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_expression_flattening.h"
38 #include "glsl_types.h"
39 #include "glsl_parser_extras.h"
40 #include "../glsl/program.h"
41 #include "ir_optimization.h"
44 #include "main/mtypes.h"
45 #include "main/shaderobj.h"
46 #include "main/uniforms.h"
47 #include "program/hash_table.h"
50 #include "main/shaderapi.h"
51 #include "program/prog_instruction.h"
52 #include "program/prog_optimize.h"
53 #include "program/prog_print.h"
54 #include "program/program.h"
55 #include "program/prog_parameter.h"
56 #include "program/sampler.h"
58 #include "pipe/p_compiler.h"
59 #include "pipe/p_context.h"
60 #include "pipe/p_screen.h"
61 #include "pipe/p_shader_tokens.h"
62 #include "pipe/p_state.h"
63 #include "util/u_math.h"
64 #include "tgsi/tgsi_ureg.h"
65 #include "tgsi/tgsi_info.h"
66 #include "st_context.h"
67 #include "st_program.h"
68 #include "st_glsl_to_tgsi.h"
69 #include "st_mesa_to_tgsi.h"
72 #define PROGRAM_IMMEDIATE PROGRAM_FILE_MAX
73 #define PROGRAM_ANY_CONST ((1 << PROGRAM_STATE_VAR) | \
74 (1 << PROGRAM_CONSTANT) | \
75 (1 << PROGRAM_UNIFORM))
78 * Maximum number of temporary registers.
80 * It is too big for stack allocated arrays -- it will cause stack overflow on
81 * Windows and likely Mac OS X.
83 #define MAX_TEMPS 4096
86 * Maximum number of arrays
88 #define MAX_ARRAYS 256
90 /* will be 4 for GLSL 4.00 */
91 #define MAX_GLSL_TEXTURE_OFFSET 1
96 static int swizzle_for_size(int size
);
99 * This struct is a corresponding struct to TGSI ureg_src.
103 st_src_reg(gl_register_file file
, int index
, const glsl_type
*type
)
107 if (type
&& (type
->is_scalar() || type
->is_vector() || type
->is_matrix()))
108 this->swizzle
= swizzle_for_size(type
->vector_elements
);
110 this->swizzle
= SWIZZLE_XYZW
;
113 this->type
= type
? type
->base_type
: GLSL_TYPE_ERROR
;
114 this->reladdr
= NULL
;
115 this->reladdr2
= NULL
;
116 this->has_index2
= false;
119 st_src_reg(gl_register_file file
, int index
, int type
)
125 this->swizzle
= SWIZZLE_XYZW
;
127 this->reladdr
= NULL
;
128 this->reladdr2
= NULL
;
129 this->has_index2
= false;
132 st_src_reg(gl_register_file file
, int index
, int type
, int index2D
)
137 this->index2D
= index2D
;
138 this->swizzle
= SWIZZLE_XYZW
;
140 this->reladdr
= NULL
;
141 this->reladdr2
= NULL
;
142 this->has_index2
= false;
147 this->type
= GLSL_TYPE_ERROR
;
148 this->file
= PROGRAM_UNDEFINED
;
153 this->reladdr
= NULL
;
154 this->reladdr2
= NULL
;
155 this->has_index2
= false;
158 explicit st_src_reg(st_dst_reg reg
);
160 gl_register_file file
; /**< PROGRAM_* from Mesa */
161 int index
; /**< temporary index, VERT_ATTRIB_*, VARYING_SLOT_*, etc. */
163 GLuint swizzle
; /**< SWIZZLE_XYZWONEZERO swizzles from Mesa. */
164 int negate
; /**< NEGATE_XYZW mask from mesa */
165 int type
; /** GLSL_TYPE_* from GLSL IR (enum glsl_base_type) */
166 /** Register index should be offset by the integer in this reg. */
168 st_src_reg
*reladdr2
;
174 st_dst_reg(gl_register_file file
, int writemask
, int type
, int index
)
178 this->writemask
= writemask
;
179 this->cond_mask
= COND_TR
;
180 this->reladdr
= NULL
;
184 st_dst_reg(gl_register_file file
, int writemask
, int type
)
188 this->writemask
= writemask
;
189 this->cond_mask
= COND_TR
;
190 this->reladdr
= NULL
;
196 this->type
= GLSL_TYPE_ERROR
;
197 this->file
= PROGRAM_UNDEFINED
;
200 this->cond_mask
= COND_TR
;
201 this->reladdr
= NULL
;
204 explicit st_dst_reg(st_src_reg reg
);
206 gl_register_file file
; /**< PROGRAM_* from Mesa */
207 int index
; /**< temporary index, VERT_ATTRIB_*, VARYING_SLOT_*, etc. */
208 int writemask
; /**< Bitfield of WRITEMASK_[XYZW] */
210 int type
; /** GLSL_TYPE_* from GLSL IR (enum glsl_base_type) */
