're baked in
994 * In the initial pass of codegen, we assign temporary numbers to
995 * intermediate results. (not SSA -- variable assignments will reuse
999 glsl_to_tgsi_visitor::get_temp(const glsl_type
*type
)
1003 src
.type
= native_integers
? type
->base_type
: GLSL_TYPE_FLOAT
;
1004 src
.file
= PROGRAM_TEMPORARY
;
1005 src
.index
= next_temp
;
1007 next_temp
+= type_size(type
);
1009 if (type
->is_array() || type
->is_record()) {
1010 src
.swizzle
= SWIZZLE_NOOP
;
1012 src
.swizzle
= swizzle_for_size(type
->vector_elements
);
1020 glsl_to_tgsi_visitor::find_variable_storage(ir_variable
*var
)
1023 variable_storage
*entry
;
1025 foreach_iter(exec_list_iterator
, iter
, this->variables
) {
1026 entry
= (variable_storage
*)iter
.get();
1028 if (entry
->var
== var
)
1036 glsl_to_tgsi_visitor::visit(ir_variable
*ir
)
1038 if (strcmp(ir
->name
, "gl_FragCoord") == 0) {
1039 struct gl_fragment_program
*fp
= (struct gl_fragment_program
*)this->prog
;
1041 fp
->OriginUpperLeft
= ir
->origin_upper_left
;
1042 fp
->PixelCenterInteger
= ir
->pixel_center_integer
;
1045 if (ir
->mode
== ir_var_uniform
&& strncmp(ir
->name
, "gl_", 3) == 0) {
1047 const ir_state_slot
*const slots
= ir
->state_slots
;
1048 assert(ir
->state_slots
!= NULL
);
1050 /* Check if this statevar's setup in the STATE file exactly
1051 * matches how we'll want to reference it as a
1052 * struct/array/whatever. If not, then we need to move it into
1053 * temporary storage and hope that it'll get copy-propagated
1056 for (i
= 0; i
< ir
->num_state_slots
; i
++) {
1057 if (slots
[i
].swizzle
!= SWIZZLE_XYZW
) {
1062 variable_storage
*storage
;
1064 if (i
== ir
->num_state_slots
) {
1065 /* We'll set the index later. */
1066 storage
= new(mem_ctx
) variable_storage(ir
, PROGRAM_STATE_VAR
, -1);
1067 this->variables
.push_tail(storage
);
1071 /* The variable_storage constructor allocates slots based on the size
1072 * of the type. However, this had better match the number of state
1073 * elements that we're going to copy into the new temporary.
1075 assert((int) ir
->num_state_slots
== type_size(ir
->type
));
1077 storage
= new(mem_ctx
) variable_storage(ir
, PROGRAM_TEMPORARY
,
1079 this->variables
.push_tail(storage
);
1080 this->next_temp
+= type_size(ir
->type
);
1082 dst
= st_dst_reg(st_src_reg(PROGRAM_TEMPORARY
, storage
->index
,
1083 native_integers
? ir
->type
->base_type
: GLSL_TYPE_FLOAT
));
1087 for (unsigned int i
= 0; i
< ir
->num_state_slots
; i
++) {
1088 int index
= _mesa_add_state_reference(this->prog
->Parameters
,
1089 (gl_state_index
*)slots
[i
].tokens
);
1091 if (storage
->file
== PROGRAM_STATE_VAR
) {
1092 if (storage
->index
== -1) {
1093 storage
->index
= index
;
1095 assert(index
== storage
->index
+ (int)i
);
1098 /* We use GLSL_TYPE_FLOAT here regardless of the actual type of
1099 * the data being moved since MOV does not care about the type of
1100 * data it is moving, and we don't want to declare registers with
1101 * array or struct types.
1103 st_src_reg
src(PROGRAM_STATE_VAR
, index
, GLSL_TYPE_FLOAT
);
1104 src
.swizzle
= slots
[i
].swizzle
;
1105 emit(ir
, TGSI_OPCODE_MOV
, dst
, src
);
1106 /* even a float takes up a whole vec4 reg in a struct/array. */
1111 if (storage
->file
== PROGRAM_TEMPORARY
&&
1112 dst
.index
!= storage
->index
+ (int) ir
->num_state_slots
) {
1113 fail_link(this->shader_program
,
1114 "failed to load builtin uniform `%s' (%d/%d regs loaded)\n",
1115 ir
->name
, dst
.index
- storage
->index
,
1116 type_size(ir
->type
));
1122 glsl_to_tgsi_visitor::visit(ir_loop
*ir
)
1124 ir_dereference_variable
*counter
= NULL
;
1126 if (ir
->counter
!= NULL
)
1127 counter
= new(ir
) ir_dereference_variable(ir
->counter
);
1129 if (ir
->from
!= NULL
) {
1130 assert(ir
->counter
!= NULL
);
1132 ir_assignment
*a
= new(ir
) ir_assignment(counter
, ir
->from
, NULL
);
1138 emit(NULL
, TGSI_OPCODE_BGNLOOP
);
1142 new(ir
) ir_expression(ir
->cmp
, glsl_type::bool_type
,
1144 ir_if
*if_stmt
= new(ir
) ir_if(e
);
1146 ir_loop_jump
*brk
= new(ir
) ir_loop_jump(ir_loop_jump::jump_break
);
1148 if_stmt
->then_instructions
.push_tail(brk
);
1150 if_stmt
->accept(this);
1157 visit_exec_list(&ir
->body_instructions
, this);
1159 if (ir
->increment
) {
1161 new(ir
) ir_expression(ir_binop_add
, counter
->type
,
1162 counter
, ir
->increment
);
1164 ir_assignment
*a
= new(ir
) ir_assignment(counter
, e
, NULL
);
1171 emit(NULL
, TGSI_OPCODE_ENDLOOP
);
1175 glsl_to_tgsi_visitor::visit(ir_loop_jump
*ir
)
1178 case ir_loop_jump::jump_break
:
1179 emit(NULL
, TGSI_OPCODE_BRK
);
1181 case ir_loop_jump::jump_continue
:
1182 emit(NULL
, TGSI_OPCODE_CONT
);
1189 glsl_to_tgsi_visitor::visit(ir_function_signature
*ir
)
1196 glsl_to_tgsi_visitor::visit(ir_function
*ir
)
1198 /* Ignore function bodies other than main() -- we shouldn't see calls to
1199 * them since they should all be inlined before we get to glsl_to_tgsi.
1201 if (strcmp(ir
->name
, "main") == 0) {
1202 const ir_function_signature
*sig
;
1205 sig
= ir
->matching_signature(&empty
);
1209 foreach_iter(exec_list_iterator
, iter
, sig
->body
) {
1210 ir_instruction
*ir
= (ir_instruction
*)iter
.get();
1218 glsl_to_tgsi_visitor::try_emit_mad(ir_expression
*ir
, int mul_operand
)
1220 int nonmul_operand
= 1 - mul_operand
;
1222 st_dst_reg result_dst
;
1224 ir_expression
*expr
= ir
->operands
[mul_operand
]->as_expression();
1225 if (!expr
|| expr
->operation
!= ir_binop_mul
)
1228 expr
->operands
[0]->accept(this);
1230 expr
->operands
[1]->accept(this);
1232 ir
->operands
[nonmul_operand
]->accept(this);
1235 this->result
= get_temp(ir
->type
);
1236 result_dst
= st_dst_reg(this->result
);
1237 result_dst
.writemask
= (1 << ir
->type
->vector_elements
) - 1;
1238 emit(ir
, TGSI_OPCODE_MAD
, result_dst
, a
, b
, c
);
1244 * Emit MAD(a, -b, a) instead of AND(a, NOT(b))
1246 * The logic values are 1.0 for true and 0.0 for false. Logical-and is
1247 * implemented using multiplication, and logical-or is implemented using
1248 * addition. Logical-not can be implemented as (true - x), or (1.0 - x).
1249 * As result, the logical expression (a & !b) can be rewritten as:
1253 * - (a * 1) - (a * b)
1257 * This final expression can be implemented as a single MAD(a, -b, a)
1261 glsl_to_tgsi_visitor::try_emit_mad_for_and_not(ir_expression
*ir
, int try_operand
)
1263 const int other_operand
= 1 - try_operand
;
1266 ir_expression
*expr
= ir
->operands
[try_operand
]->as_expression();
1267 if (!expr
|| expr
->operation
!= ir_unop_logic_not
)
1270 ir
->operands
[other_operand
]->accept(this);
1272 expr
->operands
[0]->accept(this);
1275 b
.negate
= ~b
.negate
;
1277 this->result
= get_temp(ir
->type
);
1278 emit(ir
, TGSI_OPCODE_MAD
, st_dst_reg(this->result
), a
, b
, a
);
1284 glsl_to_tgsi_visitor::try_emit_sat(ir_expression
*ir
)
1286 /* Saturates were only introduced to vertex programs in
1287 * NV_vertex_program3, so don't give them to drivers in the VP.
1289 if (this->prog
->Target
== GL_VERTEX_PROGRAM_ARB
)
1292 ir_rvalue
*sat_src
= ir
->as_rvalue_to_saturate();
1296 sat_src
->accept(this);
1297 st_src_reg src
= this->result
;
1299 /* If we generated an expression instruction into a temporary in
1300 * processing the saturate's operand, apply the saturate to that
1301 * instruction. Otherwise, generate a MOV to do the saturate.
1303 * Note that we have to be careful to only do this optimization if
1304 * the instruction in question was what generated src->result. For
1305 * example, ir_dereference_array might generate a MUL instruction
1306 * to create the reladdr, and return us a src reg using that
1307 * reladdr. That MUL result is not the value we're trying to
1310 ir_expression
*sat_src_expr
= sat_src
->as_expression();
1311 if (sat_src_expr
&& (sat_src_expr
->operation
== ir_binop_mul
||
1312 sat_src_expr
->operation
== ir_binop_add
||
1313 sat_src_expr
->operation
== ir_binop_dot
)) {
1314 glsl_to_tgsi_instruction
*new_inst
;
1315 new_inst
= (glsl_to_tgsi_instruction
*)this->instructions
.get_tail();
1316 new_inst
->saturate
= true;
1318 this->result
= get_temp(ir
->type
);
1319 st_dst_reg result_dst
= st_dst_reg(this->result
);
1320 result_dst
.writemask
= (1 << ir
->type
->vector_elements
) - 1;
1321 glsl_to_tgsi_instruction
*inst
;
1322 inst
= emit(ir
, TGSI_OPCODE_MOV
, result_dst
, src
);
1323 inst
->saturate
= true;
1330 glsl_to_tgsi_visitor::reladdr_to_temp(ir_instruction
*ir
,
1331 st_src_reg
*reg
, int *num_reladdr
)
1336 emit_arl(ir
, address_reg
, *reg
->reladdr
);
1338 if (*num_reladdr
!= 1) {
1339 st_src_reg temp
= get_temp(glsl_type::vec4_type
);
1341 emit(ir
, TGSI_OPCODE_MOV
, st_dst_reg(temp
), *reg
);
1349 glsl_to_tgsi_visitor::visit(ir_expression
*ir
)
1351 unsigned int operand
;
1352 st_src_reg op
[Elements(ir
->operands
)];
1353 st_src_reg result_src
;
1354 st_dst_reg result_dst
;
1356 /* Quick peephole: Emit MAD(a, b, c) instead of ADD(MUL(a, b), c)
1358 if (ir
->operation
== ir_binop_add
) {
1359 if (try_emit_mad(ir
, 1))
1361 if (try_emit_mad(ir
, 0))
1365 /* Quick peephole: Emit OPCODE_MAD(-a, -b, a) instead of AND(a, NOT(b))
1367 if (ir
->operation
== ir_binop_logic_and
) {
1368 if (try_emit_mad_for_and_not(ir
, 1))
1370 if (try_emit_mad_for_and_not(ir
, 0))
1374 if (try_emit_sat(ir
))
1377 if (ir
->operation
== ir_quadop_vector
)
1378 assert(!"ir_quadop_vector should have been lowered");
1380 for (operand
= 0; operand
< ir
->get_num_operands(); operand
++) {
1381 this->result
.file
= PROGRAM_UNDEFINED
;
1382 ir
->operands
[operand
]->accept(this);
1383 if (this->result
.file
== PROGRAM_UNDEFINED
) {
1385 printf("Failed to get tree for expression operand:\n");
1386 ir
->operands
[operand
]->accept(&v
);
1389 op
[operand
] = this->result
;
1391 /* Matrix expression operands should have been broken down to vector
1392 * operations already.
1394 assert(!ir
->operands
[operand
]->type
->is_matrix());
1397 int vector_elements
= ir
->operands
[0]->type
->vector_elements
;
1398 if (ir
->operands
[1]) {
1399 vector_elements
= MAX2(vector_elements
,
1400 ir
->operands
[1]->type
->vector_elements
);
1403 this->result
.file
= PROGRAM_UNDEFINED
;
1405 /* Storage for our result. Ideally for an assignment we'd be using
1406 * the actual storage for the result here, instead.
1408 result_src
= get_temp(ir
->type
);
1409 /* convenience for the emit functions below. */
1410 result_dst
= st_dst_reg(result_src
);
1411 /* Limit writes to the channels that will be used by result_src later.
1412 * This does limit this temp's use as a temporary for multi-instruction
1415 result_dst
.writemask
= (1 << ir
->type
->vector_elements
) - 1;
1417 switch (ir
->operation
) {
1418 case ir_unop_logic_not
:
1419 if (result_dst
.type
!= GLSL_TYPE_FLOAT
)
1420 emit(ir
, TGSI_OPCODE_NOT
, result_dst
, op
[0]);
1422 /* Previously 'SEQ dst, src, 0.0' was used for this. However, many
1423 * older GPUs implement SEQ using multiple instructions (i915 uses two
1424 * SGE instructions and a MUL instruction). Since our logic values are
1425 * 0.0 and 1.0, 1-x also implements !x.
1427 op
[0].negate
= ~op
[0].negate
;
1428 emit(ir
, TGSI_OPCODE_ADD
, result_dst
, op
[0], st_src_reg_for_float(1.0));
1432 if (result_dst
.type
== GLSL_TYPE_INT
|| result_dst
.type
== GLSL_TYPE_UINT
)
1433 emit(ir
, TGSI_OPCODE_INEG
, result_dst
, op
[0]);
1435 op
[0].negate
= ~op
[0].negate
;
1440 emit(ir
, TGSI_OPCODE_ABS
, result_dst
, op
[0]);
1443 emit(ir
, TGSI_OPCODE_SSG
, result_dst
, op
[0]);
1446 emit_scalar(ir
, TGSI_OPCODE_RCP
, result_dst
, op
[0]);
1450 emit_scalar(ir
, TGSI_OPCODE_EX2
, result_dst
, op
[0]);
1454 assert(!"not reached: should be handled by ir_explog_to_explog2");
1457 emit_scalar(ir
, TGSI_OPCODE_LG2
, result_dst
, op
[0]);
1460 emit_scalar(ir
, TGSI_OPCODE_SIN
, result_dst
, op
[0]);
1463 emit_scalar(ir
, TGSI_OPCODE_COS
, result_dst
, op
[0]);
1465 case ir_unop_sin_reduced
:
1466 emit_scs(ir
, TGSI_OPCODE_SIN
, result_dst
, op
[0]);
1468 case ir_unop_cos_reduced
:
1469 emit_scs(ir
, TGSI_OPCODE_COS
, result_dst
, op
[0]);
1473 emit(ir
, TGSI_OPCODE_DDX
, result_dst
, op
[0]);
1477 /* The X component contains 1 or -1 depending on whether the framebuffer
1478 * is a FBO or the window system buffer, respectively.
1479 * It is then multiplied with the source operand of DDY.
1481 static const gl_state_index transform_y_state
[STATE_LENGTH
]
1482 = { STATE_INTERNAL
, STATE_FB_WPOS_Y_TRANSFORM
};
1484 unsigned transform_y_index
=
1485 _mesa_add_state_reference(this->prog
->Parameters
,
1488 st_src_reg transform_y
= st_src_reg(PROGRAM_STATE_VAR
,
1490 glsl_type::vec4_type
);
1491 transform_y
.swizzle
= SWIZZLE_XXXX
;
1493 st_src_reg temp
= get_temp(glsl_type::vec4_type
);
1495 emit(ir
, TGSI_OPCODE_MUL
, st_dst_reg(temp
), transform_y
, op
[0]);
1496 emit(ir
, TGSI_OPCODE_DDY
, result_dst
, temp
);
1500 case ir_unop_noise
: {
1501 /* At some point, a motivated person could add a better
1502 * implementation of noise. Currently not even the nvidia
1503 * binary drivers do anything more than this. In any case, the
1504 * place to do this is in the GL state tracker, not the poor
1507 emit(ir
, TGSI_OPCODE_MOV
, result_dst
, st_src_reg_for_float(0.5));
1512 emit(ir
, TGSI_OPCODE_ADD
, result_dst
, op
[0], op
[1]);
1515 emit(ir
, TGSI_OPCODE_SUB
, result_dst
, op
[0], op
[1]);
1519 emit(ir
, TGSI_OPCODE_MUL
, result_dst
, op
[0], op
[1]);
1522 if (result_dst
.type
== GLSL_TYPE_FLOAT
)
1523 assert(!"not reached: should be handled by ir_div_to_mul_rcp");
1525 emit(ir
, TGSI_OPCODE_DIV
, result_dst
, op
[0], op
[1]);
1528 if (result_dst
.type
== GLSL_TYPE_FLOAT
)
1529 assert(!"ir_binop_mod should have been converted to b * fract(a/b)");
1531 emit(ir
, TGSI_OPCODE_MOD
, result_dst
, op
[0], op
[1]);
1535 emit(ir
, TGSI_OPCODE_SLT
, result_dst
, op
[0], op
[1]);
1537 case ir_binop_greater
:
1538 emit(ir
, TGSI_OPCODE_SLT
, result_dst
, op
[1], op
[0]);
1540 case ir_binop_lequal
:
1541 emit(ir
, TGSI_OPCODE_SGE
, result_dst
, op
[1], op
[0]);
1543 case ir_binop_gequal
:
1544 emit(ir
, TGSI_OPCODE_SGE
, result_dst
, op
[0], op
[1]);
1546 case ir_binop_equal
:
1547 emit(ir
, TGSI_OPCODE_SEQ
, result_dst
, op
[0], op
[1]);
1549 case ir_binop_nequal
:
1550 emit(ir
, TGSI_OPCODE_SNE
, result_dst
, op
[0], op
[1]);
1552 case ir_binop_all_equal
:
1553 /* "==" operator producing a scalar boolean. */
1554 if (ir
->operands
[0]->type
->is_vector() ||
1555 ir
->operands
[1]->type
->is_vector()) {
1556 st_src_reg temp
= get_temp(native_integers
?