211 /** Register index should be offset by the integer in this reg. */
215 st_src_reg::st_src_reg(st_dst_reg reg
)
217 this->type
= reg
.type
;
218 this->file
= reg
.file
;
219 this->index
= reg
.index
;
220 this->swizzle
= SWIZZLE_XYZW
;
222 this->reladdr
= reg
.reladdr
;
224 this->reladdr2
= NULL
;
225 this->has_index2
= false;
228 st_dst_reg::st_dst_reg(st_src_reg reg
)
230 this->type
= reg
.type
;
231 this->file
= reg
.file
;
232 this->index
= reg
.index
;
233 this->writemask
= WRITEMASK_XYZW
;
234 this->cond_mask
= COND_TR
;
235 this->reladdr
= reg
.reladdr
;
238 class glsl_to_tgsi_instruction
: public exec_node
{
240 DECLARE_RALLOC_CXX_OPERATORS(glsl_to_tgsi_instruction
)
245 /** Pointer to the ir source this tree came from for debugging */
247 GLboolean cond_update
;
249 int sampler
; /**< sampler index */
250 int tex_target
; /**< One of TEXTURE_*_INDEX */
251 GLboolean tex_shadow
;
252 struct tgsi_texture_offset tex_offsets
[MAX_GLSL_TEXTURE_OFFSET
];
253 unsigned tex_offset_num_offset
;
254 int dead_mask
; /**< Used in dead code elimination */
256 class function_entry
*function
; /* Set on TGSI_OPCODE_CAL or TGSI_OPCODE_BGNSUB */
259 class variable_storage
: public exec_node
{
261 variable_storage(ir_variable
*var
, gl_register_file file
, int index
)
262 : file(file
), index(index
), var(var
)
267 gl_register_file file
;
269 ir_variable
*var
; /* variable that maps to this, if any */
272 class immediate_storage
: public exec_node
{
274 immediate_storage(gl_constant_value
*values
, int size
, int type
)
276 memcpy(this->values
, values
, size
* sizeof(gl_constant_value
));
281 gl_constant_value values
[4];
282 int size
; /**< Number of components (1-4) */
283 int type
; /**< GL_FLOAT, GL_INT, GL_BOOL, or GL_UNSIGNED_INT */
286 class function_entry
: public exec_node
{
288 ir_function_signature
*sig
;
291 * identifier of this function signature used by the program.
293 * At the point that TGSI instructions for function calls are
294 * generated, we don't know the address of the first instruction of
295 * the function body. So we make the BranchTarget that is called a
296 * small integer and rewrite them during set_branchtargets().
301 * Pointer to first instruction of the function body.
303 * Set during function body emits after main() is processed.
305 glsl_to_tgsi_instruction
*bgn_inst
;
308 * Index of the first instruction of the function body in actual TGSI.
310 * Set after conversion from glsl_to_tgsi_instruction to TGSI.
314 /** Storage for the return value. */
315 st_src_reg return_reg
;
318 struct glsl_to_tgsi_visitor
: public ir_visitor
{
320 glsl_to_tgsi_visitor();
321 ~glsl_to_tgsi_visitor();
323 function_entry
*current_function
;
325 struct gl_context
*ctx
;
326 struct gl_program
*prog
;
327 struct gl_shader_program
*shader_program
;
328 struct gl_shader_compiler_options
*options
;
332 unsigned array_sizes
[MAX_ARRAYS
];
335 int num_address_regs
;
337 bool indirect_addr_consts
;
340 bool native_integers
;
343 variable_storage
*find_variable_storage(ir_variable
*var
);
345 int add_constant(gl_register_file file
, gl_constant_value values
[4],
346 int size
, int datatype
, GLuint
*swizzle_out
);
348 function_entry
*get_function_signature(ir_function_signature
*sig
);
350 st_src_reg
get_temp(const glsl_type
*type
);
351 void reladdr_to_temp(ir_instruction
*ir
, st_src_reg
*reg
, int *num_reladdr
);
353 st_src_reg
st_src_reg_for_float(float val
);
354 st_src_reg
st_src_reg_for_int(int val
);
355 st_src_reg
st_src_reg_for_type(int type
, int val
);
358 * \name Visit methods
360 * As typical for the visitor pattern, there must be one \c visit method for
361 * each concrete subclass of \c ir_instruction. Virtual base classes within
362 * the hierarchy should not have \c visit methods.