1557 glsl_type::get_instance(ir
->operands
[0]->type
->base_type
, 4, 1) :
1558 glsl_type::vec4_type
);
1560 if (native_integers
) {
1561 st_dst_reg temp_dst
= st_dst_reg(temp
);
1562 st_src_reg temp1
= st_src_reg(temp
), temp2
= st_src_reg(temp
);
1564 emit(ir
, TGSI_OPCODE_SEQ
, st_dst_reg(temp
), op
[0], op
[1]);
1566 /* Emit 1-3 AND operations to combine the SEQ results. */
1567 switch (ir
->operands
[0]->type
->vector_elements
) {
1571 temp_dst
.writemask
= WRITEMASK_Y
;
1572 temp1
.swizzle
= SWIZZLE_YYYY
;
1573 temp2
.swizzle
= SWIZZLE_ZZZZ
;
1574 emit(ir
, TGSI_OPCODE_AND
, temp_dst
, temp1
, temp2
);
1577 temp_dst
.writemask
= WRITEMASK_X
;
1578 temp1
.swizzle
= SWIZZLE_XXXX
;
1579 temp2
.swizzle
= SWIZZLE_YYYY
;
1580 emit(ir
, TGSI_OPCODE_AND
, temp_dst
, temp1
, temp2
);
1581 temp_dst
.writemask
= WRITEMASK_Y
;
1582 temp1
.swizzle
= SWIZZLE_ZZZZ
;
1583 temp2
.swizzle
= SWIZZLE_WWWW
;
1584 emit(ir
, TGSI_OPCODE_AND
, temp_dst
, temp1
, temp2
);
1587 temp1
.swizzle
= SWIZZLE_XXXX
;
1588 temp2
.swizzle
= SWIZZLE_YYYY
;
1589 emit(ir
, TGSI_OPCODE_AND
, result_dst
, temp1
, temp2
);
1591 emit(ir
, TGSI_OPCODE_SNE
, st_dst_reg(temp
), op
[0], op
[1]);
1593 /* After the dot-product, the value will be an integer on the
1594 * range [0,4]. Zero becomes 1.0, and positive values become zero.
1596 emit_dp(ir
, result_dst
, temp
, temp
, vector_elements
);
1598 /* Negating the result of the dot-product gives values on the range
1599 * [-4, 0]. Zero becomes 1.0, and negative values become zero.
1600 * This is achieved using SGE.
1602 st_src_reg sge_src
= result_src
;
1603 sge_src
.negate
= ~sge_src
.negate
;
1604 emit(ir
, TGSI_OPCODE_SGE
, result_dst
, sge_src
, st_src_reg_for_float(0.0));
1607 emit(ir
, TGSI_OPCODE_SEQ
, result_dst
, op
[0], op
[1]);
1610 case ir_binop_any_nequal
:
1611 /* "!=" operator producing a scalar boolean. */
1612 if (ir
->operands
[0]->type
->is_vector() ||
1613 ir
->operands
[1]->type
->is_vector()) {
1614 st_src_reg temp
= get_temp(native_integers
?
1615 glsl_type::get_instance(ir
->operands
[0]->type
->base_type
, 4, 1) :
1616 glsl_type::vec4_type
);
1617 emit(ir
, TGSI_OPCODE_SNE
, st_dst_reg(temp
), op
[0], op
[1]);
1619 if (native_integers
) {
1620 st_dst_reg temp_dst
= st_dst_reg(temp
);
1621 st_src_reg temp1
= st_src_reg(temp
), temp2
= st_src_reg(temp
);
1623 /* Emit 1-3 OR operations to combine the SNE results. */
1624 switch (ir
->operands
[0]->type
->vector_elements
) {
1628 temp_dst
.writemask
= WRITEMASK_Y
;
1629 temp1
.swizzle
= SWIZZLE_YYYY
;
1630 temp2
.swizzle
= SWIZZLE_ZZZZ
;
1631 emit(ir
, TGSI_OPCODE_OR
, temp_dst
, temp1
, temp2
);
1634 temp_dst
.writemask
= WRITEMASK_X
;
1635 temp1
.swizzle
= SWIZZLE_XXXX
;
1636 temp2
.swizzle
= SWIZZLE_YYYY
;
1637 emit(ir
, TGSI_OPCODE_OR
, temp_dst
, temp1
, temp2
);
1638 temp_dst
.writemask
= WRITEMASK_Y
;
1639 temp1
.swizzle
= SWIZZLE_ZZZZ
;
1640 temp2
.swizzle
= SWIZZLE_WWWW
;
1641 emit(ir
, TGSI_OPCODE_OR
, temp_dst
, temp1
, temp2
);
1644 temp1
.swizzle
= SWIZZLE_XXXX
;
1645 temp2
.swizzle
= SWIZZLE_YYYY
;
1646 emit(ir
, TGSI_OPCODE_OR
, result_dst
, temp1
, temp2
);
1648 /* After the dot-product, the value will be an integer on the
1649 * range [0,4]. Zero stays zero, and positive values become 1.0.
1651 glsl_to_tgsi_instruction
*const dp
=
1652 emit_dp(ir
, result_dst
, temp
, temp
, vector_elements
);
1653 if (this->prog
->Target
== GL_FRAGMENT_PROGRAM_ARB
) {
1654 /* The clamping to [0,1] can be done for free in the fragment
1655 * shader with a saturate.
1657 dp
->saturate
= true;
1659 /* Negating the result of the dot-product gives values on the range
1660 * [-4, 0]. Zero stays zero, and negative values become 1.0. This
1661 * achieved using SLT.
1663 st_src_reg slt_src
= result_src
;
1664 slt_src
.negate
= ~slt_src
.negate
;
1665 emit(ir
, TGSI_OPCODE_SLT
, result_dst
, slt_src
, st_src_reg_for_float(0.0));
1669 emit(ir
, TGSI_OPCODE_SNE
, result_dst
, op
[0], op
[1]);
1674 assert(ir
->operands
[0]->type
->is_vector());
1676 /* After the dot-product, the value will be an integer on the
1677 * range [0,4]. Zero stays zero, and positive values become 1.0.
1679 glsl_to_tgsi_instruction
*const dp
=
1680 emit_dp(ir
, result_dst
, op
[0], op
[0],
1681 ir
->operands
[0]->type
->vector_elements
);
1682 if (this->prog
->Target
== GL_FRAGMENT_PROGRAM_ARB
&&
1683 result_dst
.type
== GLSL_TYPE_FLOAT
) {
1684 /* The clamping to [0,1] can be done for free in the fragment
1685 * shader with a saturate.
1687 dp
->saturate
= true;
1688 } else if (result_dst
.type
== GLSL_TYPE_FLOAT
) {
1689 /* Negating the result of the dot-product gives values on the range
1690 * [-4, 0]. Zero stays zero, and negative values become 1.0. This
1691 * is achieved using SLT.
1693 st_src_reg slt_src
= result_src
;
1694 slt_src
.negate
= ~slt_src
.negate
;
1695 emit(ir
, TGSI_OPCODE_SLT
, result_dst
, slt_src
, st_src_reg_for_float(0.0));
1698 /* Use SNE 0 if integers are being used as boolean values. */
1699 emit(ir
, TGSI_OPCODE_SNE
, result_dst
, result_src
, st_src_reg_for_int(0));
1704 case ir_binop_logic_xor
:
1705 if (native_integers
)
1706 emit(ir
, TGSI_OPCODE_XOR
, result_dst
, op
[0], op
[1]);
1708 emit(ir
, TGSI_OPCODE_SNE
, result_dst
, op
[0], op
[1]);
1711 case ir_binop_logic_or
: {
1712 if (native_integers
) {
1713 /* If integers are used as booleans, we can use an actual "or"
1716 assert(native_integers
);
1717 emit(ir
, TGSI_OPCODE_OR
, result_dst
, op
[0], op
[1]);
1719 /* After the addition, the value will be an integer on the
1720 * range [0,2]. Zero stays zero, and positive values become 1.0.
1722 glsl_to_tgsi_instruction
*add
=
1723 emit(ir
, TGSI_OPCODE_ADD
, result_dst
, op
[0], op
[1]);
1724 if (this->prog
->Target
== GL_FRAGMENT_PROGRAM_ARB
) {
1725 /* The clamping to [0,1] can be done for free in the fragment
1726 * shader with a saturate if floats are being used as boolean values.
1728 add
->saturate
= true;
1730 /* Negating the result of the addition gives values on the range
1731 * [-2, 0]. Zero stays zero, and negative values become 1.0. This
1732 * is achieved using SLT.
1734 st_src_reg slt_src
= result_src
;
1735 slt_src
.negate
= ~slt_src
.negate
;
1736 emit(ir
, TGSI_OPCODE_SLT
, result_dst
, slt_src
, st_src_reg_for_float(0.0));
1742 case ir_binop_logic_and
:
1743 /* If native integers are disabled, the bool args are stored as float 0.0
1744 * or 1.0, so "mul" gives us "and". If they're enabled, just use the
1745 * actual AND opcode.
1747 if (native_integers
)
1748 emit(ir
, TGSI_OPCODE_AND
, result_dst
, op
[0], op
[1]);
1750 emit(ir
, TGSI_OPCODE_MUL
, result_dst
, op
[0], op
[1]);
1754 assert(ir
->operands
[0]->type
->is_vector());
1755 assert(ir
->operands
[0]->type
== ir
->operands
[1]->type
);
1756 emit_dp(ir
, result_dst
, op
[0], op
[1],
1757 ir
->operands
[0]->type
->vector_elements
);
1761 /* sqrt(x) = x * rsq(x). */
1762 emit_scalar(ir
, TGSI_OPCODE_RSQ
, result_dst
, op
[0]);
1763 emit(ir
, TGSI_OPCODE_MUL
, result_dst
, result_src
, op
[0]);
1764 /* For incoming channels <= 0, set the result to 0. */
1765 op
[0].negate
= ~op
[0].negate
;
1766 emit(ir
, TGSI_OPCODE_CMP
, result_dst
,
1767 op
[0], result_src
, st_src_reg_for_float(0.0));
1770 emit_scalar(ir
, TGSI_OPCODE_RSQ
, result_dst
, op
[0]);
1773 if (native_integers
) {
1774 emit(ir
, TGSI_OPCODE_I2F
, result_dst
, op
[0]);
1777 /* fallthrough to next case otherwise */
1779 if (native_integers
) {
1780 emit(ir
, TGSI_OPCODE_AND
, result_dst
, op
[0], st_src_reg_for_float(1.0));
1783 /* fallthrough to next case otherwise */
1786 /* Converting between signed and unsigned integers is a no-op. */
1790 if (native_integers
) {
1791 /* Booleans are stored as integers using ~0 for true and 0 for false.
1792 * GLSL requires that int(bool) return 1 for true and 0 for false.
1793 * This conversion is done with AND, but it could be done with NEG.
1795 emit(ir
, TGSI_OPCODE_AND
, result_dst
, op
[0], st_src_reg_for_int(1));
1797 /* Booleans and integers are both stored as floats when native
1798 * integers are disabled.
1804 if (native_integers
)
1805 emit(ir
, TGSI_OPCODE_F2I
, result_dst
, op
[0]);
1807 emit(ir
, TGSI_OPCODE_TRUNC
, result_dst
, op
[0]);
1810 if (native_integers
)
1811 emit(ir
, TGSI_OPCODE_F2U
, result_dst
, op
[0]);
1813 emit(ir
, TGSI_OPCODE_TRUNC
, result_dst
, op
[0]);
1815 case ir_unop_bitcast_f2i
:
1816 case ir_unop_bitcast_f2u
:
1817 case ir_unop_bitcast_i2f
:
1818 case ir_unop_bitcast_u2f
:
1822 emit(ir
, TGSI_OPCODE_SNE
, result_dst
, op
[0], st_src_reg_for_float(0.0));
1825 if (native_integers
)
1826 emit(ir
, TGSI_OPCODE_INEG
, result_dst
, op
[0]);
1828 emit(ir
, TGSI_OPCODE_SNE
, result_dst
, op
[0], st_src_reg_for_float(0.0));
1831 emit(ir
, TGSI_OPCODE_TRUNC
, result_dst
, op
[0]);
1834 emit(ir
, TGSI_OPCODE_CEIL
, result_dst
, op
[0]);
1837 emit(ir
, TGSI_OPCODE_FLR
, result_dst
, op
[0]);
1839 case ir_unop_round_even
:
1840 emit(ir
, TGSI_OPCODE_ROUND
, result_dst
, op
[0]);
1843 emit(ir
, TGSI_OPCODE_FRC
, result_dst
, op
[0]);
1847 emit(ir
, TGSI_OPCODE_MIN
, result_dst
, op
[0], op
[1]);
1850 emit(ir
, TGSI_OPCODE_MAX
, result_dst
, op
[0], op
[1]);
1853 emit_scalar(ir
, TGSI_OPCODE_POW
, result_dst
, op
[0], op
[1]);
1856 case ir_unop_bit_not
:
1857 if (native_integers
) {
1858 emit(ir
, TGSI_OPCODE_NOT
, result_dst
, op
[0]);
1862 if (native_integers
) {
1863 emit(ir
, TGSI_OPCODE_U2F
, result_dst
, op
[0]);
1866 case ir_binop_lshift
:
1867 if (native_integers
) {
1868 emit(ir
, TGSI_OPCODE_SHL
, result_dst
, op
[0], op
[1]);
1871 case ir_binop_rshift
:
1872 if (native_integers
) {
1873 emit(ir
, TGSI_OPCODE_ISHR
, result_dst
, op
[0], op
[1]);
1876 case ir_binop_bit_and
:
1877 if (native_integers
) {
1878 emit(ir
, TGSI_OPCODE_AND
, result_dst
, op
[0], op
[1]);
1881 case ir_binop_bit_xor
:
1882 if (native_integers
) {
1883 emit(ir
, TGSI_OPCODE_XOR
, result_dst
, op
[0], op
[1]);
1886 case ir_binop_bit_or
:
1887 if (native_integers
) {
1888 emit(ir
, TGSI_OPCODE_OR
, result_dst
, op
[0], op
[1]);
1892 assert(!"GLSL 1.30 features unsupported");
1895 case ir_binop_ubo_load
: {
1896 ir_constant
*uniform_block
= ir
->operands
[0]->as_constant();
1897 ir_constant
*const_offset_ir
= ir
->operands
[1]->as_constant();
1898 unsigned const_offset
= const_offset_ir
? const_offset_ir
->value
.u
[0] : 0;
1899 st_src_reg index_reg
= get_temp(glsl_type::uint_type
);
1902 cbuf
.type
= glsl_type::vec4_type
->base_type
;
1903 cbuf
.file
= PROGRAM_CONSTANT
;
1905 cbuf
.index2D
= uniform_block
->value
.u
[0] + 1;
1906 cbuf
.reladdr
= NULL
;
1909 assert(ir
->type
->is_vector() || ir
->type
->is_scalar());
1911 if (const_offset_ir
) {
1912 index_reg
= st_src_reg_for_int(const_offset
/ 16);
1914 emit(ir
, TGSI_OPCODE_USHR
, st_dst_reg(index_reg
), op
[1], st_src_reg_for_int(4));
1917 cbuf
.swizzle
= swizzle_for_size(ir
->type
->vector_elements
);
1918 cbuf
.swizzle
+= MAKE_SWIZZLE4(const_offset
% 16 / 4,
1919 const_offset
% 16 / 4,
1920 const_offset
% 16 / 4,
1921 const_offset
% 16 / 4);
1923 cbuf
.reladdr
= ralloc(mem_ctx
, st_src_reg
);
1924 memcpy(cbuf
.reladdr
, &index_reg
, sizeof(index_reg
));
1926 if (ir
->type
->base_type
== GLSL_TYPE_BOOL
) {
1927 emit(ir
, TGSI_OPCODE_USNE
, result_dst
, cbuf
, st_src_reg_for_int(0));
1928 result_src
.negate
= 1;
1929 emit(ir
, TGSI_OPCODE_UCMP
, result_dst
, result_src
, st_src_reg_for_int(~0), st_src_reg_for_int(0));
1931 emit(ir
, TGSI_OPCODE_MOV
, result_dst
, cbuf
);
1935 case ir_quadop_vector
:
1936 /* This operation should have already been handled.
1938 assert(!"Should not get here.");
1942 this->result
= result_src
;
1947 glsl_to_tgsi_visitor::visit(ir_swizzle
*ir
)
1953 /* Note that this is only swizzles in expressions, not those on the left
1954 * hand side of an assignment, which do write masking. See ir_assignment
1958 ir
->val
->accept(this);
1960 assert(src
.file
!= PROGRAM_UNDEFINED
);
1962 for (i
= 0; i
< 4; i
++) {
1963 if (i
< ir
->type
->vector_elements
) {
1966 swizzle
[i
] = GET_SWZ(src
.swizzle
, ir
->mask
.x
);
1969 swizzle
[i
] = GET_SWZ(src
.swizzle
, ir
->mask
.y
);
1972 swizzle
[i
] = GET_SWZ(src
.swizzle
, ir
->mask
.z
);
1975 swizzle
[i
] = GET_SWZ(src
.swizzle
, ir
->mask
.w
);
1979 /* If the type is smaller than a vec4, replicate the last
1982 swizzle
[i
] = swizzle
[ir
->type
->vector_elements
- 1];
1986 src
.swizzle
= MAKE_SWIZZLE4(swizzle
[0], swizzle
[1], swizzle
[2], swizzle
[3]);
1992 glsl_to_tgsi_visitor::visit(ir_dereference_variable
*ir
)
1994 variable_storage
*entry
= find_variable_storage(ir
->var
);
1995 ir_variable
*var
= ir
->var
;
1998 switch (var
->mode
) {
1999 case ir_var_uniform
:
2000 entry
= new(mem_ctx
) variable_storage(var
, PROGRAM_UNIFORM
,
2002 this->variables
.push_tail(entry
);
2004 case ir_var_shader_in
:
2005 /* The linker assigns locations for varyings and attributes,
2006 * including deprecated builtins (like gl_Color), user-assign
2007 * generic attributes (glBindVertexLocation), and
2008 * user-defined varyings.
2010 assert(var
->location
!= -1);
2011 entry
= new(mem_ctx
) variable_storage(var
,
2015 case ir_var_shader_out
:
2016 assert(var
->location
!= -1);
2017 entry
= new(mem_ctx
) variable_storage(var
,
2019 var
->location
+ var
->index
);
2021 case ir_var_system_value
:
2022 entry
= new(mem_ctx
) variable_storage(var
,
2023 PROGRAM_SYSTEM_VALUE
,
2027 case ir_var_temporary
:
2028 entry
= new(mem_ctx
) variable_storage(var
, PROGRAM_TEMPORARY
,
2030 this->variables
.push_tail(entry
);
2032 next_temp
+= type_size(var
->type
);
2037 printf("Failed to make storage for %s\n", var
->name
);
2042 this->result
= st_src_reg(entry
->file
, entry
->index
, var
->type
);
2043 if (!native_integers
)
2044 this->result
.type
= GLSL_TYPE_FLOAT
;
2048 glsl_to_tgsi_visitor::visit(ir_dereference_array
*ir
)
2052 int element_size
= type_size(ir
->type
);
2054 index
= ir
->array_index
->constant_expression_value();
2056 ir
->array
->accept(this);
2060 src
.index
+= index
->value
.i
[0] * element_size
;
2062 /* Variable index array dereference. It eats the "vec4" of the
2063 * base of the array and an index that offsets the TGSI register
2066 ir
->array_index
->accept(this);
2068 st_src_reg index_reg
;
2070 if (element_size
== 1) {
2071 index_reg
= this->result
;
2073 index_reg
= get_temp(native_integers
?
2074 glsl_type::int_type
: glsl_type::float_type
);
2076 emit(ir
, TGSI_OPCODE_MUL
, st_dst_reg(index_reg
),
2077 this->result
, st_src_reg_for_type(index_reg
.type
, element_size
));
2080 /* If there was already a relative address register involved, add the
2081 * new and the old together to get the new offset.
2083 if (src
.reladdr
!= NULL
) {
2084 st_src_reg accum_reg
= get_temp(native_integers
?
2085 glsl_type::int_type
: glsl_type::float_type
);
2087 emit(ir
, TGSI_OPCODE_ADD
, st_dst_reg(accum_reg
),
2088 index_reg
, *src
.reladdr
);
2090 index_reg
= accum_reg
;
2093 src
.reladdr
= ralloc(mem_ctx
, st_src_reg
);
2094 memcpy(src
.reladdr
, &index_reg
, sizeof(index_reg
));
2097 /* If the type is smaller than a vec4, replicate the last channel out. */
2098 if (ir
->type
->is_scalar() || ir
->type
->is_vector())
2099 src
.swizzle
= swizzle_for_size(ir
->type
->vector_elements
);
2101 src
.swizzle
= SWIZZLE_NOOP
;
2103 /* Change the register type to the element type of the array. */
2104 src
.type
= ir
->type
->base_type
;
2110 glsl_to_tgsi_visitor::visit(ir_dereference_record
*ir
)
2113 const glsl_type
*struct_type
= ir
->record
->type
;
2116 ir
->record
->accept(this);
2118 for (i
= 0; i
< struct_type
->length
; i
++) {
2119 if (strcmp(struct_type
->fields
.structure
[i
].name
, ir
->field
) == 0)
2121 offset
+= type_size(struct_type
->fields
.structure
[i
].type
);
2124 /* If the type is smaller than a vec4, replicate the last channel out. */
2125 if (ir
->type
->is_scalar() || ir
->type
->is_vector())
2126 this->result
.swizzle
= swizzle_for_size(ir
->type
->vector_elements
);
2128 this->result
.swizzle
= SWIZZLE_NOOP
;
2130 this->result
.index
+= offset
;
2131 this->result
.type
= ir
->type
->base_type
;
2135 * We want to be careful in assignment setup to hit the actual storage
2136 * instead of potentially using a temporary like we might with the
2137 * ir_dereference handler.
2140 get_assignment_lhs(ir_dereference
*ir
, glsl_to_tgsi_visitor
*v
)
2142 /* The LHS must be a dereference. If the LHS is a variable indexed array
2143 * access of a vector, it must be separated into a series conditional moves
2144 * before reaching this point (see ir_vec_index_to_cond_assign).
2146 assert(ir
->as_dereference());
2147 ir_dereference_array
*deref_array
= ir
->as_dereference_array();
2149 assert(!deref_array
->array
->type
->is_vector());
2152 /* Use the rvalue deref handler for the most part. We'll ignore
2153 * swizzles in it and write swizzles using writemask, though.
2156 return st_dst_reg(v
->result
);
2160 * Process the condition of a conditional assignment
2162 * Examines the condition of a conditional assignment to generate the optimal
2163 * first operand of a \c CMP instruction. If the condition is a relational
2164 * operator with 0 (e.g., \c ir_binop_less), the value being compared will be
2165 * used as the source for the \c CMP instruction. Otherwise the comparison
2166 * is processed to a boolean result, and the boolean result is used as the
2167 * operand to the CMP instruction.
2170 glsl_to_tgsi_visitor::process_move_condition(ir_rvalue
*ir
)
2172 ir_rvalue
*src_ir
= ir
;
2174 bool switch_order
= false;
2176 ir_expression
*const expr
= ir
->as_expression();
2177 if ((expr
!= NULL
) && (expr
->get_num_operands() == 2)) {
2178 bool zero_on_left
= false;
2180 if (expr
->operands
[0]->is_zero()) {
2181 src_ir
= expr
->operands
[1];
2182 zero_on_left
= true;
2183 } else if (expr
->operands
[1]->is_zero()) {
2184 src_ir
= expr
->operands
[0];
2185 zero_on_left
= false;
2189 * (a < 0) T F F ( a < 0) T F F
2190 * (0 < a) F F T (-a < 0) F F T
2191 * (a <= 0) T T F (-a < 0) F F T (swap order of other operands)
2192 * (0 <= a) F T T ( a < 0) T F F (swap order of other operands)
2193 * (a > 0) F F T (-a < 0) F F T
2194 * (0 > a) T F F ( a < 0) T F F
2195 * (a >= 0) F T T ( a < 0) T F F (swap order of other operands)
2196 * (0 >= a) T T F (-a < 0) F F T (swap order of other operands)
2198 * Note that exchanging the order of 0 and 'a' in the comparison simply
2199 * means that the value of 'a' should be negated.
2202 switch (expr
->operation
) {
2204 switch_order
= false;
2205 negate
= zero_on_left
;
2208 case ir_binop_greater
:
2209 switch_order
= false;
2210 negate
= !zero_on_left
;
2213 case ir_binop_lequal
:
2214 switch_order
= true;
2215 negate
= !zero_on_left
;
2218 case ir_binop_gequal
:
2219 switch_order
= true;
2220 negate
= zero_on_left
;
2224 /* This isn't the right kind of comparison afterall, so make sure
2225 * the whole condition is visited.
2233 src_ir
->accept(this);
2235 /* We use the TGSI_OPCODE_CMP (a < 0 ? b : c) for conditional moves, and the
2236 * condition we produced is 0.0 or 1.0. By flipping the sign, we can
2237 * choose which value TGSI_OPCODE_CMP produces without an extra instruction
2238 * computing the condition.
2241 this->result
.negate
= ~this->result
.negate
;
2243 return switch_order
;
2247 glsl_to_tgsi_visitor::emit_block_mov(ir_assignment
*ir
, const struct glsl_type
*type
,
2248 st_dst_reg
*l
, st_src_reg
*r
)
2250 if (type
->base_type
== GLSL_TYPE_STRUCT
) {
2251 for (unsigned int i
= 0; i
< type
->length
; i
++) {
2252 emit_block_mov(ir
, type
->fields
.structure
[i
].type
, l
, r
);
2257 if (type
->is_array()) {
2258 for (unsigned int i
= 0; i
< type
->length
; i
++) {
2259 emit_block_mov(ir
, type
->fields
.array
, l
, r
);
2264 if (type
->is_matrix()) {
2265 const struct glsl_type
*vec_type
;
2267 vec_type
= glsl_type::get_instance(GLSL_TYPE_FLOAT
,
2268 type
->vector_elements
, 1);
2270 for (int i
= 0; i
< type
->matrix_columns
; i
++) {
2271 emit_block_mov(ir
, vec_type
, l
, r
);
2276 assert(type
->is_scalar() || type
->is_vector());
2278 r
->type
= type
->base_type
;
2279 emit(ir
, TGSI_OPCODE_MOV
, *l
, *r
);
2285 glsl_to_tgsi_visitor::visit(ir_assignment
*ir
)
2291 ir
->rhs
->accept(this);
2294 l
= get_assignment_lhs(ir
->lhs
, this);
2296 /* FINISHME: This should really set to the correct maximal writemask for each
2297 * FINISHME: component written (in the loops below). This case can only
2298 * FINISHME: occur for matrices, arrays, and structures.
2300 if (ir
->write_mask
== 0) {
2301 assert(!ir
->lhs
->type
->is_scalar() && !ir
->lhs
->type
->is_vector());
2302 l
.writemask
= WRITEMASK_XYZW
;
2303 } else if (ir
->lhs
->type
->is_scalar() &&
2304 ir
->lhs
->variable_referenced()->mode
== ir_var_shader_out
) {
2305 /* FINISHME: This hack makes writing to gl_FragDepth, which lives in the
2306 * FINISHME: W component of fragment shader output zero, work correctly.
2308 l
.writemask
= WRITEMASK_XYZW
;
2311 int first_enabled_chan
= 0;
2314 l
.writemask
= ir
->write_mask
;
2316 for (int i
= 0; i
< 4; i
++) {
2317 if (l
.writemask
& (1 << i
)) {
2318 first_enabled_chan
= GET_SWZ(r
.swizzle
, i
);
2323 /* Swizzle a small RHS vector into the channels being written.
2325 * glsl ir treats write_mask as dictating how many channels are
2326 * present on the RHS while TGSI treats write_mask as just
2327 * showing which channels of the vec4 RHS get written.
2329 for (int i
= 0; i
< 4; i
++) {
2330 if (l
.writemask
& (1 << i
))
2331 swizzles
[i
] = GET_SWZ(r
.swizzle
, rhs_chan
++);
2333 swizzles
[i
] = first_enabled_chan
;
2335 r
.swizzle
= MAKE_SWIZZLE4(swizzles
[0], swizzles
[1],
2336 swizzles
[2], swizzles
[3]);
2339 assert(l
.file
!= PROGRAM_UNDEFINED
);
2340 assert(r
.file
!= PROGRAM_UNDEFINED
);
2342 if (ir
->condition
) {
2343 const bool switch_order
= this->process_move_condition(ir
->condition
);
2344 st_src_reg condition
= this->result
;
2346 for (i
= 0; i
< type_size(ir
->lhs
->type
); i
++) {
2347 st_src_reg l_src
= st_src_reg(l
);
2348 st_src_reg condition_temp
= condition
;
2349 l_src
.swizzle
= swizzle_for_size(ir
->lhs
->type
->vector_elements
);
2351 if (native_integers
) {
2352 /* This is necessary because TGSI's CMP instruction expects the
2353 * condition to be a float, and we store booleans as integers.
2354 * If TGSI had a UCMP instruction or similar, this extra
2355 * instruction would not be necessary.
2357 condition_temp
= get_temp(glsl_type::vec4_type
);
2358 condition
.negate
= 0;
2359 emit(ir
, TGSI_OPCODE_I2F
, st_dst_reg(condition_temp
), condition
);
2360 condition_temp
.swizzle
= condition
.swizzle
;
2364 emit(ir
, TGSI_OPCODE_CMP
, l
, condition_temp
, l_src
, r
);
2366 emit(ir
, TGSI_OPCODE_CMP
, l
, condition_temp
, r
, l_src
);
2372 } else if (ir
->rhs
->as_expression() &&
2373 this->instructions
.get_tail() &&
2374 ir
->rhs
== ((glsl_to_tgsi_instruction
*)this->instructions
.get_tail())->ir
&&
2375 type_size(ir
->lhs
->type
) == 1 &&
2376 l
.writemask
== ((glsl_to_tgsi_instruction
*)this->instructions
.get_tail())->dst
.writemask
) {
2377 /* To avoid emitting an extra MOV when assigning an expression to a
2378 * variable, emit the last instruction of the expression again, but
2379 * replace the destination register with the target of the assignment.
2380 * Dead code elimination will remove the original instruction.
2382 glsl_to_tgsi_instruction
*inst
, *new_inst
;
2383 inst
= (glsl_to_tgsi_instruction
*)this->instructions
.get_tail();
2384 new_inst
= emit(ir
, inst
->op
, l
, inst
->src
[0], inst
->src
[1], inst
->src
[2]);
2385 new_inst
->saturate
= inst
->saturate
;
2386 inst
->dead_mask
= inst
->dst
.writemask
;
2388 emit_block_mov(ir
, ir
->rhs
->type
, &l
, &r
);
2394 glsl_to_tgsi_visitor::visit(ir_constant
*ir
)
2397 GLfloat stack_vals
[4] = { 0 };
2398 gl_constant_value
*values
= (gl_constant_value
*) stack_vals
;
2399 GLenum gl_type
= GL_NONE
;
2401 static int in_array
= 0;
2402 gl_register_file file
= in_array
? PROGRAM_CONSTANT
: PROGRAM_IMMEDIATE
;
2404 /* Unfortunately, 4 floats is all we can get into
2405 * _mesa_add_typed_unnamed_constant. So, make a temp to store an
2406 * aggregate constant and move each constant value into it. If we
2407 * get lucky, copy propagation will eliminate the extra moves.
2409 if (ir
->type
->base_type
== GLSL_TYPE_STRUCT
) {
2410 st_src_reg temp_base
= get_temp(ir
->type
);
2411 st_dst_reg temp
= st_dst_reg(temp_base
);
2413 foreach_iter(exec_list_iterator
, iter
, ir
->components
) {
2414 ir_constant
*field_value
= (ir_constant
*)iter
.get();
2415 int size
= type_size(field_value
->type
);
2419 field_value
->accept(this);
2422 for (i
= 0; i
< (unsigned int)size
; i
++) {
2423 emit(ir
, TGSI_OPCODE_MOV
, temp
, src
);
2429 this->result
= temp_base
;
2433 if (ir
->type
->is_array()) {
2434 st_src_reg temp_base
= get_temp(ir
->type
);
2435 st_dst_reg temp
= st_dst_reg(temp_base
);
2436 int size
= type_size(ir
->type
->fields
.array
);
2441 for (i
= 0; i
< ir
->type
->length
; i
++) {
2442 ir
->array_elements
[i
]->accept(this);
2444 for (int j
= 0; j
< size
; j
++) {
2445 emit(ir
, TGSI_OPCODE_MOV
, temp
, src
);
2451 this->result
= temp_base
;
2456 if (ir
->type
->is_matrix()) {
2457 st_src_reg mat
= get_temp(ir
->type
);
2458 st_dst_reg mat_column
= st_dst_reg(mat
);
2460 for (i
= 0; i
< ir
->type
->matrix_columns
; i
++) {
2461 assert(ir
->type
->base_type
== GLSL_TYPE_FLOAT
);
2462 values
= (gl_constant_value
*) &ir
->value
.f
[i
* ir
->type
->vector_elements
];
2464 src
= st_src_reg(file
, -1, ir
->type
->base_type
);
2465 src
.index
= add_constant(file
,
2467 ir
->type
->vector_elements
,
2470 emit(ir
, TGSI_OPCODE_MOV
, mat_column
, src
);
2479 switch (ir
->type
->base_type
) {
2480 case GLSL_TYPE_FLOAT
:
2482 for (i
= 0; i
< ir
->type
->vector_elements
; i
++) {
2483 values
[i
].f
= ir
->value
.f
[i
];
2486 case GLSL_TYPE_UINT
:
2487 gl_type
= native_integers
? GL_UNSIGNED_INT
: GL_FLOAT
;
2488 for (i
= 0; i
< ir
->type
->vector_elements
; i
++) {
2489 if (native_integers
)
2490 values
[i
].u
= ir
->value
.u
[i
];
2492 values
[i
].f
= ir
->value
.u
[i
];
2496 gl_type
= native_integers
? GL_INT
: GL_FLOAT
;
2497 for (i
= 0; i
< ir
->type
->vector_elements
; i
++) {
2498 if (native_integers
)
2499 values
[i
].i
= ir
->value
.i
[i
];
2501 values
[i
].f
= ir
->value
.i
[i
];
2504 case GLSL_TYPE_BOOL
:
2505 gl_type
= native_integers
? GL_BOOL
: GL_FLOAT
;
2506 for (i
= 0; i
< ir
->type
->vector_elements
; i
++) {
2507 if (native_integers
)
2508 values
[i
].u
= ir
->value
.b
[i
] ? ~0 : 0;
2510 values
[i
].f
= ir
->value
.b
[i
];
2514 assert(!"Non-float/uint/int/bool constant");
2517 this->result
= st_src_reg(file
, -1, ir
->type
);
2518 this->result
.index
= add_constant(file
,
2520 ir
->type
->vector_elements
,
2522 &this->result
.swizzle
);
2526 glsl_to_tgsi_visitor::get_function_signature(ir_function_signature
*sig
)
2528 function_entry
*entry
;
2530 foreach_iter(exec_list_iterator
, iter
, this->function_signatures
) {
2531 entry
= (function_entry
*)iter
.get();
2533 if (entry
->sig
== sig
)
2537 entry
= ralloc(mem_ctx
, function_entry
);
2539 entry
->sig_id
= this->next_signature_id
++;
2540 entry
->bgn_inst
= NULL
;
2542 /* Allocate storage for all the parameters. */
2543 foreach_iter(exec_list_iterator
, iter
, sig
->parameters
) {
2544 ir_variable
*param
= (ir_variable
*)iter
.get();
2545 variable_storage
*storage
;
2547 storage
= find_variable_storage(param
);
2550 storage
= new(mem_ctx
) variable_storage(param
, PROGRAM_TEMPORARY
,
2552 this->variables
.push_tail(storage
);
2554 this->next_temp
+= type_size(param
->type
);
2557 if (!sig
->return_type
->is_void()) {
2558 entry
->return_reg
= get_temp(sig
->return_type
);
2560 entry
->return_reg
= undef_src
;
2563 this->function_signatures
.push_tail(entry
);
2568 glsl_to_tgsi_visitor::visit(ir_call
*ir
)
2570 glsl_to_tgsi_instruction
*call_inst
;
2571 ir_function_signature
*sig
= ir
->callee
;
2572 function_entry
*entry
= get_function_signature(sig
);
2575 /* Process in parameters. */
2576 exec_list_iterator sig_iter
= sig
->parameters
.iterator();
2577 foreach_iter(exec_list_iterator
, iter
, *ir
) {
2578 ir_rvalue
*param_rval
= (ir_rvalue
*)iter
.get();
2579 ir_variable
*param
= (ir_variable
*)sig_iter
.get();
2581 if (param
->mode
== ir_var_function_in
||
2582 param
->mode
== ir_var_function_inout
) {
2583 variable_storage
*storage
= find_variable_storage(param
);
2586 param_rval
->accept(this);
2587 st_src_reg r
= this->result
;
2590 l
.file
= storage
->file
;
2591 l
.index
= storage
->index
;
2593 l
.writemask
= WRITEMASK_XYZW
;
2594 l
.cond_mask
= COND_TR
;
2596 for (i
= 0; i
< type_size(param
->type
); i
++) {
2597 emit(ir
, TGSI_OPCODE_MOV
, l
, r
);
2605 assert(!sig_iter
.has_next());
2607 /* Emit call instruction */
2608 call_inst
= emit(ir
, TGSI_OPCODE_CAL
);
2609 call_inst
->function
= entry
;
2611 /* Process out parameters. */
2612 sig_iter
= sig
->parameters
.iterator();
2613 foreach_iter(exec_list_iterator
, iter
, *ir
) {
2614 ir_rvalue
*param_rval
= (ir_rvalue
*)iter
.get();
2615 ir_variable
*param
= (ir_variable
*)sig_iter
.get();
2617 if (param
->mode
== ir_var_function_out
||
2618 param
->mode
== ir_var_function_inout
) {
2619 variable_storage
*storage
= find_variable_storage(param
);
2623 r
.file
= storage
->file
;
2624 r
.index
= storage
->index
;
2626 r
.swizzle
= SWIZZLE_NOOP
;
2629 param_rval
->accept(this);
2630 st_dst_reg l
= st_dst_reg(this->result
);
2632 for (i
= 0; i
< type_size(param
->type
); i
++) {
2633 emit(ir
, TGSI_OPCODE_MOV
, l
, r
);
2641 assert(!sig_iter
.has_next());
2643 /* Process return value. */
2644 this->result
= entry
->return_reg
;
2648 glsl_to_tgsi_visitor::visit(ir_texture
*ir
)
2650 st_src_reg result_src
, coord
, cube_sc
, lod_info
, projector
, dx
, dy
, offset
;
2651 st_dst_reg result_dst
, coord_dst
, cube_sc_dst
;
2652 glsl_to_tgsi_instruction
*inst
= NULL
;
2653 unsigned opcode
= TGSI_OPCODE_NOP
;
2654 const glsl_type
*sampler_type
= ir
->sampler
->type
;
2655 bool is_cube_array
= false;
2657 /* if we are a cube array sampler */
2658 if ((sampler_type
->sampler_dimensionality
== GLSL_SAMPLER_DIM_CUBE
&&
2659 sampler_type
->sampler_array
)) {
2660 is_cube_array
= true;
2663 if (ir
->coordinate
) {
2664 ir
->coordinate
->accept(this);
2666 /* Put our coords in a temp. We'll need to modify them for shadow,
2667 * projection, or LOD, so the only case we'd use it as is is if
2668 * we're doing plain old texturing. The optimization passes on
2669 * glsl_to_tgsi_visitor should handle cleaning up our mess in that case.