365 virtual void visit(ir_variable
*);
366 virtual void visit(ir_loop
*);
367 virtual void visit(ir_loop_jump
*);
368 virtual void visit(ir_function_signature
*);
369 virtual void visit(ir_function
*);
370 virtual void visit(ir_expression
*);
371 virtual void visit(ir_swizzle
*);
372 virtual void visit(ir_dereference_variable
*);
373 virtual void visit(ir_dereference_array
*);
374 virtual void visit(ir_dereference_record
*);
375 virtual void visit(ir_assignment
*);
376 virtual void visit(ir_constant
*);
377 virtual void visit(ir_call
*);
378 virtual void visit(ir_return
*);
379 virtual void visit(ir_discard
*);
380 virtual void visit(ir_texture
*);
381 virtual void visit(ir_if
*);
382 virtual void visit(ir_emit_vertex
*);
383 virtual void visit(ir_end_primitive
*);
388 /** List of variable_storage */
391 /** List of immediate_storage */
392 exec_list immediates
;
393 unsigned num_immediates
;
395 /** List of function_entry */
396 exec_list function_signatures
;
397 int next_signature_id
;
399 /** List of glsl_to_tgsi_instruction */
400 exec_list instructions
;
402 glsl_to_tgsi_instruction
*emit(ir_instruction
*ir
, unsigned op
);
404 glsl_to_tgsi_instruction
*emit(ir_instruction
*ir
, unsigned op
,
405 st_dst_reg dst
, st_src_reg src0
);
407 glsl_to_tgsi_instruction
*emit(ir_instruction
*ir
, unsigned op
,
408 st_dst_reg dst
, st_src_reg src0
, st_src_reg src1
);
410 glsl_to_tgsi_instruction
*emit(ir_instruction
*ir
, unsigned op
,
412 st_src_reg src0
, st_src_reg src1
, st_src_reg src2
);
414 unsigned get_opcode(ir_instruction
*ir
, unsigned op
,
416 st_src_reg src0
, st_src_reg src1
);
419 * Emit the correct dot-product instruction for the type of arguments
421 glsl_to_tgsi_instruction
*emit_dp(ir_instruction
*ir
,
427 void emit_scalar(ir_instruction
*ir
, unsigned op
,
428 st_dst_reg dst
, st_src_reg src0
);
430 void emit_scalar(ir_instruction
*ir
, unsigned op
,
431 st_dst_reg dst
, st_src_reg src0
, st_src_reg src1
);
433 void emit_arl(ir_instruction
*ir
, st_dst_reg dst
, st_src_reg src0
);
435 void emit_scs(ir_instruction
*ir
, unsigned op
,
436 st_dst_reg dst
, const st_src_reg
&src
);
438 bool try_emit_mad(ir_expression
*ir
,
440 bool try_emit_mad_for_and_not(ir_expression
*ir
,
442 bool try_emit_sat(ir_expression
*ir
);
444 void emit_swz(ir_expression
*ir
);
446 bool process_move_condition(ir_rvalue
*ir
);
448 void simplify_cmp(void);
450 void rename_temp_register(int index
, int new_index
);
451 int get_first_temp_read(int index
);
452 int get_first_temp_write(int index
);
453 int get_last_temp_read(int index
);
454 int get_last_temp_write(int index
);
456 void copy_propagate(void);
457 void eliminate_dead_code(void);
458 int eliminate_dead_code_advanced(void);
459 void merge_registers(void);
460 void renumber_registers(void);
462 void emit_block_mov(ir_assignment
*ir
, const struct glsl_type
*type
,
463 st_dst_reg
*l
, st_src_reg
*r
);
468 static st_src_reg undef_src
= st_src_reg(PROGRAM_UNDEFINED
, 0, GLSL_TYPE_ERROR
);
470 static st_dst_reg undef_dst
= st_dst_reg(PROGRAM_UNDEFINED
, SWIZZLE_NOOP
, GLSL_TYPE_ERROR
);
472 static st_dst_reg address_reg
= st_dst_reg(PROGRAM_ADDRESS
, WRITEMASK_X
, GLSL_TYPE_FLOAT
, 0);
473 static st_dst_reg address_reg2
= st_dst_reg(PROGRAM_ADDRESS
, WRITEMASK_X
, GLSL_TYPE_FLOAT
, 1);
476 fail_link(struct gl_shader_program
*prog
, const char *fmt
, ...) PRINTFLIKE(2, 3);
479 fail_link(struct gl_shader_program
*prog
, const char *fmt
, ...)
483 ralloc_vasprintf_append(&prog
->InfoLog
, fmt
, args
);
486 prog
->LinkStatus
= GL_FALSE
;
490 swizzle_for_size(int size
)
492 int size_swizzles
[4] = {
493 MAKE_SWIZZLE4(SWIZZLE_X
, SWIZZLE_X
, SWIZZLE_X
, SWIZZLE_X
),
494 MAKE_SWIZZLE4(SWIZZLE_X
, SWIZZLE_Y
, SWIZZLE_Y
, SWIZZLE_Y
),
495 MAKE_SWIZZLE4(SWIZZLE_X
, SWIZZLE_Y
, SWIZZLE_Z
, SWIZZLE_Z
),
496 MAKE_SWIZZLE4(SWIZZLE_X
, SWIZZLE_Y
, SWIZZLE_Z
, SWIZZLE_W
),
499 assert((size
>= 1) && (size
<= 4));
500 return size_swizzles
[size
- 1];
504 is_tex_instruction(unsigned opcode
)
506 const tgsi_opcode_info
* info
= tgsi_get_opcode_info(opcode
);
511 num_inst_dst_regs(unsigned opcode
)
513 const tgsi_opcode_info
* info
= tgsi_get_opcode_info(opcode
);
514 return info
->num_dst
;
518 num_inst_src_regs(unsigned opcode
)
520 const tgsi_opcode_info
* info
= tgsi_get_opcode_info(opcode
);
521 return info
->is_tex
? info
->num_src
- 1 : info
->num_src
;
524 glsl_to_tgsi_instruction
*
525 glsl_to_tgsi_visitor::emit(ir_instruction
*ir
, unsigned op
,
527 st_src_reg src0
, st_src_reg src1
, st_src_reg src2
)
529 glsl_to_tgsi_instruction
*inst
= new(mem_ctx
) glsl_to_tgsi_instruction();
530 int num_reladdr
= 0, i
;
532 op
= get_opcode(ir
, op
, dst
, src0
, src1
);
534 /* If we have to do relative addressing, we want to load the ARL
535 * reg directly for one of the regs, and preload the other reladdr
536 * sources into temps.
538 num_reladdr
+= dst
.reladdr
!= NULL
;
539 num_reladdr
+= src0
.reladdr
!= NULL
|| src0
.reladdr2
!= NULL
;
540 num_reladdr
+= src1
.reladdr
!= NULL
|| src1
.reladdr2
!= NULL
;
541 num_reladdr
+= src2
.reladdr
!= NULL
|| src2
.reladdr2
!= NULL
;
543 reladdr_to_temp(ir
, &src2
, &num_reladdr
);
544 reladdr_to_temp(ir
, &src1
, &num_reladdr
);
545 reladdr_to_temp(ir
, &src0
, &num_reladdr
);
548 emit_arl(ir
, address_reg
, *dst
.reladdr
);
551 assert(num_reladdr
== 0);
561 inst
->function
= NULL
;
563 /* Update indirect addressing status used by TGSI */
566 case PROGRAM_STATE_VAR
:
567 case PROGRAM_CONSTANT
:
568 case PROGRAM_UNIFORM
:
569 this->indirect_addr_consts
= true;
571 case PROGRAM_IMMEDIATE
:
572 assert(!"immediates should not have indirect addressing");
579 for (i
=0; i
<3; i
++) {
580 if(inst
->src
[i
].reladdr
) {
581 switch(inst
->src
[i
].file
) {
582 case PROGRAM_STATE_VAR
:
583 case PROGRAM_CONSTANT
:
584 case PROGRAM_UNIFORM
:
585 this->indirect_addr_consts
= true;
587 case PROGRAM_IMMEDIATE
:
588 assert(!"immediates should not have indirect addressing");
597 this->instructions
.push_tail(inst
);
603 glsl_to_tgsi_instruction
*