2671 coord
= get_temp(glsl_type::vec4_type
);
2672 coord_dst
= st_dst_reg(coord
);
2673 emit(ir
, TGSI_OPCODE_MOV
, coord_dst
, this->result
);
2676 if (ir
->projector
) {
2677 ir
->projector
->accept(this);
2678 projector
= this->result
;
2681 /* Storage for our result. Ideally for an assignment we'd be using
2682 * the actual storage for the result here, instead.
2684 result_src
= get_temp(ir
->type
);
2685 result_dst
= st_dst_reg(result_src
);
2689 opcode
= (is_cube_array
&& ir
->shadow_comparitor
) ? TGSI_OPCODE_TEX2
: TGSI_OPCODE_TEX
;
2692 opcode
= is_cube_array
? TGSI_OPCODE_TXB2
: TGSI_OPCODE_TXB
;
2693 ir
->lod_info
.bias
->accept(this);
2694 lod_info
= this->result
;
2697 opcode
= is_cube_array
? TGSI_OPCODE_TXL2
: TGSI_OPCODE_TXL
;
2698 ir
->lod_info
.lod
->accept(this);
2699 lod_info
= this->result
;
2702 opcode
= TGSI_OPCODE_TXD
;
2703 ir
->lod_info
.grad
.dPdx
->accept(this);
2705 ir
->lod_info
.grad
.dPdy
->accept(this);
2709 opcode
= TGSI_OPCODE_TXQ
;
2710 ir
->lod_info
.lod
->accept(this);
2711 lod_info
= this->result
;
2714 opcode
= TGSI_OPCODE_TXF
;
2715 ir
->lod_info
.lod
->accept(this);
2716 lod_info
= this->result
;
2718 ir
->offset
->accept(this);
2719 offset
= this->result
;
2724 if (ir
->projector
) {
2725 if (opcode
== TGSI_OPCODE_TEX
) {
2726 /* Slot the projector in as the last component of the coord. */
2727 coord_dst
.writemask
= WRITEMASK_W
;
2728 emit(ir
, TGSI_OPCODE_MOV
, coord_dst
, projector
);
2729 coord_dst
.writemask
= WRITEMASK_XYZW
;
2730 opcode
= TGSI_OPCODE_TXP
;
2732 st_src_reg coord_w
= coord
;
2733 coord_w
.swizzle
= SWIZZLE_WWWW
;
2735 /* For the other TEX opcodes there's no projective version
2736 * since the last slot is taken up by LOD info. Do the
2737 * projective divide now.
2739 coord_dst
.writemask
= WRITEMASK_W
;
2740 emit(ir
, TGSI_OPCODE_RCP
, coord_dst
, projector
);
2742 /* In the case where we have to project the coordinates "by hand,"
2743 * the shadow comparator value must also be projected.
2745 st_src_reg tmp_src
= coord
;
2746 if (ir
->shadow_comparitor
) {
2747 /* Slot the shadow value in as the second to last component of the
2750 ir
->shadow_comparitor
->accept(this);
2752 tmp_src
= get_temp(glsl_type::vec4_type
);
2753 st_dst_reg tmp_dst
= st_dst_reg(tmp_src
);
2755 /* Projective division not allowed for array samplers. */
2756 assert(!sampler_type
->sampler_array
);
2758 tmp_dst
.writemask
= WRITEMASK_Z
;
2759 emit(ir
, TGSI_OPCODE_MOV
, tmp_dst
, this->result
);
2761 tmp_dst
.writemask
= WRITEMASK_XY
;
2762 emit(ir
, TGSI_OPCODE_MOV
, tmp_dst
, coord
);
2765 coord_dst
.writemask
= WRITEMASK_XYZ
;
2766 emit(ir
, TGSI_OPCODE_MUL
, coord_dst
, tmp_src
, coord_w
);
2768 coord_dst
.writemask
= WRITEMASK_XYZW
;
2769 coord
.swizzle
= SWIZZLE_XYZW
;
2773 /* If projection is done and the opcode is not TGSI_OPCODE_TXP, then the shadow
2774 * comparator was put in the correct place (and projected) by the code,
2775 * above, that handles by-hand projection.
2777 if (ir
->shadow_comparitor
&& (!ir
->projector
|| opcode
== TGSI_OPCODE_TXP
)) {
2778 /* Slot the shadow value in as the second to last component of the
2781 ir
->shadow_comparitor
->accept(this);
2783 if (is_cube_array
) {
2784 cube_sc
= get_temp(glsl_type::float_type
);
2785 cube_sc_dst
= st_dst_reg(cube_sc
);
2786 cube_sc_dst
.writemask
= WRITEMASK_X
;
2787 emit(ir
, TGSI_OPCODE_MOV
, cube_sc_dst
, this->result
);
2788 cube_sc_dst
.writemask
= WRITEMASK_X
;
2791 if ((sampler_type
->sampler_dimensionality
== GLSL_SAMPLER_DIM_2D
&&
2792 sampler_type
->sampler_array
) ||
2793 sampler_type
->sampler_dimensionality
== GLSL_SAMPLER_DIM_CUBE
) {
2794 coord_dst
.writemask
= WRITEMASK_W
;
2796 coord_dst
.writemask
= WRITEMASK_Z
;
2799 emit(ir
, TGSI_OPCODE_MOV
, coord_dst
, this->result
);
2800 coord_dst
.writemask
= WRITEMASK_XYZW
;
2804 if (opcode
== TGSI_OPCODE_TXL
|| opcode
== TGSI_OPCODE_TXB
||
2805 opcode
== TGSI_OPCODE_TXF
) {
2806 /* TGSI stores LOD or LOD bias in the last channel of the coords. */
2807 coord_dst
.writemask
= WRITEMASK_W
;
2808 emit(ir
, TGSI_OPCODE_MOV
, coord_dst
, lod_info
);
2809 coord_dst
.writemask
= WRITEMASK_XYZW
;
2812 if (opcode
== TGSI_OPCODE_TXD
)
2813 inst
= emit(ir
, opcode
, result_dst
, coord
, dx
, dy
);
2814 else if (opcode
== TGSI_OPCODE_TXQ
)
2815 inst
= emit(ir
, opcode
, result_dst
, lod_info
);
2816 else if (opcode
== TGSI_OPCODE_TXF
) {
2817 inst
= emit(ir
, opcode
, result_dst
, coord
);
2818 } else if (opcode
== TGSI_OPCODE_TXL2
|| opcode
== TGSI_OPCODE_TXB2
) {
2819 inst
= emit(ir
, opcode
, result_dst
, coord
, lod_info
);
2820 } else if (opcode
== TGSI_OPCODE_TEX2
) {
2821 inst
= emit(ir
, opcode
, result_dst
, coord
, cube_sc
);
2823 inst
= emit(ir
, opcode
, result_dst
, coord
);
2825 if (ir
->shadow_comparitor
)
2826 inst
->tex_shadow
= GL_TRUE
;
2828 inst
->sampler
= _mesa_get_sampler_uniform_value(ir
->sampler
,
2829 this->shader_program
,
2833 inst
->tex_offset_num_offset
= 1;
2834 inst
->tex_offsets
[0].Index
= offset
.index
;
2835 inst
->tex_offsets
[0].File
= offset
.file
;
2836 inst
->tex_offsets
[0].SwizzleX
= GET_SWZ(offset
.swizzle
, 0);
2837 inst
->tex_offsets
[0].SwizzleY
= GET_SWZ(offset
.swizzle
, 1);
2838 inst
->tex_offsets
[0].SwizzleZ
= GET_SWZ(offset
.swizzle
, 2);
2841 switch (sampler_type
->sampler_dimensionality
) {
2842 case GLSL_SAMPLER_DIM_1D
:
2843 inst
->tex_target
= (sampler_type
->sampler_array
)
2844 ? TEXTURE_1D_ARRAY_INDEX
: TEXTURE_1D_INDEX
;
2846 case GLSL_SAMPLER_DIM_2D
:
2847 inst
->tex_target
= (sampler_type
->sampler_array
)
2848 ? TEXTURE_2D_ARRAY_INDEX
: TEXTURE_2D_INDEX
;
2850 case GLSL_SAMPLER_DIM_3D
:
2851 inst
->tex_target
= TEXTURE_3D_INDEX
;
2853 case GLSL_SAMPLER_DIM_CUBE
:
2854 inst
->tex_target
= (sampler_type
->sampler_array
)
2855 ? TEXTURE_CUBE_ARRAY_INDEX
: TEXTURE_CUBE_INDEX
;
2857 case GLSL_SAMPLER_DIM_RECT
:
2858 inst
->tex_target
= TEXTURE_RECT_INDEX
;
2860 case GLSL_SAMPLER_DIM_BUF
:
2861 inst
->tex_target
= TEXTURE_BUFFER_INDEX
;
2863 case GLSL_SAMPLER_DIM_EXTERNAL
:
2864 inst
->tex_target
= TEXTURE_EXTERNAL_INDEX
;
2867 assert(!"Should not get here.");
2870 this->result
= result_src
;
2874 glsl_to_tgsi_visitor::visit(ir_return
*ir
)
2876 if (ir
->get_value()) {
2880 assert(current_function
);
2882 ir
->get_value()->accept(this);
2883 st_src_reg r
= this->result
;
2885 l
= st_dst_reg(current_function
->return_reg
);
2887 for (i
= 0; i
< type_size(current_function
->sig
->return_type
); i
++) {
2888 emit(ir
, TGSI_OPCODE_MOV
, l
, r
);
2894 emit(ir
, TGSI_OPCODE_RET
);
2898 glsl_to_tgsi_visitor::visit(ir_discard
*ir
)
2900 if (ir
->condition
) {
2901 ir
->condition
->accept(this);
2902 this->result
.negate
= ~this->result
.negate
;
2903 emit(ir
, TGSI_OPCODE_KIL
, undef_dst
, this->result
);
2905 emit(ir
, TGSI_OPCODE_KILP
);
2910 glsl_to_tgsi_visitor::visit(ir_if
*ir
)
2912 glsl_to_tgsi_instruction
*cond_inst
, *if_inst
;
2913 glsl_to_tgsi_instruction
*prev_inst
;
2915 prev_inst
= (glsl_to_tgsi_instruction
*)this->instructions
.get_tail();
2917 ir
->condition
->accept(this);
2918 assert(this->result
.file
!= PROGRAM_UNDEFINED
);
2920 if (this->options
->EmitCondCodes
) {
2921 cond_inst
= (glsl_to_tgsi_instruction
*)this->instructions
.get_tail();
2923 /* See if we actually generated any instruction for generating
2924 * the condition. If not, then cook up a move to a temp so we
2925 * have something to set cond_update on.
2927 if (cond_inst
== prev_inst
) {
2928 st_src_reg temp
= get_temp(glsl_type::bool_type
);
2929 cond_inst
= emit(ir
->condition
, TGSI_OPCODE_MOV
, st_dst_reg(temp
), result
);
2931 cond_inst
->cond_update
= GL_TRUE
;
2933 if_inst
= emit(ir
->condition
, TGSI_OPCODE_IF
);
2934 if_inst
->dst
.cond_mask
= COND_NE
;
2936 if_inst
= emit(ir
->condition
, TGSI_OPCODE_IF
, undef_dst
, this->result
);
2939 this->instructions
.push_tail(if_inst
);
2941 visit_exec_list(&ir
->then_instructions
, this);
2943 if (!ir
->else_instructions
.is_empty()) {
2944 emit(ir
->condition
, TGSI_OPCODE_ELSE
);
2945 visit_exec_list(&ir
->else_instructions
, this);
2948 if_inst
= emit(ir
->condition
, TGSI_OPCODE_ENDIF
);
2951 glsl_to_tgsi_visitor::glsl_to_tgsi_visitor()
2953 result
.file
= PROGRAM_UNDEFINED
;
2955 next_signature_id
= 1;
2957 current_function
= NULL
;
2958 num_address_regs
= 0;
2960 indirect_addr_temps
= false;
2961 indirect_addr_consts
= false;
2963 native_integers
= false;
2964 mem_ctx
= ralloc_context(NULL
);
2967 shader_program
= NULL
;
2971 glsl_to_tgsi_visitor::~glsl_to_tgsi_visitor()
2973 ralloc_free(mem_ctx
);
2976 extern "C" void free_glsl_to_tgsi_visitor(glsl_to_tgsi_visitor
*v
)
2983 * Count resources used by the given gpu program (number of texture
2987 count_resources(glsl_to_tgsi_visitor
*v
, gl_program
*prog
)
2989 v
->samplers_used
= 0;
2991 foreach_iter(exec_list_iterator
, iter
, v
->instructions
) {
2992 glsl_to_tgsi_instruction
*inst
= (glsl_to_tgsi_instruction
*)iter
.get();
2994 if (is_tex_instruction(inst
->op
)) {
2995 v
->samplers_used
|= 1 << inst
->sampler
;
2997 if (inst
->tex_shadow
) {
2998 prog
->ShadowSamplers
|= 1 << inst
->sampler
;
3003 prog
->SamplersUsed
= v
->samplers_used
;
3005 if (v
->shader_program
!= NULL
)
3006 _mesa_update_shader_textures_used(v
->shader_program
, prog
);
3010 set_uniform_initializer(struct gl_context
*ctx
, void *mem_ctx
,
3011 struct gl_shader_program
*shader_program
,
3012 const char *name
, const glsl_type
*type
,
3015 if (type
->is_record()) {
3016 ir_constant
*field_constant
;
3018 field_constant
= (ir_constant
*)val
->components
.get_head();
3020 for (unsigned int i
= 0; i
< type
->length
; i
++) {
3021 const glsl_type
*field_type
= type
->fields
.structure
[i
].type
;
3022 const char *field_name
= ralloc_asprintf(mem_ctx
, "%s.%s", name
,
3023 type
->fields
.structure
[i
].name
);
3024 set_uniform_initializer(ctx
, mem_ctx
, shader_program
, field_name
,
3025 field_type
, field_constant
);
3026 field_constant
= (ir_constant
*)field_constant
->next
;
3032 unsigned index
= _mesa_get_uniform_location(ctx
, shader_program
, name
,
3034 if (offset
== GL_INVALID_INDEX
) {
3035 fail_link(shader_program
,
3036 "Couldn't find uniform for initializer %s\n", name
);
3039 int loc
= _mesa_uniform_merge_location_offset(index
, offset
);
3041 for (unsigned int i
= 0; i
< (type
->is_array() ? type
->length
: 1); i
++) {
3042 ir_constant
*element
;
3043 const glsl_type
*element_type
;
3044 if (type
->is_array()) {
3045 element
= val
->array_elements
[i
];
3046 element_type
= type
->fields
.array
;
3049 element_type
= type
;
3054 if (element_type
->base_type
== GLSL_TYPE_BOOL
) {
3055 int *conv
= ralloc_array(mem_ctx
, int, element_type
->components());
3056 for (unsigned int j
= 0; j
< element_type
->components(); j
++) {
3057 conv
[j
] = element
->value
.b
[j
];
3059 values
= (void *)conv
;
3060 element_type
= glsl_type::get_instance(GLSL_TYPE_INT
,
3061 element_type
->vector_elements
,
3064 values
= &element
->value
;
3067 if (element_type
->is_matrix()) {
3068 _mesa_uniform_matrix(ctx
, shader_program
,
3069 element_type
->matrix_columns
,
3070 element_type
->vector_elements
,
3071 loc
, 1, GL_FALSE
, (GLfloat
*)values
);
3073 _mesa_uniform(ctx
, shader_program
, loc
, element_type
->matrix_columns
,
3074 values
, element_type
->gl_type
);
3082 * Returns the mask of channels (bitmask of WRITEMASK_X,Y,Z,W) which
3083 * are read from the given src in this instruction
3086 get_src_arg_mask(st_dst_reg dst
, st_src_reg src
)
3088 int read_mask
= 0, comp
;
3090 /* Now, given the src swizzle and the written channels, find which
3091 * components are actually read
3093 for (comp
= 0; comp
< 4; ++comp
) {
3094 const unsigned coord
= GET_SWZ(src
.swizzle
, comp
);
3096 if (dst
.writemask
& (1 << comp
) && coord
<= SWIZZLE_W
)
3097 read_mask
|= 1 << coord
;
3104 * This pass replaces CMP T0, T1 T2 T0 with MOV T0, T2 when the CMP
3105 * instruction is the first instruction to write to register T0. There are
3106 * several lowering passes done in GLSL IR (e.g. branches and
3107 * relative addressing) that create a large number of conditional assignments
3108 * that ir_to_mesa converts to CMP instructions like the one mentioned above.
3110 * Here is why this conversion is safe:
3111 * CMP T0, T1 T2 T0 can be expanded to:
3117 * If (T1 < 0.0) evaluates to true then our replacement MOV T0, T2 is the same
3118 * as the original program. If (T1 < 0.0) evaluates to false, executing
3119 * MOV T0, T0 will store a garbage value in T0 since T0 is uninitialized.
3120 * Therefore, it doesn't matter that we are replacing MOV T0, T0 with MOV T0, T2
3121 * because any instruction that was going to read from T0 after this was going
3122 * to read a garbage value anyway.
3125 glsl_to_tgsi_visitor::simplify_cmp(void)
3127 unsigned *tempWrites
;
3128 unsigned outputWrites
[MAX_PROGRAM_OUTPUTS
];
3130 tempWrites
= new unsigned[MAX_TEMPS
];
3134 memset(tempWrites
, 0, sizeof(unsigned) * MAX_TEMPS
);
3135 memset(outputWrites
, 0, sizeof(outputWrites
));
3137 foreach_iter(exec_list_iterator
, iter
, this->instructions
) {
3138 glsl_to_tgsi_instruction
*inst
= (glsl_to_tgsi_instruction
*)iter
.get();
3139 unsigned prevWriteMask
= 0;
3141 /* Give up if we encounter relative addressing or flow control. */
3142 if (inst
->dst
.reladdr
||
3143 tgsi_get_opcode_info(inst
->op
)->is_branch
||
3144 inst
->op
== TGSI_OPCODE_BGNSUB
||
3145 inst
->op
== TGSI_OPCODE_CONT
||
3146 inst
->op
== TGSI_OPCODE_END
||
3147 inst
->op
== TGSI_OPCODE_ENDSUB
||
3148 inst
->op
== TGSI_OPCODE_RET
) {
3152 if (inst
->dst
.file
== PROGRAM_OUTPUT
) {
3153 assert(inst
->dst
.index
< MAX_PROGRAM_OUTPUTS
);
3154 prevWriteMask
= outputWrites
[inst
->dst
.index
];
3155 outputWrites
[inst
->dst
.index
] |= inst
->dst
.writemask
;
3156 } else if (inst
->dst
.file
== PROGRAM_TEMPORARY
) {
3157 assert(inst
->dst
.index
< MAX_TEMPS
);
3158 prevWriteMask
= tempWrites
[inst
->dst
.index
];
3159 tempWrites
[inst
->dst
.index
] |= inst
->dst
.writemask
;
3162 /* For a CMP to be considered a conditional write, the destination
3163 * register and source register two must be the same. */
3164 if (inst
->op
== TGSI_OPCODE_CMP
3165 && !(inst
->dst
.writemask
& prevWriteMask
)
3166 && inst
->src
[2].file
== inst
->dst
.file
3167 && inst
->src
[2].index
== inst
->dst
.index
3168 && inst
->dst
.writemask
== get_src_arg_mask(inst
->dst
, inst
->src
[2])) {
3170 inst
->op
= TGSI_OPCODE_MOV
;
3171 inst
->src
[0] = inst
->src
[1];
3175 delete [] tempWrites
;
3178 /* Replaces all references to a temporary register index with another index. */
3180 glsl_to_tgsi_visitor::rename_temp_register(int index
, int new_index
)
3182 foreach_iter(exec_list_iterator
, iter
, this->instructions
) {
3183 glsl_to_tgsi_instruction
*inst
= (glsl_to_tgsi_instruction
*)iter
.get();
3186 for (j
=0; j
< num_inst_src_regs(inst
->op
); j
++) {
3187 if (inst
->src
[j
].file
== PROGRAM_TEMPORARY
&&
3188 inst
->src
[j
].index
== index
) {
3189 inst
->src
[j
].index
= new_index
;
3193 if (inst
->dst
.file
== PROGRAM_TEMPORARY
&& inst
->dst
.index
== index
) {
3194 inst
->dst
.index
= new_index
;
3200 glsl_to_tgsi_visitor::get_first_temp_read(int index
)
3202 int depth
= 0; /* loop depth */
3203 int loop_start
= -1; /* index of the first active BGNLOOP (if any) */
3206 foreach_iter(exec_list_iterator
, iter
, this->instructions
) {
3207 glsl_to_tgsi_instruction
*inst
= (glsl_to_tgsi_instruction
*)iter
.get();
3209 for (j
=0; j
< num_inst_src_regs(inst
->op
); j
++) {
3210 if (inst
->src
[j
].file
== PROGRAM_TEMPORARY
&&
3211 inst
->src
[j
].index
== index
) {
3212 return (depth
== 0) ? i
: loop_start
;
3216 if (inst
->op
== TGSI_OPCODE_BGNLOOP
) {
3219 } else if (inst
->op
== TGSI_OPCODE_ENDLOOP
) {
3232 glsl_to_tgsi_visitor::get_first_temp_write(int index
)
3234 int depth
= 0; /* loop depth */
3235 int loop_start
= -1; /* index of the first active BGNLOOP (if any) */
3238 foreach_iter(exec_list_iterator
, iter
, this->instructions
) {
3239 glsl_to_tgsi_instruction
*inst
= (glsl_to_tgsi_instruction
*)iter
.get();
3241 if (inst
->dst
.file
== PROGRAM_TEMPORARY
&& inst
->dst
.index
== index
) {
3242 return (depth
== 0) ? i
: loop_start
;
3245 if (inst
->op
== TGSI_OPCODE_BGNLOOP
) {
3248 } else if (inst
->op
== TGSI_OPCODE_ENDLOOP
) {
3261 glsl_to_tgsi_visitor::get_last_temp_read(int index
)
3263 int depth
= 0; /* loop depth */
3264 int last
= -1; /* index of last instruction that reads the temporary */
3267 foreach_iter(exec_list_iterator
, iter
, this->instructions
) {
3268 glsl_to_tgsi_instruction
*inst
= (glsl_to_tgsi_instruction
*)iter
.get();
3270 for (j
=0; j
< num_inst_src_regs(inst
->op
); j
++) {
3271 if (inst
->src
[j
].file
== PROGRAM_TEMPORARY
&&
3272 inst
->src
[j
].index
== index
) {
3273 last
= (depth
== 0) ? i
: -2;
3277 if (inst
->op
== TGSI_OPCODE_BGNLOOP
)
3279 else if (inst
->op
== TGSI_OPCODE_ENDLOOP
)
3280 if (--depth
== 0 && last
== -2)
3292 glsl_to_tgsi_visitor::get_last_temp_write(int index
)
3294 int depth
= 0; /* loop depth */
3295 int last
= -1; /* index of last instruction that writes to the temporary */
3298 foreach_iter(exec_list_iterator
, iter
, this->instructions
) {
3299 glsl_to_tgsi_instruction
*inst
= (glsl_to_tgsi_instruction
*)iter
.get();
3301 if (inst
->dst
.file
== PROGRAM_TEMPORARY
&& inst
->dst
.index
== index
)
3302 last
= (depth
== 0) ? i
: -2;
3304 if (inst
->op
== TGSI_OPCODE_BGNLOOP
)
3306 else if (inst
->op
== TGSI_OPCODE_ENDLOOP
)
3307 if (--depth
== 0 && last
== -2)
3319 * On a basic block basis, tracks available PROGRAM_TEMPORARY register
3320 * channels for copy propagation and updates following instructions to
3321 * use the original versions.
3323 * The glsl_to_tgsi_visitor lazily produces code assuming that this pass
3324 * will occur. As an example, a TXP production before this pass:
3326 * 0: MOV TEMP[1], INPUT[4].xyyy;
3327 * 1: MOV TEMP[1].w, INPUT[4].wwww;
3328 * 2: TXP TEMP[2], TEMP[1], texture[0], 2D;
3332 * 0: MOV TEMP[1], INPUT[4].xyyy;
3333 * 1: MOV TEMP[1].w, INPUT[4].wwww;
3334 * 2: TXP TEMP[2], INPUT[4].xyyw, texture[0], 2D;
3336 * which allows for dead code elimination on TEMP[1]'s writes.
3339 glsl_to_tgsi_visitor::copy_propagate(void)
3341 glsl_to_tgsi_instruction
**acp
= rzalloc_array(mem_ctx
,
3342 glsl_to_tgsi_instruction
*,
3343 this->next_temp
* 4);
3344 int *acp_level
= rzalloc_array(mem_ctx
, int, this->next_temp
* 4);
3347 foreach_iter(exec_list_iterator
, iter
, this->instructions
) {
3348 glsl_to_tgsi_instruction
*inst
= (glsl_to_tgsi_instruction
*)iter
.get();
3350 assert(inst
->dst
.file
!= PROGRAM_TEMPORARY
3351 || inst
->dst
.index
< this->next_temp
);
3353 /* First, do any copy propagation possible into the src regs. */
3354 for (int r
= 0; r
< 3; r
++) {
3355 glsl_to_tgsi_instruction
*first
= NULL
;
3357 int acp_base
= inst
->src
[r
].index
* 4;
3359 if (inst
->src
[r
].file
!= PROGRAM_TEMPORARY
||
3360 inst
->src
[r
].reladdr
)
3363 /* See if we can find entries in the ACP consisting of MOVs
3364 * from the same src register for all the swizzled channels
3365 * of this src register reference.
3367 for (int i
= 0; i
< 4; i
++) {
3368 int src_chan
= GET_SWZ(inst
->src
[r
].swizzle
, i
);
3369 glsl_to_tgsi_instruction
*copy_chan
= acp
[acp_base
+ src_chan
];
3376 assert(acp_level
[acp_base
+ src_chan
] <= level
);
3381 if (first
->src
[0].file
!= copy_chan
->src
[0].file
||
3382 first
->src
[0].index
!= copy_chan
->src
[0].index
) {
3390 /* We've now validated that we can copy-propagate to
3391 * replace this src register reference. Do it.
3393 inst
->src
[r
].file
= first
->src
[0].file
;
3394 inst
->src
[r
].index
= first
->src
[0].index
;
3397 for (int i
= 0; i
< 4; i
++) {
3398 int src_chan
= GET_SWZ(inst
->src
[r
].swizzle
, i
);
3399 glsl_to_tgsi_instruction
*copy_inst
= acp
[acp_base
+ src_chan
];
3400 swizzle
|= (GET_SWZ(copy_inst
->src
[0].swizzle
, src_chan
) <<
3403 inst
->src
[r
].swizzle
= swizzle
;
3408 case TGSI_OPCODE_BGNLOOP
:
3409 case TGSI_OPCODE_ENDLOOP
:
3410 /* End of a basic block, clear the ACP entirely. */
3411 memset(acp
, 0, sizeof(*acp
) * this->next_temp
* 4);
3414 case TGSI_OPCODE_IF
:
3418 case TGSI_OPCODE_ENDIF
:
3419 case TGSI_OPCODE_ELSE
:
3420 /* Clear all channels written inside the block from the ACP, but
3421 * leaving those that were not touched.
3423 for (int r
= 0; r
< this->next_temp
; r
++) {
3424 for (int c
= 0; c
< 4; c
++) {
3425 if (!acp
[4 * r
+ c
])
3428 if (acp_level
[4 * r
+ c
] >= level
)
3429 acp
[4 * r
+ c
] = NULL
;
3432 if (inst
->op
== TGSI_OPCODE_ENDIF
)
3437 /* Continuing the block, clear any written channels from
3440 if (inst
->dst
.file
== PROGRAM_TEMPORARY
&& inst
->dst
.reladdr
) {
3441 /* Any temporary might be written, so no copy propagation
3442 * across this instruction.
3444 memset(acp
, 0, sizeof(*acp
) * this->next_temp
* 4);
3445 } else if (inst
->dst
.file
== PROGRAM_OUTPUT
&&
3446 inst
->dst
.reladdr
) {
3447 /* Any output might be written, so no copy propagation
3448 * from outputs across this instruction.
3450 for (int r
= 0; r
< this->next_temp
; r
++) {
3451 for (int c
= 0; c
< 4; c
++) {
3452 if (!acp
[4 * r
+ c
])
3455 if (acp
[4 * r
+ c
]->src
[0].file
== PROGRAM_OUTPUT
)
3456 acp
[4 * r
+ c
] = NULL
;
3459 } else if (inst
->dst
.file
== PROGRAM_TEMPORARY
||
3460 inst
->dst
.file
== PROGRAM_OUTPUT
) {
3461 /* Clear where it's used as dst. */
3462 if (inst
->dst
.file
== PROGRAM_TEMPORARY
) {
3463 for (int c
= 0; c
< 4; c
++) {
3464 if (inst
->dst
.writemask
& (1 << c
)) {
3465 acp
[4 * inst
->dst
.index
+ c
] = NULL
;
3470 /* Clear where it's used as src. */
3471 for (int r
= 0; r
< this->next_temp
; r
++) {
3472 for (int c
= 0; c
< 4; c
++) {
3473 if (!acp
[4 * r
+ c
])
3476 int src_chan
= GET_SWZ(acp
[4 * r
+ c
]->src
[0].swizzle
, c
);
3478 if (acp
[4 * r
+ c
]->src
[0].file
== inst
->dst
.file
&&
3479 acp
[4 * r
+ c
]->src
[0].index
== inst
->dst
.index
&&
3480 inst
->dst
.writemask
& (1 << src_chan
))
3482 acp
[4 * r
+ c
] = NULL
;
3490 /* If this is a copy, add it to the ACP. */
3491 if (inst
->op
== TGSI_OPCODE_MOV
&&
3492 inst
->dst
.file
== PROGRAM_TEMPORARY
&&
3493 !inst
->dst
.reladdr
&&
3495 !inst
->src
[0].reladdr
&&
3496 !inst
->src
[0].negate
) {
3497 for (int i
= 0; i
< 4; i
++) {
3498 if (inst
->dst
.writemask
& (1 << i
)) {
3499 acp
[4 * inst
->dst
.index
+ i
] = inst
;
3500 acp_level
[4 * inst
->dst
.index
+ i
] = level
;
3506 ralloc_free(acp_level
);
3511 * Tracks available PROGRAM_TEMPORARY registers for dead code elimination.
3513 * The glsl_to_tgsi_visitor lazily produces code assuming that this pass
3514 * will occur. As an example, a TXP production after copy propagation but
3517 * 0: MOV TEMP[1], INPUT[4].xyyy;
3518 * 1: MOV TEMP[1].w, INPUT[4].wwww;
3519 * 2: TXP TEMP[2], INPUT[4].xyyw, texture[0], 2D;
3521 * and after this pass:
3523 * 0: TXP TEMP[2], INPUT[4].xyyw, texture[0], 2D;
3525 * FIXME: assumes that all functions are inlined (no support for BGNSUB/ENDSUB)
3526 * FIXME: doesn't eliminate all dead code inside of loops; it steps around them
3529 glsl_to_tgsi_visitor::eliminate_dead_code(void)
3533 for (i
=0; i
< this->next_temp
; i
++) {
3534 int last_read
= get_last_temp_read(i
);
3537 foreach_iter(exec_list_iterator
, iter
, this->instructions
) {
3538 glsl_to_tgsi_instruction
*inst
= (glsl_to_tgsi_instruction
*)iter
.get();
3540 if (inst
->dst
.file
== PROGRAM_TEMPORARY
&& inst
->dst
.index
== i
&&
3553 * On a basic block basis, tracks available PROGRAM_TEMPORARY registers for dead
3554 * code elimination. This is less primitive than eliminate_dead_code(), as it
3555 * is per-channel and can detect consecutive writes without a read between them
3556 * as dead code. However, there is some dead code that can be eliminated by
3557 * eliminate_dead_code() but not this function - for example, this function
3558 * cannot eliminate an instruction writing to a register that is never read and
3559 * is the only instruction writing to that register.
3561 * The glsl_to_tgsi_visitor lazily produces code assuming that this pass
3565 glsl_to_tgsi_visitor::eliminate_dead_code_advanced(void)
3567 glsl_to_tgsi_instruction
**writes
= rzalloc_array(mem_ctx
,
3568 glsl_to_tgsi_instruction
*,
3569 this->next_temp
* 4);
3570 int *write_level
= rzalloc_array(mem_ctx
, int, this->next_temp
* 4);
3574 foreach_iter(exec_list_iterator
, iter
, this->instructions
) {
3575 glsl_to_tgsi_instruction
*inst
= (glsl_to_tgsi_instruction
*)iter
.get();
3577 assert(inst
->dst
.file
!= PROGRAM_TEMPORARY
3578 || inst
->dst
.index
< this->next_temp
);
3581 case TGSI_OPCODE_BGNLOOP
:
3582 case TGSI_OPCODE_ENDLOOP
:
3583 case TGSI_OPCODE_CONT
:
3584 case TGSI_OPCODE_BRK
:
3585 /* End of a basic block, clear the write array entirely.
3587 * This keeps us from killing dead code when the writes are
3588 * on either side of a loop, even when the register isn't touched
3589 * inside the loop. However, glsl_to_tgsi_visitor doesn't seem to emit
3590 * dead code of this type, so it shouldn't make a difference as long as
3591 * the dead code elimination pass in the GLSL compiler does its job.
3593 memset(writes
, 0, sizeof(*writes
) * this->next_temp
* 4);
3596 case TGSI_OPCODE_ENDIF
:
3597 case TGSI_OPCODE_ELSE
:
3598 /* Promote the recorded level of all channels written inside the
3599 * preceding if or else block to the level above the if/else block.
3601 for (int r
= 0; r
< this->next_temp
; r
++) {
3602 for (int c
= 0; c
< 4; c
++) {
3603 if (!writes
[4 * r
+ c
])
3606 if (write_level
[4 * r
+ c
] == level
)
3607 write_level
[4 * r
+ c
] = level
-1;
3611 if(inst
->op
== TGSI_OPCODE_ENDIF
)
3616 case TGSI_OPCODE_IF
:
3618 /* fallthrough to default case to mark the condition as read */
3621 /* Continuing the block, clear any channels from the write array that
3622 * are read by this instruction.
3624 for (unsigned i
= 0; i
< Elements(inst
->src
); i
++) {
3625 if (inst
->src
[i
].file
== PROGRAM_TEMPORARY
&& inst
->src
[i
].reladdr
){
3626 /* Any temporary might be read, so no dead code elimination
3627 * across this instruction.
3629 memset(writes
, 0, sizeof(*writes
) * this->next_temp
* 4);
3630 } else if (inst
->src
[i
].file
== PROGRAM_TEMPORARY
) {
3631 /* Clear where it's used as src. */
3632 int src_chans
= 1 << GET_SWZ(inst
->src
[i
].swizzle
, 0);
3633 src_chans
|= 1 << GET_SWZ(inst
->src
[i
].swizzle
, 1);
3634 src_chans
|= 1 << GET_SWZ(inst
->src
[i
].swizzle
, 2);
3635 src_chans
|= 1 << GET_SWZ(inst
->src
[i
].swizzle
, 3);
3637 for (int c
= 0; c
< 4; c
++) {
3638 if (src_chans
& (1 << c
)) {
3639 writes
[4 * inst
->src
[i
].index
+ c
] = NULL
;
3647 /* If this instruction writes to a temporary, add it to the write array.
3648 * If there is already an instruction in the write array for one or more
3649 * of the channels, flag that channel write as dead.
3651 if (inst
->dst
.file
== PROGRAM_TEMPORARY
&&
3652 !inst
->dst
.reladdr
&&
3654 for (int c
= 0; c
< 4; c
++) {
3655 if (inst
->dst
.writemask
& (1 << c
)) {
3656 if (writes
[4 * inst
->dst
.index
+ c
]) {
3657 if (write_level
[4 * inst
->dst
.index
+ c
] < level
)
3660 writes
[4 * inst
->dst
.index
+ c
]->dead_mask
|= (1 << c
);
3662 writes
[4 * inst
->dst
.index
+ c
] = inst
;
3663 write_level
[4 * inst
->dst
.index
+ c
] = level
;
3669 /* Anything still in the write array at this point is dead code. */
3670 for (int r
= 0; r
< this->next_temp
; r
++) {
3671 for (int c
= 0; c
< 4; c
++) {
3672 glsl_to_tgsi_instruction
*inst
= writes
[4 * r
+ c
];
3674 inst
->dead_mask
|= (1 << c
);
3678 /* Now actually remove the instructions that are completely dead and update
3679 * the writemask of other instructions with dead channels.
3681 foreach_iter(exec_list_iterator
, iter
, this->instructions
) {
3682 glsl_to_tgsi_instruction
*inst
= (glsl_to_tgsi_instruction
*)iter
.get();
3684 if (!inst
->dead_mask
|| !inst
->dst
.writemask
)
3686 else if ((inst
->dst
.writemask
& ~inst
->dead_mask
) == 0) {
3691 inst
->dst
.writemask
&= ~(inst
->dead_mask
);
3694 ralloc_free(write_level
);
3695 ralloc_free(writes
);
3700 /* Merges temporary registers together where possible to reduce the number of
3701 * registers needed to run a program.
3703 * Produces optimal code only after copy propagation and dead code elimination
3706 glsl_to_tgsi_visitor::merge_registers(void)
3708 int *last_reads
= rzalloc_array(mem_ctx
, int, this->next_temp
);
3709 int *first_writes
= rzalloc_array(mem_ctx
, int, this->next_temp
);
3712 /* Read the indices of the last read and first write to each temp register
3713 * into an array so that we don't have to traverse the instruction list as
3715 for (i
=0; i
< this->next_temp
; i
++) {
3716 last_reads
[i
] = get_last_temp_read(i
);
3717 first_writes
[i
] = get_first_temp_write(i
);
3720 /* Start looking for registers with non-overlapping usages that can be
3721 * merged together. */
3722 for (i
=0; i
< this->next_temp
; i
++) {
3723 /* Don't touch unused registers. */
3724 if (last_reads
[i
] < 0 || first_writes
[i
] < 0) continue;
3726 for (j
=0; j
< this->next_temp
; j
++) {
3727 /* Don't touch unused registers. */
3728 if (last_reads
[j
] < 0 || first_writes
[j
] < 0) continue;
3730 /* We can merge the two registers if the first write to j is after or
3731 * in the same instruction as the last read from i. Note that the
3732 * register at index i will always be used earlier or at the same time
3733 * as the register at index j. */
3734 if (first_writes
[i
] <= first_writes
[j
] &&
3735 last_reads
[i
] <= first_writes
[j
])
3737 rename_temp_register(j
, i
); /* Replace all references to j with i.*/
3739 /* Update the first_writes and last_reads arrays with the new
3740 * values for the merged register index, and mark the newly unused
3741 * register index as such. */
3742 last_reads
[i
] = last_reads
[j
];
3743 first_writes
[j
] = -1;
3749 ralloc_free(last_reads
);
3750 ralloc_free(first_writes
);
3753 /* Reassign indices to temporary registers by reusing unused indices created
3754 * by optimization passes. */
3756 glsl_to_tgsi_visitor::renumber_registers(void)
3761 for (i
=0; i
< this->next_temp
; i
++) {
3762 if (get_first_temp_read(i
) < 0) continue;
3764 rename_temp_register(i
, new_index
);
3768 this->next_temp
= new_index
;
3772 * Returns a fragment program which implements the current pixel transfer ops.