604 glsl_to_tgsi_visitor::emit(ir_instruction
*ir
, unsigned op
,
605 st_dst_reg dst
, st_src_reg src0
, st_src_reg src1
)
607 return emit(ir
, op
, dst
, src0
, src1
, undef_src
);
610 glsl_to_tgsi_instruction
*
611 glsl_to_tgsi_visitor::emit(ir_instruction
*ir
, unsigned op
,
612 st_dst_reg dst
, st_src_reg src0
)
614 assert(dst
.writemask
!= 0);
615 return emit(ir
, op
, dst
, src0
, undef_src
, undef_src
);
618 glsl_to_tgsi_instruction
*
619 glsl_to_tgsi_visitor::emit(ir_instruction
*ir
, unsigned op
)
621 return emit(ir
, op
, undef_dst
, undef_src
, undef_src
, undef_src
);
625 * Determines whether to use an integer, unsigned integer, or float opcode
626 * based on the operands and input opcode, then emits the result.
629 glsl_to_tgsi_visitor::get_opcode(ir_instruction
*ir
, unsigned op
,
631 st_src_reg src0
, st_src_reg src1
)
633 int type
= GLSL_TYPE_FLOAT
;
635 assert(src0
.type
!= GLSL_TYPE_ARRAY
);
636 assert(src0
.type
!= GLSL_TYPE_STRUCT
);
637 assert(src1
.type
!= GLSL_TYPE_ARRAY
);
638 assert(src1
.type
!= GLSL_TYPE_STRUCT
);
640 if (src0
.type
== GLSL_TYPE_FLOAT
|| src1
.type
== GLSL_TYPE_FLOAT
)
641 type
= GLSL_TYPE_FLOAT
;
642 else if (native_integers
)
643 type
= src0
.type
== GLSL_TYPE_BOOL
? GLSL_TYPE_INT
: src0
.type
;
645 #define case4(c, f, i, u) \
646 case TGSI_OPCODE_##c: \
647 if (type == GLSL_TYPE_INT) \
648 op = TGSI_OPCODE_##i; \
649 else if (type == GLSL_TYPE_UINT) \
650 op = TGSI_OPCODE_##u; \
652 op = TGSI_OPCODE_##f; \
655 #define case3(f, i, u) case4(f, f, i, u)
656 #define case2fi(f, i) case4(f, f, i, i)
657 #define case2iu(i, u) case4(i, LAST, i, u)
659 #define casecomp(c, f, i, u) \
660 case TGSI_OPCODE_##c: \
661 if (type == GLSL_TYPE_INT) \
662 op = TGSI_OPCODE_##i; \
663 else if (type == GLSL_TYPE_UINT) \
664 op = TGSI_OPCODE_##u; \
665 else if (native_integers) \
666 op = TGSI_OPCODE_##f; \
668 op = TGSI_OPCODE_##c; \
675 case3(DIV
, IDIV
, UDIV
);
676 case3(MAX
, IMAX
, UMAX
);
677 case3(MIN
, IMIN
, UMIN
);
680 casecomp(SEQ
, FSEQ
, USEQ
, USEQ
);
681 casecomp(SNE
, FSNE
, USNE
, USNE
);
682 casecomp(SGE
, FSGE
, ISGE
, USGE
);
683 casecomp(SLT
, FSLT
, ISLT
, USLT
);
688 case3(ABS
, IABS
, IABS
);
693 assert(op
!= TGSI_OPCODE_LAST
);
697 glsl_to_tgsi_instruction
*
698 glsl_to_tgsi_visitor::emit_dp(ir_instruction
*ir
,
699 st_dst_reg dst
, st_src_reg src0
, st_src_reg src1
,
702 static const unsigned dot_opcodes
[] = {
703 TGSI_OPCODE_DP2
, TGSI_OPCODE_DP3
, TGSI_OPCODE_DP4
706 return emit(ir
, dot_opcodes
[elements
- 2], dst
, src0
, src1
);
710 * Emits TGSI scalar opcodes to produce unique answers across channels.