3773 * Based on get_pixel_transfer_program in st_atom_pixeltransfer.c.
3776 get_pixel_transfer_visitor(struct st_fragment_program
*fp
,
3777 glsl_to_tgsi_visitor
*original
,
3778 int scale_and_bias
, int pixel_maps
)
3780 glsl_to_tgsi_visitor
*v
= new glsl_to_tgsi_visitor();
3781 struct st_context
*st
= st_context(original
->ctx
);
3782 struct gl_program
*prog
= &fp
->Base
.Base
;
3783 struct gl_program_parameter_list
*params
= _mesa_new_parameter_list();
3784 st_src_reg coord
, src0
;
3786 glsl_to_tgsi_instruction
*inst
;
3788 /* Copy attributes of the glsl_to_tgsi_visitor in the original shader. */
3789 v
->ctx
= original
->ctx
;
3791 v
->shader_program
= NULL
;
3792 v
->glsl_version
= original
->glsl_version
;
3793 v
->native_integers
= original
->native_integers
;
3794 v
->options
= original
->options
;
3795 v
->next_temp
= original
->next_temp
;
3796 v
->num_address_regs
= original
->num_address_regs
;
3797 v
->samplers_used
= prog
->SamplersUsed
= original
->samplers_used
;
3798 v
->indirect_addr_temps
= original
->indirect_addr_temps
;
3799 v
->indirect_addr_consts
= original
->indirect_addr_consts
;
3800 memcpy(&v
->immediates
, &original
->immediates
, sizeof(v
->immediates
));
3801 v
->num_immediates
= original
->num_immediates
;
3804 * Get initial pixel color from the texture.
3805 * TEX colorTemp, fragment.texcoord[0], texture[0], 2D;
3807 coord
= st_src_reg(PROGRAM_INPUT
, FRAG_ATTRIB_TEX0
, glsl_type::vec2_type
);
3808 src0
= v
->get_temp(glsl_type::vec4_type
);
3809 dst0
= st_dst_reg(src0
);
3810 inst
= v
->emit(NULL
, TGSI_OPCODE_TEX
, dst0
, coord
);
3812 inst
->tex_target
= TEXTURE_2D_INDEX
;
3814 prog
->InputsRead
|= FRAG_BIT_TEX0
;
3815 prog
->SamplersUsed
|= (1 << 0); /* mark sampler 0 as used */
3816 v
->samplers_used
|= (1 << 0);
3818 if (scale_and_bias
) {
3819 static const gl_state_index scale_state
[STATE_LENGTH
] =
3820 { STATE_INTERNAL
, STATE_PT_SCALE
,
3821 (gl_state_index
) 0, (gl_state_index
) 0, (gl_state_index
) 0 };
3822 static const gl_state_index bias_state
[STATE_LENGTH
] =
3823 { STATE_INTERNAL
, STATE_PT_BIAS
,
3824 (gl_state_index
) 0, (gl_state_index
) 0, (gl_state_index
) 0 };
3825 GLint scale_p
, bias_p
;
3826 st_src_reg scale
, bias
;
3828 scale_p
= _mesa_add_state_reference(params
, scale_state
);
3829 bias_p
= _mesa_add_state_reference(params
, bias_state
);
3831 /* MAD colorTemp, colorTemp, scale, bias; */
3832 scale
= st_src_reg(PROGRAM_STATE_VAR
, scale_p
, GLSL_TYPE_FLOAT
);
3833 bias
= st_src_reg(PROGRAM_STATE_VAR
, bias_p
, GLSL_TYPE_FLOAT
);
3834 inst
= v
->emit(NULL
, TGSI_OPCODE_MAD
, dst0
, src0
, scale
, bias
);
3838 st_src_reg temp
= v
->get_temp(glsl_type::vec4_type
);
3839 st_dst_reg temp_dst
= st_dst_reg(temp
);
3841 assert(st
->pixel_xfer
.pixelmap_texture
);
3843 /* With a little effort, we can do four pixel map look-ups with
3844 * two TEX instructions:
3847 /* TEX temp.rg, colorTemp.rgba, texture[1], 2D; */
3848 temp_dst
.writemask
= WRITEMASK_XY
; /* write R,G */
3849 inst
= v
->emit(NULL
, TGSI_OPCODE_TEX
, temp_dst
, src0
);
3851 inst
->tex_target
= TEXTURE_2D_INDEX
;
3853 /* TEX temp.ba, colorTemp.baba, texture[1], 2D; */
3854 src0
.swizzle
= MAKE_SWIZZLE4(SWIZZLE_Z
, SWIZZLE_W
, SWIZZLE_Z
, SWIZZLE_W
);
3855 temp_dst
.writemask
= WRITEMASK_ZW
; /* write B,A */
3856 inst
= v
->emit(NULL
, TGSI_OPCODE_TEX
, temp_dst
, src0
);
3858 inst
->tex_target
= TEXTURE_2D_INDEX
;
3860 prog
->SamplersUsed
|= (1 << 1); /* mark sampler 1 as used */
3861 v
->samplers_used
|= (1 << 1);
3863 /* MOV colorTemp, temp; */
3864 inst
= v
->emit(NULL
, TGSI_OPCODE_MOV
, dst0
, temp
);
3867 /* Now copy the instructions from the original glsl_to_tgsi_visitor into the
3869 foreach_iter(exec_list_iterator
, iter
, original
->instructions
) {
3870 glsl_to_tgsi_instruction
*inst
= (glsl_to_tgsi_instruction
*)iter
.get();
3871 glsl_to_tgsi_instruction
*newinst
;
3872 st_src_reg src_regs
[3];
3874 if (inst
->dst
.file
== PROGRAM_OUTPUT
)
3875 prog
->OutputsWritten
|= BITFIELD64_BIT(inst
->dst
.index
);
3877 for (int i
=0; i
<3; i
++) {
3878 src_regs
[i
] = inst
->src
[i
];
3879 if (src_regs
[i
].file
== PROGRAM_INPUT
&&
3880 src_regs
[i
].index
== FRAG_ATTRIB_COL0
)
3882 src_regs
[i
].file
= PROGRAM_TEMPORARY
;
3883 src_regs
[i
].index
= src0
.index
;
3885 else if (src_regs
[i
].file
== PROGRAM_INPUT
)
3886 prog
->InputsRead
|= BITFIELD64_BIT(src_regs
[i
].index
);
3889 newinst
= v
->emit(NULL
, inst
->op
, inst
->dst
, src_regs
[0], src_regs
[1], src_regs
[2]);
3890 newinst
->tex_target
= inst
->tex_target
;
3893 /* Make modifications to fragment program info. */
3894 prog
->Parameters
= _mesa_combine_parameter_lists(params
,
3895 original
->prog
->Parameters
);
3896 _mesa_free_parameter_list(params
);
3897 count_resources(v
, prog
);
3898 fp
->glsl_to_tgsi
= v
;
3902 * Make fragment program for glBitmap:
3903 * Sample the texture and kill the fragment if the bit is 0.
3904 * This program will be combined with the user's fragment program.
3906 * Based on make_bitmap_fragment_program in st_cb_bitmap.c.
3909 get_bitmap_visitor(struct st_fragment_program
*fp
,
3910 glsl_to_tgsi_visitor
*original
, int samplerIndex
)
3912 glsl_to_tgsi_visitor
*v
= new glsl_to_tgsi_visitor();
3913 struct st_context
*st
= st_context(original
->ctx
);
3914 struct gl_program
*prog
= &fp
->Base
.Base
;
3915 st_src_reg coord
, src0
;
3917 glsl_to_tgsi_instruction
*inst
;
3919 /* Copy attributes of the glsl_to_tgsi_visitor in the original shader. */
3920 v
->ctx
= original
->ctx
;
3922 v
->shader_program
= NULL
;
3923 v
->glsl_version
= original
->glsl_version
;
3924 v
->native_integers
= original
->native_integers
;
3925 v
->options
= original
->options
;
3926 v
->next_temp
= original
->next_temp
;
3927 v
->num_address_regs
= original
->num_address_regs
;
3928 v
->samplers_used
= prog
->SamplersUsed
= original
->samplers_used
;
3929 v
->indirect_addr_temps
= original
->indirect_addr_temps
;
3930 v
->indirect_addr_consts
= original
->indirect_addr_consts
;
3931 memcpy(&v
->immediates
, &original
->immediates
, sizeof(v
->immediates
));
3932 v
->num_immediates
= original
->num_immediates
;
3934 /* TEX tmp0, fragment.texcoord[0], texture[0], 2D; */
3935 coord
= st_src_reg(PROGRAM_INPUT
, FRAG_ATTRIB_TEX0
, glsl_type::vec2_type
);
3936 src0
= v
->get_temp(glsl_type::vec4_type
);
3937 dst0
= st_dst_reg(src0
);
3938 inst
= v
->emit(NULL
, TGSI_OPCODE_TEX
, dst0
, coord
);
3939 inst
->sampler
= samplerIndex
;
3940 inst
->tex_target
= TEXTURE_2D_INDEX
;
3942 prog
->InputsRead
|= FRAG_BIT_TEX0
;
3943 prog
->SamplersUsed
|= (1 << samplerIndex
); /* mark sampler as used */
3944 v
->samplers_used
|= (1 << samplerIndex
);
3946 /* KIL if -tmp0 < 0 # texel=0 -> keep / texel=0 -> discard */
3947 src0
.negate
= NEGATE_XYZW
;
3948 if (st
->bitmap
.tex_format
== PIPE_FORMAT_L8_UNORM
)
3949 src0
.swizzle
= SWIZZLE_XXXX
;
3950 inst
= v
->emit(NULL
, TGSI_OPCODE_KIL
, undef_dst
, src0
);
3952 /* Now copy the instructions from the original glsl_to_tgsi_visitor into the
3954 foreach_iter(exec_list_iterator
, iter
, original
->instructions
) {
3955 glsl_to_tgsi_instruction
*inst
= (glsl_to_tgsi_instruction
*)iter
.get();
3956 glsl_to_tgsi_instruction
*newinst
;
3957 st_src_reg src_regs
[3];
3959 if (inst
->dst
.file
== PROGRAM_OUTPUT
)
3960 prog
->OutputsWritten
|= BITFIELD64_BIT(inst
->dst
.index
);
3962 for (int i
=0; i
<3; i
++) {
3963 src_regs
[i
] = inst
->src
[i
];
3964 if (src_regs
[i
].file
== PROGRAM_INPUT
)
3965 prog
->InputsRead
|= BITFIELD64_BIT(src_regs
[i
].index
);
3968 newinst
= v
->emit(NULL
, inst
->op
, inst
->dst
, src_regs
[0], src_regs
[1], src_regs
[2]);
3969 newinst
->tex_target
= inst
->tex_target
;
3972 /* Make modifications to fragment program info. */
3973 prog
->Parameters
= _mesa_clone_parameter_list(original
->prog
->Parameters
);
3974 count_resources(v
, prog
);
3975 fp
->glsl_to_tgsi
= v
;
3978 /* ------------------------- TGSI conversion stuff -------------------------- */
3980 unsigned branch_target
;
3985 * Intermediate state used during shader translation.
3987 struct st_translate
{
3988 struct ureg_program
*ureg
;
3990 struct ureg_dst temps
[MAX_TEMPS
];
3991 struct ureg_src
*constants
;
3992 struct ureg_src
*immediates
;
3993 struct ureg_dst outputs
[PIPE_MAX_SHADER_OUTPUTS
];
3994 struct ureg_src inputs
[PIPE_MAX_SHADER_INPUTS
];
3995 struct ureg_dst address
[1];
3996 struct ureg_src samplers
[PIPE_MAX_SAMPLERS
];
3997 struct ureg_src systemValues
[SYSTEM_VALUE_MAX
];
3999 const GLuint
*inputMapping
;
4000 const GLuint
*outputMapping
;
4002 /* For every instruction that contains a label (eg CALL), keep
4003 * details so that we can go back afterwards and emit the correct
4004 * tgsi instruction number for each label.
4006 struct label
*labels
;
4007 unsigned labels_size
;
4008 unsigned labels_count
;
4010 /* Keep a record of the tgsi instruction number that each mesa
4011 * instruction starts at, will be used to fix up labels after
4016 unsigned insn_count
;
4018 unsigned procType
; /**< TGSI_PROCESSOR_VERTEX/FRAGMENT */
4023 /** Map Mesa's SYSTEM_VALUE_x to TGSI_SEMANTIC_x */
4024 static unsigned mesa_sysval_to_semantic
[SYSTEM_VALUE_MAX
] = {
4026 TGSI_SEMANTIC_VERTEXID
,
4027 TGSI_SEMANTIC_INSTANCEID
4031 * Make note of a branch to a label in the TGSI code.
4032 * After we've emitted all instructions, we'll go over the list
4033 * of labels built here and patch the TGSI code with the actual
4034 * location of each label.
4036 static unsigned *get_label(struct st_translate
*t
, unsigned branch_target
)
4040 if (t
->labels_count
+ 1 >= t
->labels_size
) {
4041 t
->labels_size
= 1 << (util_logbase2(t
->labels_size
) + 1);
4042 t
->labels
= (struct label
*)realloc(t
->labels
,
4043 t
->labels_size
* sizeof(struct label
));
4044 if (t
->labels
== NULL
) {
4045 static unsigned dummy
;
4051 i
= t
->labels_count
++;
4052 t
->labels
[i
].branch_target
= branch_target
;
4053 return &t
->labels
[i
].token
;
4057 * Called prior to emitting the TGSI code for each instruction.
4058 * Allocate additional space for instructions if needed.
4059 * Update the insn[] array so the next glsl_to_tgsi_instruction points to
4060 * the next TGSI instruction.
4062 static void set_insn_start(struct st_translate
*t
, unsigned start
)
4064 if (t
->insn_count
+ 1 >= t
->insn_size
) {
4065 t
->insn_size
= 1 << (util_logbase2(t
->insn_size
) + 1);
4066 t
->insn
= (unsigned *)realloc(t
->insn
, t
->insn_size
* sizeof(t
->insn
[0]));
4067 if (t
->insn
== NULL
) {
4073 t
->insn
[t
->insn_count
++] = start
;
4077 * Map a glsl_to_tgsi constant/immediate to a TGSI immediate.
4079 static struct ureg_src
4080 emit_immediate(struct st_translate
*t
,
4081 gl_constant_value values
[4],
4084 struct ureg_program
*ureg
= t
->ureg
;
4089 return ureg_DECL_immediate(ureg
, &values
[0].f
, size
);
4091 return ureg_DECL_immediate_int(ureg
, &values
[0].i
, size
);
4092 case GL_UNSIGNED_INT
:
4094 return ureg_DECL_immediate_uint(ureg
, &values
[0].u
, size
);
4096 assert(!"should not get here - type must be float, int, uint, or bool");
4097 return ureg_src_undef();
4102 * Map a glsl_to_tgsi dst register to a TGSI ureg_dst register.
4104 static struct ureg_dst
4105 dst_register(struct st_translate
*t
,
4106 gl_register_file file
,
4110 case PROGRAM_UNDEFINED
:
4111 return ureg_dst_undef();
4113 case PROGRAM_TEMPORARY
:
4114 if (ureg_dst_is_undef(t
->temps
[index
]))
4115 t
->temps
[index
] = ureg_DECL_local_temporary(t
->ureg
);
4117 return t
->temps
[index
];
4119 case PROGRAM_OUTPUT
:
4120 if (t
->procType
== TGSI_PROCESSOR_VERTEX
)
4121 assert(index
< VERT_RESULT_MAX
);
4122 else if (t
->procType
== TGSI_PROCESSOR_FRAGMENT
)
4123 assert(index
< FRAG_RESULT_MAX
);
4125 assert(index
< GEOM_RESULT_MAX
);
4127 assert(t
->outputMapping
[index
] < Elements(t
->outputs
));
4129 return t
->outputs
[t
->outputMapping
[index
]];
4131 case PROGRAM_ADDRESS
:
4132 return t
->address
[index
];
4135 assert(!"unknown dst register file");
4136 return ureg_dst_undef();
4141 * Map a glsl_to_tgsi src register to a TGSI ureg_src register.
4143 static struct ureg_src
4144 src_register(struct st_translate
*t
,
4145 gl_register_file file
,
4146 GLint index
, GLint index2D
)
4149 case PROGRAM_UNDEFINED
:
4150 return ureg_src_undef();
4152 case PROGRAM_TEMPORARY
:
4154 assert(index
< (int) Elements(t
->temps
));
4155 if (ureg_dst_is_undef(t
->temps
[index
]))
4156 t
->temps
[index
] = ureg_DECL_local_temporary(t
->ureg
);
4157 return ureg_src(t
->temps
[index
]);
4159 case PROGRAM_ENV_PARAM
:
4160 case PROGRAM_LOCAL_PARAM
:
4161 case PROGRAM_UNIFORM
:
4163 return t
->constants
[index
];
4164 case PROGRAM_STATE_VAR
:
4165 case PROGRAM_CONSTANT
: /* ie, immediate */
4167 struct ureg_src src
;
4168 src
= ureg_src_register(TGSI_FILE_CONSTANT
, 0);
4170 src
.DimensionIndex
= index2D
;
4172 } else if (index
< 0)
4173 return ureg_DECL_constant(t
->ureg
, 0);
4175 return t
->constants
[index
];
4177 case PROGRAM_IMMEDIATE
:
4178 return t
->immediates
[index
];
4181 assert(t
->inputMapping
[index
] < Elements(t
->inputs
));
4182 return t
->inputs
[t
->inputMapping
[index
]];
4184 case PROGRAM_OUTPUT
:
4185 assert(t
->outputMapping
[index
] < Elements(t
->outputs
));
4186 return ureg_src(t
->outputs
[t
->outputMapping
[index
]]); /* not needed? */
4188 case PROGRAM_ADDRESS
:
4189 return ureg_src(t
->address
[index
]);
4191 case PROGRAM_SYSTEM_VALUE
:
4192 assert(index
< (int) Elements(t
->systemValues
));
4193 return t
->systemValues
[index
];
4196 assert(!"unknown src register file");
4197 return ureg_src_undef();
4202 * Create a TGSI ureg_dst register from an st_dst_reg.
4204 static struct ureg_dst
4205 translate_dst(struct st_translate
*t
,
4206 const st_dst_reg
*dst_reg
,
4207 bool saturate
, bool clamp_color
)
4209 struct ureg_dst dst
= dst_register(t
,
4213 dst
= ureg_writemask(dst
, dst_reg
->writemask
);
4216 dst
= ureg_saturate(dst
);
4217 else if (clamp_color
&& dst_reg
->file
== PROGRAM_OUTPUT
) {
4218 /* Clamp colors for ARB_color_buffer_float. */
4219 switch (t
->procType
) {
4220 case TGSI_PROCESSOR_VERTEX
:
4221 /* XXX if the geometry shader is present, this must be done there
4222 * instead of here. */
4223 if (dst_reg
->index
== VERT_RESULT_COL0
||
4224 dst_reg
->index
== VERT_RESULT_COL1
||
4225 dst_reg
->index
== VERT_RESULT_BFC0
||
4226 dst_reg
->index
== VERT_RESULT_BFC1
) {
4227 dst
= ureg_saturate(dst
);
4231 case TGSI_PROCESSOR_FRAGMENT
:
4232 if (dst_reg
->index
>= FRAG_RESULT_COLOR
) {
4233 dst
= ureg_saturate(dst
);
4239 if (dst_reg
->reladdr
!= NULL
)
4240 dst
= ureg_dst_indirect(dst
, ureg_src(t
->address
[0]));
4246 * Create a TGSI ureg_src register from an st_src_reg.
4248 static struct ureg_src
4249 translate_src(struct st_translate
*t
, const st_src_reg
*src_reg
)
4251 struct ureg_src src
= src_register(t
, src_reg
->file
, src_reg
->index
, src_reg
->index2D
);
4253 src
= ureg_swizzle(src
,
4254 GET_SWZ(src_reg
->swizzle
, 0) & 0x3,
4255 GET_SWZ(src_reg
->swizzle
, 1) & 0x3,
4256 GET_SWZ(src_reg
->swizzle
, 2) & 0x3,
4257 GET_SWZ(src_reg
->swizzle
, 3) & 0x3);
4259 if ((src_reg
->negate
& 0xf) == NEGATE_XYZW
)
4260 src
= ureg_negate(src
);
4262 if (src_reg
->reladdr
!= NULL
) {
4263 /* Normally ureg_src_indirect() would be used here, but a stupid compiler
4264 * bug in g++ makes ureg_src_indirect (an inline C function) erroneously
4265 * set the bit for src.Negate. So we have to do the operation manually
4266 * here to work around the compiler's problems. */
4267 /*src = ureg_src_indirect(src, ureg_src(t->address[0]));*/
4268 struct ureg_src addr
= ureg_src(t
->address
[0]);
4270 src
.IndirectFile
= addr
.File
;
4271 src
.IndirectIndex
= addr
.Index
;
4272 src
.IndirectSwizzle
= addr
.SwizzleX
;
4274 if (src_reg
->file
!= PROGRAM_INPUT
&&
4275 src_reg
->file
!= PROGRAM_OUTPUT
) {
4276 /* If src_reg->index was negative, it was set to zero in
4277 * src_register(). Reassign it now. But don't do this
4278 * for input/output regs since they get remapped while
4279 * const buffers don't.
4281 src
.Index
= src_reg
->index
;
4288 static struct tgsi_texture_offset
4289 translate_tex_offset(struct st_translate
*t
,
4290 const struct tgsi_texture_offset
*in_offset
)
4292 struct tgsi_texture_offset offset
;
4293 struct ureg_src imm_src
;
4295 assert(in_offset
->File
== PROGRAM_IMMEDIATE
);
4296 imm_src
= t
->immediates
[in_offset
->Index
];
4298 offset
.File
= imm_src
.File
;
4299 offset
.Index
= imm_src
.Index
;
4300 offset
.SwizzleX
= imm_src
.SwizzleX
;
4301 offset
.SwizzleY
= imm_src
.SwizzleY
;
4302 offset
.SwizzleZ
= imm_src
.SwizzleZ
;
4303 offset
.File
= TGSI_FILE_IMMEDIATE
;
4310 compile_tgsi_instruction(struct st_translate
*t
,
4311 const glsl_to_tgsi_instruction
*inst
,
4312 bool clamp_dst_color_output
)
4314 struct ureg_program
*ureg
= t
->ureg
;
4316 struct ureg_dst dst
[1];
4317 struct ureg_src src
[4];
4318 struct tgsi_texture_offset texoffsets
[MAX_GLSL_TEXTURE_OFFSET
];
4322 unsigned tex_target
;
4324 num_dst
= num_inst_dst_regs(inst
->op
);
4325 num_src
= num_inst_src_regs(inst
->op
);
4328 dst
[0] = translate_dst(t
,
4331 clamp_dst_color_output
);
4333 for (i
= 0; i
< num_src
; i
++)
4334 src
[i
] = translate_src(t
, &inst
->src
[i
]);
4337 case TGSI_OPCODE_BGNLOOP
:
4338 case TGSI_OPCODE_CAL
:
4339 case TGSI_OPCODE_ELSE
:
4340 case TGSI_OPCODE_ENDLOOP
:
4341 case TGSI_OPCODE_IF
:
4342 assert(num_dst
== 0);
4343 ureg_label_insn(ureg
,
4347 inst
->op
== TGSI_OPCODE_CAL
? inst
->function
->sig_id
: 0));
4350 case TGSI_OPCODE_TEX
:
4351 case TGSI_OPCODE_TXB
:
4352 case TGSI_OPCODE_TXD
:
4353 case TGSI_OPCODE_TXL
:
4354 case TGSI_OPCODE_TXP
:
4355 case TGSI_OPCODE_TXQ
:
4356 case TGSI_OPCODE_TXF
:
4357 case TGSI_OPCODE_TEX2
:
4358 case TGSI_OPCODE_TXB2
:
4359 case TGSI_OPCODE_TXL2
:
4360 src
[num_src
++] = t
->samplers
[inst
->sampler
];
4361 for (i
= 0; i
< inst
->tex_offset_num_offset
; i
++) {
4362 texoffsets
[i
] = translate_tex_offset(t
, &inst
->tex_offsets
[i
]);
4364 tex_target
= st_translate_texture_target(inst
->tex_target
, inst
->tex_shadow
);
4370 texoffsets
, inst
->tex_offset_num_offset
,
4374 case TGSI_OPCODE_SCS
:
4375 dst
[0] = ureg_writemask(dst
[0], TGSI_WRITEMASK_XY
);
4376 ureg_insn(ureg
, inst
->op
, dst
, num_dst
, src
, num_src
);
4389 * Emit the TGSI instructions for inverting and adjusting WPOS.
4390 * This code is unavoidable because it also depends on whether
4391 * a FBO is bound (STATE_FB_WPOS_Y_TRANSFORM).
4394 emit_wpos_adjustment( struct st_translate
*t
,
4395 const struct gl_program
*program
,
4397 GLfloat adjX
, GLfloat adjY
[2])
4399 struct ureg_program
*ureg
= t
->ureg
;
4401 /* Fragment program uses fragment position input.
4402 * Need to replace instances of INPUT[WPOS] with temp T
4403 * where T = INPUT[WPOS] by y is inverted.
4405 static const gl_state_index wposTransformState
[STATE_LENGTH
]
4406 = { STATE_INTERNAL
, STATE_FB_WPOS_Y_TRANSFORM
,
4407 (gl_state_index
)0, (gl_state_index
)0, (gl_state_index
)0 };
4409 /* XXX: note we are modifying the incoming shader here! Need to
4410 * do this before emitting the constant decls below, or this
4413 unsigned wposTransConst
= _mesa_add_state_reference(program
->Parameters
,
4414 wposTransformState
);
4416 struct ureg_src wpostrans
= ureg_DECL_constant( ureg
, wposTransConst
);
4417 struct ureg_dst wpos_temp
= ureg_DECL_temporary( ureg
);
4418 struct ureg_src wpos_input
= t
->inputs
[t
->inputMapping
[FRAG_ATTRIB_WPOS
]];
4420 /* First, apply the coordinate shift: */
4421 if (adjX
|| adjY
[0] || adjY
[1]) {
4422 if (adjY
[0] != adjY
[1]) {
4423 /* Adjust the y coordinate by adjY[1] or adjY[0] respectively
4424 * depending on whether inversion is actually going to be applied
4425 * or not, which is determined by testing against the inversion
4426 * state variable used below, which will be either +1 or -1.
4428 struct ureg_dst adj_temp
= ureg_DECL_local_temporary(ureg
);
4430 ureg_CMP(ureg
, adj_temp
,
4431 ureg_scalar(wpostrans
, invert
? 2 : 0),
4432 ureg_imm4f(ureg
, adjX
, adjY
[0], 0.0f
, 0.0f
),
4433 ureg_imm4f(ureg
, adjX
, adjY
[1], 0.0f
, 0.0f
));
4434 ureg_ADD(ureg
, wpos_temp
, wpos_input
, ureg_src(adj_temp
));
4436 ureg_ADD(ureg
, wpos_temp
, wpos_input
,
4437 ureg_imm4f(ureg
, adjX
, adjY
[0], 0.0f
, 0.0f
));
4439 wpos_input
= ureg_src(wpos_temp
);
4441 /* MOV wpos_temp, input[wpos]
4443 ureg_MOV( ureg
, wpos_temp
, wpos_input
);
4446 /* Now the conditional y flip: STATE_FB_WPOS_Y_TRANSFORM.xy/zw will be
4447 * inversion/identity, or the other way around if we're drawing to an FBO.
4450 /* MAD wpos_temp.y, wpos_input, wpostrans.xxxx, wpostrans.yyyy
4453 ureg_writemask(wpos_temp
, TGSI_WRITEMASK_Y
),
4455 ureg_scalar(wpostrans
, 0),
4456 ureg_scalar(wpostrans
, 1));
4458 /* MAD wpos_temp.y, wpos_input, wpostrans.zzzz, wpostrans.wwww
4461 ureg_writemask(wpos_temp
, TGSI_WRITEMASK_Y
),
4463 ureg_scalar(wpostrans
, 2),
4464 ureg_scalar(wpostrans
, 3));
4467 /* Use wpos_temp as position input from here on:
4469 t
->inputs
[t
->inputMapping
[FRAG_ATTRIB_WPOS
]] = ureg_src(wpos_temp
);
4474 * Emit fragment position/ooordinate code.
4477 emit_wpos(struct st_context
*st
,
4478 struct st_translate
*t
,
4479 const struct gl_program
*program
,
4480 struct ureg_program
*ureg
)
4482 const struct gl_fragment_program
*fp
=
4483 (const struct gl_fragment_program
*) program
;
4484 struct pipe_screen
*pscreen
= st
->pipe
->screen
;
4485 GLfloat adjX
= 0.0f
;
4486 GLfloat adjY
[2] = { 0.0f
, 0.0f
};
4487 boolean invert
= FALSE
;
4489 /* Query the pixel center conventions supported by the pipe driver and set
4490 * adjX, adjY to help out if it cannot handle the requested one internally.
4492 * The bias of the y-coordinate depends on whether y-inversion takes place
4493 * (adjY[1]) or not (adjY[0]), which is in turn dependent on whether we are
4494 * drawing to an FBO (causes additional inversion), and whether the the pipe
4495 * driver origin and the requested origin differ (the latter condition is
4496 * stored in the 'invert' variable).
4498 * For height = 100 (i = integer, h = half-integer, l = lower, u = upper):
4500 * center shift only:
4505 * l,i -> u,i: ( 0.0 + 1.0) * -1 + 100 = 99
4506 * l,h -> u,h: ( 0.5 + 0.0) * -1 + 100 = 99.5
4507 * u,i -> l,i: (99.0 + 1.0) * -1 + 100 = 0
4508 * u,h -> l,h: (99.5 + 0.0) * -1 + 100 = 0.5
4510 * inversion and center shift:
4511 * l,i -> u,h: ( 0.0 + 0.5) * -1 + 100 = 99.5
4512 * l,h -> u,i: ( 0.5 + 0.5) * -1 + 100 = 99
4513 * u,i -> l,h: (99.0 + 0.5) * -1 + 100 = 0.5
4514 * u,h -> l,i: (99.5 + 0.5) * -1 + 100 = 0
4516 if (fp
->OriginUpperLeft
) {
4517 /* Fragment shader wants origin in upper-left */
4518 if (pscreen
->get_param(pscreen
, PIPE_CAP_TGSI_FS_COORD_ORIGIN_UPPER_LEFT
)) {
4519 /* the driver supports upper-left origin */
4521 else if (pscreen
->get_param(pscreen
, PIPE_CAP_TGSI_FS_COORD_ORIGIN_LOWER_LEFT
)) {
4522 /* the driver supports lower-left origin, need to invert Y */
4523 ureg_property_fs_coord_origin(ureg
, TGSI_FS_COORD_ORIGIN_LOWER_LEFT
);
4530 /* Fragment shader wants origin in lower-left */
4531 if (pscreen
->get_param(pscreen
, PIPE_CAP_TGSI_FS_COORD_ORIGIN_LOWER_LEFT
))
4532 /* the driver supports lower-left origin */
4533 ureg_property_fs_coord_origin(ureg
, TGSI_FS_COORD_ORIGIN_LOWER_LEFT
);
4534 else if (pscreen
->get_param(pscreen
, PIPE_CAP_TGSI_FS_COORD_ORIGIN_UPPER_LEFT
))
4535 /* the driver supports upper-left origin, need to invert Y */
4541 if (fp
->PixelCenterInteger
) {
4542 /* Fragment shader wants pixel center integer */
4543 if (pscreen
->get_param(pscreen
, PIPE_CAP_TGSI_FS_COORD_PIXEL_CENTER_INTEGER
)) {
4544 /* the driver supports pixel center integer */
4546 ureg_property_fs_coord_pixel_center(ureg
, TGSI_FS_COORD_PIXEL_CENTER_INTEGER
);
4548 else if (pscreen
->get_param(pscreen
, PIPE_CAP_TGSI_FS_COORD_PIXEL_CENTER_HALF_INTEGER
)) {
4549 /* the driver supports pixel center half integer, need to bias X,Y */
4558 /* Fragment shader wants pixel center half integer */
4559 if (pscreen
->get_param(pscreen
, PIPE_CAP_TGSI_FS_COORD_PIXEL_CENTER_HALF_INTEGER
)) {
4560 /* the driver supports pixel center half integer */
4562 else if (pscreen
->get_param(pscreen
, PIPE_CAP_TGSI_FS_COORD_PIXEL_CENTER_INTEGER
)) {
4563 /* the driver supports pixel center integer, need to bias X,Y */
4564 adjX
= adjY
[0] = adjY
[1] = 0.5f
;
4565 ureg_property_fs_coord_pixel_center(ureg
, TGSI_FS_COORD_PIXEL_CENTER_INTEGER
);
4571 /* we invert after adjustment so that we avoid the MOV to temporary,
4572 * and reuse the adjustment ADD instead */
4573 emit_wpos_adjustment(t
, program
, invert
, adjX
, adjY
);
4577 * OpenGL's fragment gl_FrontFace input is 1 for front-facing, 0 for back.
4578 * TGSI uses +1 for front, -1 for back.
4579 * This function converts the TGSI value to the GL value. Simply clamping/
4580 * saturating the value to [0,1] does the job.
4583 emit_face_var(struct st_translate
*t
)
4585 struct ureg_program
*ureg
= t
->ureg
;
4586 struct ureg_dst face_temp
= ureg_DECL_temporary(ureg
);
4587 struct ureg_src face_input
= t
->inputs
[t
->inputMapping
[FRAG_ATTRIB_FACE
]];
4589 /* MOV_SAT face_temp, input[face] */
4590 face_temp
= ureg_saturate(face_temp
);
4591 ureg_MOV(ureg
, face_temp
, face_input
);
4593 /* Use face_temp as face input from here on: */
4594 t
->inputs
[t
->inputMapping
[FRAG_ATTRIB_FACE
]] = ureg_src(face_temp
);
4598 emit_edgeflags(struct st_translate
*t
)
4600 struct ureg_program
*ureg
= t
->ureg
;
4601 struct ureg_dst edge_dst
= t
->outputs
[t
->outputMapping
[VERT_RESULT_EDGE
]];
4602 struct ureg_src edge_src
= t
->inputs
[t
->inputMapping
[VERT_ATTRIB_EDGEFLAG
]];
4604 ureg_MOV(ureg
, edge_dst
, edge_src
);
4608 * Translate intermediate IR (glsl_to_tgsi_instruction) to TGSI format.
4609 * \param program the program to translate
4610 * \param numInputs number of input registers used
4611 * \param inputMapping maps Mesa fragment program inputs to TGSI generic
4613 * \param inputSemanticName the TGSI_SEMANTIC flag for each input
4614 * \param inputSemanticIndex the semantic index (ex: which texcoord) for
4616 * \param interpMode the TGSI_INTERPOLATE_LINEAR/PERSP mode for each input
4617 * \param numOutputs number of output registers used
4618 * \param outputMapping maps Mesa fragment program outputs to TGSI
4620 * \param outputSemanticName the TGSI_SEMANTIC flag for each output
4621 * \param outputSemanticIndex the semantic index (ex: which texcoord) for
4624 * \return PIPE_OK or PIPE_ERROR_OUT_OF_MEMORY
4626 extern "C" enum pipe_error
4627 st_translate_program(
4628 struct gl_context
*ctx
,
4630 struct ureg_program
*ureg
,
4631 glsl_to_tgsi_visitor
*program
,
4632 const struct gl_program
*proginfo
,
4634 const GLuint inputMapping
[],
4635 const ubyte inputSemanticName
[],
4636 const ubyte inputSemanticIndex
[],
4637 const GLuint interpMode
[],
4638 const GLboolean is_centroid
[],
4640 const GLuint outputMapping
[],
4641 const ubyte outputSemanticName
[],
4642 const ubyte outputSemanticIndex
[],
4643 boolean passthrough_edgeflags
,
4644 boolean clamp_color
)
4646 struct st_translate
*t
;
4648 enum pipe_error ret
= PIPE_OK
;
4650 assert(numInputs
<= Elements(t
->inputs
));
4651 assert(numOutputs
<= Elements(t
->outputs
));
4653 t
= CALLOC_STRUCT(st_translate
);
4655 ret
= PIPE_ERROR_OUT_OF_MEMORY
;
4659 memset(t
, 0, sizeof *t
);
4661 t
->procType
= procType
;
4662 t
->inputMapping
= inputMapping
;
4663 t
->outputMapping
= outputMapping
;
4666 if (program
->shader_program
) {
4667 for (i
= 0; i
< program
->shader_program
->NumUserUniformStorage
; i
++) {
4668 struct gl_uniform_storage
*const storage
=
4669 &program
->shader_program
->UniformStorage
[i
];
4671 _mesa_uniform_detach_all_driver_storage(storage
);
4676 * Declare input attributes.
4678 if (procType
== TGSI_PROCESSOR_FRAGMENT
) {
4679 for (i
= 0; i
< numInputs
; i
++) {
4680 t
->inputs
[i
] = ureg_DECL_fs_input_cyl_centroid(ureg
,
4681 inputSemanticName
[i
],
4682 inputSemanticIndex
[i
],
4687 if (proginfo
->InputsRead
& FRAG_BIT_WPOS
) {
4688 /* Must do this after setting up t->inputs, and before
4689 * emitting constant references, below:
4691 emit_wpos(st_context(ctx
), t
, proginfo
, ureg
);
4694 if (proginfo
->InputsRead
& FRAG_BIT_FACE
)
4698 * Declare output attributes.
4700 for (i
= 0; i
< numOutputs
; i
++) {
4701 switch (outputSemanticName
[i
]) {
4702 case TGSI_SEMANTIC_POSITION
:
4703 t
->outputs
[i
] = ureg_DECL_output(ureg
,
4704 TGSI_SEMANTIC_POSITION
, /* Z/Depth */
4705 outputSemanticIndex
[i
]);
4706 t
->outputs
[i
] = ureg_writemask(t
->outputs
[i
], TGSI_WRITEMASK_Z
);
4708 case TGSI_SEMANTIC_STENCIL
:
4709 t
->outputs
[i
] = ureg_DECL_output(ureg
,
4710 TGSI_SEMANTIC_STENCIL
, /* Stencil */
4711 outputSemanticIndex
[i
]);
4712 t
->outputs
[i
] = ureg_writemask(t
->outputs
[i
], TGSI_WRITEMASK_Y
);
4714 case TGSI_SEMANTIC_COLOR
:
4715 t
->outputs
[i
] = ureg_DECL_output(ureg
,
4716 TGSI_SEMANTIC_COLOR
,
4717 outputSemanticIndex
[i
]);
4720 assert(!"fragment shader outputs must be POSITION/STENCIL/COLOR");
4721 ret
= PIPE_ERROR_BAD_INPUT
;
4726 else if (procType
== TGSI_PROCESSOR_GEOMETRY
) {
4727 for (i
= 0; i
< numInputs
; i
++) {
4728 t
->inputs
[i
] = ureg_DECL_gs_input(ureg
,
4730 inputSemanticName
[i
],
4731 inputSemanticIndex
[i
]);
4734 for (i
= 0; i
< numOutputs
; i
++) {
4735 t
->outputs
[i
] = ureg_DECL_output(ureg
,
4736 outputSemanticName
[i
],
4737 outputSemanticIndex
[i
]);
4741 assert(procType
== TGSI_PROCESSOR_VERTEX
);
4743 for (i
= 0; i
< numInputs
; i
++) {
4744 t
->inputs
[i
] = ureg_DECL_vs_input(ureg
, i
);
4747 for (i
= 0; i
< numOutputs
; i
++) {
4748 t
->outputs
[i
] = ureg_DECL_output(ureg
,
4749 outputSemanticName
[i
],
4750 outputSemanticIndex
[i
]);
4752 if (passthrough_edgeflags
)
4756 /* Declare address register.
4758 if (program
->num_address_regs
> 0) {
4759 assert(program
->num_address_regs
== 1);
4760 t
->address
[0] = ureg_DECL_address(ureg
);
4763 /* Declare misc input registers
4766 GLbitfield sysInputs
= proginfo
->SystemValuesRead
;
4767 unsigned numSys
= 0;
4768 for (i
= 0; sysInputs
; i
++) {
4769 if (sysInputs
& (1 << i
)) {
4770 unsigned semName
= mesa_sysval_to_semantic
[i
];
4771 t
->systemValues
[i
] = ureg_DECL_system_value(ureg
, numSys
, semName
, 0);
4772 if (semName
== TGSI_SEMANTIC_INSTANCEID
||
4773 semName
== TGSI_SEMANTIC_VERTEXID
) {
4774 /* From Gallium perspective, these system values are always
4775 * integer, and require native integer support. However, if
4776 * native integer is supported on the vertex stage but not the
4777 * pixel stage (e.g, i915g + draw), Mesa will generate IR that
4778 * assumes these system values are floats. To resolve the
4779 * inconsistency, we insert a U2F.
4781 struct st_context
*st
= st_context(ctx
);
4782 struct pipe_screen
*pscreen
= st
->pipe
->screen
;
4783 assert(procType
== TGSI_PROCESSOR_VERTEX
);
4784 assert(pscreen
->get_shader_param(pscreen
, PIPE_SHADER_VERTEX
, PIPE_SHADER_CAP_INTEGERS
));
4785 if (!ctx
->Const
.NativeIntegers
) {
4786 struct ureg_dst temp
= ureg_DECL_local_temporary(t
->ureg
);
4787 ureg_U2F( t
->ureg
, ureg_writemask(temp
, TGSI_WRITEMASK_X
), t
->systemValues
[i
]);
4788 t
->systemValues
[i
] = ureg_scalar(ureg_src(temp
), 0);
4792 sysInputs
&= ~(1 << i
);
4797 if (program
->indirect_addr_temps
) {
4798 /* If temps are accessed with indirect addressing, declare temporaries
4799 * in sequential order. Else, we declare them on demand elsewhere.
4800 * (Note: the number of temporaries is equal to program->next_temp)
4802 for (i
= 0; i
< (unsigned)program
->next_temp
; i
++) {
4803 /* XXX use TGSI_FILE_TEMPORARY_ARRAY when it's supported by ureg */
4804 t
->temps
[i
] = ureg_DECL_local_temporary(t
->ureg
);
4808 /* Emit constants and uniforms. TGSI uses a single index space for these,
4809 * so we put all the translated regs in t->constants.
4811 if (proginfo
->Parameters
) {
4812 t
->constants
= (struct ureg_src
*)
4813 calloc(proginfo
->Parameters
->NumParameters
, sizeof(t
->constants
[0]));
4814 if (t
->constants
== NULL
) {
4815 ret
= PIPE_ERROR_OUT_OF_MEMORY
;
4819 for (i
= 0; i
< proginfo
->Parameters
->NumParameters
; i
++) {
4820 switch (proginfo
->Parameters
->Parameters
[i
].Type
) {
4821 case PROGRAM_ENV_PARAM
:
4822 case PROGRAM_LOCAL_PARAM
:
4823 case PROGRAM_STATE_VAR
:
4824 case PROGRAM_UNIFORM
:
4825 t
->constants
[i
] = ureg_DECL_constant(ureg
, i
);
4828 /* Emit immediates for PROGRAM_CONSTANT only when there's no indirect
4829 * addressing of the const buffer.
4830 * FIXME: Be smarter and recognize param arrays:
4831 * indirect addressing is only valid within the referenced
4834 case PROGRAM_CONSTANT
:
4835 if (program
->indirect_addr_consts
)
4836 t
->constants
[i
] = ureg_DECL_constant(ureg
, i
);
4838 t
->constants
[i
] = emit_immediate(t
,
4839 proginfo
->Parameters
->ParameterValues
[i
],
4840 proginfo
->Parameters
->Parameters
[i
].DataType
,
4849 if (program
->shader_program
) {
4850 unsigned num_ubos
= program
->shader_program
->NumUniformBlocks
;
4852 for (i
= 0; i
< num_ubos
; i
++) {
4853 ureg_DECL_constant2D(t
->ureg
, 0, program
->shader_program
->UniformBlocks
[i
].UniformBufferSize
/ 4, i
+ 1);
4857 /* Emit immediate values.
4859 t
->immediates
= (struct ureg_src
*)
4860 calloc(program
->num_immediates
, sizeof(struct ureg_src
));
4861 if (t
->immediates
== NULL
) {
4862 ret
= PIPE_ERROR_OUT_OF_MEMORY
;
4866 foreach_iter(exec_list_iterator
, iter
, program
->immediates
) {
4867 immediate_storage
*imm
= (immediate_storage
*)iter
.get();
4868 assert(i
< program
->num_immediates
);
4869 t
->immediates
[i
++] = emit_immediate(t
, imm
->values
, imm
->type
, imm
->size
);
4871 assert(i
== program
->num_immediates
);
4873 /* texture samplers */
4874 for (i
= 0; i
< ctx
->Const
.MaxTextureImageUnits
; i
++) {
4875 if (program
->samplers_used
& (1 << i
)) {
4876 t
->samplers
[i
] = ureg_DECL_sampler(ureg
, i
);
4880 /* Emit each instruction in turn:
4882 foreach_iter(exec_list_iterator
, iter
, program
->instructions
) {
4883 set_insn_start(t
, ureg_get_instruction_number(ureg
));
4884 compile_tgsi_instruction(t
, (glsl_to_tgsi_instruction
*)iter
.get(),
4888 /* Fix up all emitted labels:
4890 for (i
= 0; i
< t
->labels_count
; i
++) {
4891 ureg_fixup_label(ureg
, t
->labels
[i
].token
,
4892 t
->insn
[t
->labels
[i
].branch_target
]);
4895 if (program
->shader_program
) {
4896 /* This has to be done last. Any operation the can cause
4897 * prog->ParameterValues to get reallocated (e.g., anything that adds a
4898 * program constant) has to happen before creating this linkage.
4900 for (unsigned i
= 0; i
< MESA_SHADER_TYPES
; i
++) {
4901 if (program
->shader_program
->_LinkedShaders
[i
] == NULL
)
4904 _mesa_associate_uniform_storage(ctx
, program
->shader_program
,
4905 program
->shader_program
->_LinkedShaders
[i
]->Program
->Parameters
);
4914 free(t
->immediates
);
4917 debug_printf("%s: translate error flag set\n", __FUNCTION__
);
4925 /* ----------------------------- End TGSI code ------------------------------ */
4928 * Convert a shader's GLSL IR into a Mesa gl_program, although without
4929 * generating Mesa IR.
4931 static struct gl_program
*
4932 get_mesa_program(struct gl_context
*ctx
,
4933 struct gl_shader_program
*shader_program
,
4934 struct gl_shader
*shader
)
4936 glsl_to_tgsi_visitor
* v
;
4937 struct gl_program
*prog
;
4939 const char *target_string
;
4941 struct gl_shader_compiler_options
*options
=
4942 &ctx
->ShaderCompilerOptions
[_mesa_shader_type_to_index(shader
->Type
)];
4944 switch (shader
->Type
) {
4945 case GL_VERTEX_SHADER
:
4946 target
= GL_VERTEX_PROGRAM_ARB
;
4947 target_string
= "vertex";
4949 case GL_FRAGMENT_SHADER
:
4950 target
= GL_FRAGMENT_PROGRAM_ARB
;
4951 target_string
= "fragment";
4953 case GL_GEOMETRY_SHADER
:
4954 target
= GL_GEOMETRY_PROGRAM_NV
;
4955 target_string
= "geometry";
4958 assert(!"should not be reached");
4962 validate_ir_tree(shader
->ir
);
4964 prog
= ctx
->Driver
.NewProgram(ctx
, target
, shader_program
->Name
);
4967 prog
->Parameters
= _mesa_new_parameter_list();
4968 v
= new glsl_to_tgsi_visitor();
4971 v
->shader_program
= shader_program
;
4972 v
->options
= options
;
4973 v
->glsl_version
= ctx
->Const
.GLSLVersion
;
4974 v
->native_integers
= ctx
->Const
.NativeIntegers
;
4976 _mesa_generate_parameters_list_for_uniforms(shader_program
, shader
,
4979 /* Remove reads from output registers. */
4980 lower_output_reads(shader
->ir
);
4982 /* Emit intermediate IR for main(). */
4983 visit_exec_list(shader
->ir
, v
);
4985 /* Now emit bodies for any functions that were used. */
4987 progress
= GL_FALSE
;
4989 foreach_iter(exec_list_iterator
, iter
, v
->function_signatures
) {
4990 function_entry
*entry
= (function_entry
*)iter
.get();
4992 if (!entry
->bgn_inst
) {
4993 v
->current_function
= entry
;
4995 entry
->bgn_inst
= v
->emit(NULL
, TGSI_OPCODE_BGNSUB
);
4996 entry
->bgn_inst
->function
= entry
;
4998 visit_exec_list(&entry
->sig
->body
, v
);
5000 glsl_to_tgsi_instruction
*last
;
5001 last
= (glsl_to_tgsi_instruction
*)v
->instructions
.get_tail();
5002 if (last
->op
!= TGSI_OPCODE_RET
)
5003 v
->emit(NULL
, TGSI_OPCODE_RET
);
5005 glsl_to_tgsi_instruction
*end
;
5006 end
= v
->emit(NULL
, TGSI_OPCODE_ENDSUB
);
5007 end
->function
= entry
;
5015 /* Print out some information (for debugging purposes) used by the
5016 * optimization passes. */
5017 for (i
=0; i
< v
->next_temp
; i
++) {
5018 int fr
= v
->get_first_temp_read(i
);
5019 int fw
= v
->get_first_temp_write(i
);
5020 int lr
= v
->get_last_temp_read(i
);
5021 int lw
= v
->get_last_temp_write(i
);
5023 printf("Temp %d: FR=%3d FW=%3d LR=%3d LW=%3d\n", i
, fr
, fw
, lr
, lw
);
5028 /* Perform optimizations on the instructions in the glsl_to_tgsi_visitor. */
5030 v
->copy_propagate();
5031 while (v
->eliminate_dead_code_advanced());
5033 /* FIXME: These passes to optimize temporary registers don't work when there
5034 * is indirect addressing of the temporary register space. We need proper
5035 * array support so that we don't have to give up these passes in every
5036 * shader that uses arrays.
5038 if (!v
->indirect_addr_temps
) {
5039 v
->eliminate_dead_code();
5040 v
->merge_registers();
5041 v
->renumber_registers();
5044 /* Write the END instruction. */
5045 v
->emit(NULL
, TGSI_OPCODE_END
);
5047 if (ctx
->Shader
.Flags
& GLSL_DUMP
) {
5049 printf("GLSL IR for linked %s program %d:\n", target_string
,
5050 shader_program
->Name
);
5051 _mesa_print_ir(shader
->ir
, NULL
);
5057 prog
->Instructions
= NULL
;
5058 prog
->NumInstructions
= 0;
5060 do_set_program_inouts(shader
->ir
, prog
, shader
->Type
== GL_FRAGMENT_SHADER
);
5061 count_resources(v
, prog
);
5063 _mesa_reference_program(ctx
, &shader
->Program
, prog
);
5065 /* This has to be done last. Any operation the can cause
5066 * prog->ParameterValues to get reallocated (e.g., anything that adds a
5067 * program constant) has to happen before creating this linkage.
5069 _mesa_associate_uniform_storage(ctx
, shader_program
, prog
->Parameters
);
5070 if (!shader_program
->LinkStatus
) {
5074 struct st_vertex_program
*stvp
;
5075 struct st_fragment_program
*stfp
;
5076 struct st_geometry_program
*stgp
;
5078 switch (shader
->Type
) {
5079 case GL_VERTEX_SHADER
:
5080 stvp
= (struct st_vertex_program
*)prog
;
5081 stvp
->glsl_to_tgsi
= v
;
5083 case GL_FRAGMENT_SHADER
:
5084 stfp
= (struct st_fragment_program
*)prog
;
5085 stfp
->glsl_to_tgsi
= v
;
5087 case GL_GEOMETRY_SHADER
:
5088 stgp
= (struct st_geometry_program
*)prog
;
5089 stgp
->glsl_to_tgsi
= v
;
5092 assert(!"should not be reached");
5102 st_new_shader(struct gl_context
*ctx
, GLuint name
, GLuint type
)
5104 struct gl_shader
*shader
;
5105 assert(type
== GL_FRAGMENT_SHADER
|| type
== GL_VERTEX_SHADER
||
5106 type
== GL_GEOMETRY_SHADER_ARB
);
5107 shader
= rzalloc(NULL
, struct gl_shader
);
5109 shader
->Type
= type
;
5110 shader
->Name
= name
;
5111 _mesa_init_shader(ctx
, shader
);
5116 struct gl_shader_program
*
5117 st_new_shader_program(struct gl_context
*ctx
, GLuint name
)
5119 struct gl_shader_program
*shProg
;
5120 shProg
= rzalloc(NULL
, struct gl_shader_program
);
5122 shProg
->Name
= name
;
5123 _mesa_init_shader_program(ctx
, shProg
);
5130 * Called via ctx->Driver.LinkShader()
5131 * This actually involves converting GLSL IR into an intermediate TGSI-like IR
5132 * with code lowering and other optimizations.
5135 st_link_shader(struct gl_context
*ctx
, struct gl_shader_program
*prog
)
5137 assert(prog
->LinkStatus
);
5139 for (unsigned i
= 0; i
< MESA_SHADER_TYPES
; i
++) {
5140 if (prog
->_LinkedShaders
[i
] == NULL
)
5144 exec_list
*ir
= prog
->_LinkedShaders
[i
]->ir
;
5145 const struct gl_shader_compiler_options
*options
=
5146 &ctx
->ShaderCompilerOptions
[_mesa_shader_type_to_index(prog
->_LinkedShaders
[i
]->Type
)];
5149 unsigned what_to_lower
= MOD_TO_FRACT
| DIV_TO_MUL_RCP
|
5150 EXP_TO_EXP2
| LOG_TO_LOG2
;
5151 if (options
->EmitNoPow
)
5152 what_to_lower
|= POW_TO_EXP2
;
5153 if (!ctx
->Const
.NativeIntegers
)
5154 what_to_lower
|= INT_DIV_TO_MUL_RCP
;
5159 do_mat_op_to_vec(ir
);
5160 lower_instructions(ir
, what_to_lower
);
5162 lower_ubo_reference(prog
->_LinkedShaders
[i
], ir
);
5164 progress
= do_lower_jumps(ir
, true, true, options
->EmitNoMainReturn
, options
->EmitNoCont
, options
->EmitNoLoops
) || progress
;
5166 progress
= do_common_optimization(ir
, true, true,
5167 options
->MaxUnrollIterations
)
5170 progress
= lower_quadop_vector(ir
, false) || progress
;
5172 if (options
->MaxIfDepth
== 0)
5173 progress
= lower_discard(ir
) || progress
;
5175 progress
= lower_if_to_cond_assign(ir
, options
->MaxIfDepth
) || progress
;
5177 if (options
->EmitNoNoise
)
5178 progress
= lower_noise(ir
) || progress
;
5180 /* If there are forms of indirect addressing that the driver
5181 * cannot handle, perform the lowering pass.
5183 if (options
->EmitNoIndirectInput
|| options
->EmitNoIndirectOutput
5184 || options
->EmitNoIndirectTemp
|| options
->EmitNoIndirectUniform
)
5186 lower_variable_index_to_cond_assign(ir
,
5187 options
->EmitNoIndirectInput
,
5188 options
->EmitNoIndirectOutput
,
5189 options
->EmitNoIndirectTemp
,
5190 options
->EmitNoIndirectUniform
)
5193 progress
= do_vec_index_to_cond_assign(ir
) || progress
;
5197 validate_ir_tree(ir
);
5200 for (unsigned i
= 0; i
< MESA_SHADER_TYPES
; i
++) {
5201 struct gl_program
*linked_prog
;
5203 if (prog
->_LinkedShaders
[i
] == NULL
)
5206 linked_prog
= get_mesa_program(ctx
, prog
, prog
->_LinkedShaders
[i
]);
5209 static const GLenum targets
[] = {
5210 GL_VERTEX_PROGRAM_ARB
,
5211 GL_FRAGMENT_PROGRAM_ARB
,
5212 GL_GEOMETRY_PROGRAM_NV
5215 _mesa_reference_program(ctx
, &prog
->_LinkedShaders
[i
]->Program
,
5217 if (!ctx
->Driver
.ProgramStringNotify(ctx
, targets
[i
], linked_prog
)) {
5218 _mesa_reference_program(ctx
, &prog
->_LinkedShaders
[i
]->Program
,
5220 _mesa_reference_program(ctx
, &linked_prog
, NULL
);
5225 _mesa_reference_program(ctx
, &linked_prog
, NULL
);
5232 st_translate_stream_output_info(glsl_to_tgsi_visitor
*glsl_to_tgsi
,
5233 const GLuint outputMapping
[],
5234 struct pipe_stream_output_info
*so
)
5237 struct gl_transform_feedback_info
*info
=
5238 &glsl_to_tgsi
->shader_program
->LinkedTransformFeedback
;
5240 for (i
= 0; i
< info
->NumOutputs
; i
++) {
5241 so
->output
[i
].register_index
=
5242 outputMapping
[info
->Outputs
[i
].OutputRegister
];
5243 so
->output
[i
].start_component
= info
->Outputs
[i
].ComponentOffset
;
5244 so
->output
[i
].num_components
= info
->Outputs
[i
].NumComponents
;
5245 so
->output
[i
].output_buffer
= info
->Outputs
[i
].OutputBuffer
;
5246 so
->output
[i
].dst_offset
= info
->Outputs
[i
].DstOffset
;
5249 for (i
= 0; i
< PIPE_MAX_SO_BUFFERS
; i
++) {
5250 so
->stride
[i
] = info
->BufferStride
[i
];
5252 so
->num_outputs
= info
->NumOutputs
;