712 * Some TGSI opcodes are scalar-only, like ARB_fp/vp. The src X
713 * channel determines the result across all channels. So to do a vec4
714 * of this operation, we want to emit a scalar per source channel used
715 * to produce dest channels.
718 glsl_to_tgsi_visitor::emit_scalar(ir_instruction
*ir
, unsigned op
,
720 st_src_reg orig_src0
, st_src_reg orig_src1
)
723 int done_mask
= ~dst
.writemask
;
725 /* TGSI RCP is a scalar operation splatting results to all channels,
726 * like ARB_fp/vp. So emit as many RCPs as necessary to cover our
729 for (i
= 0; i
< 4; i
++) {
730 GLuint this_mask
= (1 << i
);
731 glsl_to_tgsi_instruction
*inst
;
732 st_src_reg src0
= orig_src0
;
733 st_src_reg src1
= orig_src1
;
735 if (done_mask
& this_mask
)
738 GLuint src0_swiz
= GET_SWZ(src0
.swizzle
, i
);
739 GLuint src1_swiz
= GET_SWZ(src1
.swizzle
, i
);
740 for (j
= i
+ 1; j
< 4; j
++) {
741 /* If there is another enabled component in the destination that is
742 * derived from the same inputs, generate its value on this pass as
745 if (!(done_mask
& (1 << j
)) &&
746 GET_SWZ(src0
.swizzle
, j
) == src0_swiz
&&
747 GET_SWZ(src1
.swizzle
, j
) == src1_swiz
) {
748 this_mask
|= (1 << j
);
751 src0
.swizzle
= MAKE_SWIZZLE4(src0_swiz
, src0_swiz
,
752 src0_swiz
, src0_swiz
);
753 src1
.swizzle
= MAKE_SWIZZLE4(src1_swiz
, src1_swiz
,
754 src1_swiz
, src1_swiz
);
756 inst
= emit(ir
, op
, dst
, src0
, src1
);
757 inst
->dst
.writemask
= this_mask
;
758 done_mask
|= this_mask
;
763 glsl_to_tgsi_visitor::emit_scalar(ir_instruction
*ir
, unsigned op
,
764 st_dst_reg dst
, st_src_reg src0
)
766 st_src_reg undef
= undef_src
;
768 undef
.swizzle
= SWIZZLE_XXXX
;
770 emit_scalar(ir
, op
, dst
, src0
, undef
);
774 glsl_to_tgsi_visitor::emit_arl(ir_instruction
*ir
,
775 st_dst_reg dst
, st_src_reg src0
)
777 int op
= TGSI_OPCODE_ARL
;
779 if (src0
.type
== GLSL_TYPE_INT
|| src0
.type
== GLSL_TYPE_UINT
)
780 op
= TGSI_OPCODE_UARL
;
782 assert(dst
.file
== PROGRAM_ADDRESS
);
783 if (dst
.index
>= this->num_address_regs
)
784 this->num_address_regs
= dst
.index
+ 1;
786 emit(NULL
, op
, dst
, src0
);
790 * Emit an TGSI_OPCODE_SCS instruction
792 * The \c SCS opcode functions a bit differently than the other TGSI opcodes.
793 * Instead of splatting its result across all four components of the
794 * destination, it writes one value to the \c x component and another value to
795 * the \c y component.
797 * \param ir IR instruction being processed
798 * \param op Either \c TGSI_OPCODE_SIN or \c TGSI_OPCODE_COS depending
799 * on which value is desired.
800 * \param dst Destination register
801 * \param src Source register
804 glsl_to_tgsi_visitor::emit_scs(ir_instruction
*ir
, unsigned op
,
806 const st_src_reg
&src
)
808 /* Vertex programs cannot use the SCS opcode.
810 if (this->prog
->Target
== GL_VERTEX_PROGRAM_ARB
) {
811 emit_scalar(ir
, op
, dst
, src
);
815 const unsigned component
= (op
== TGSI_OPCODE_SIN
) ? 0 : 1;
816 const unsigned scs_mask
= (1U << component
);
817 int done_mask
= ~dst
.writemask
;
820 assert(op
== TGSI_OPCODE_SIN
|| op
== TGSI_OPCODE_COS
);
822 /* If there are compnents in the destination that differ from the component
823 * that will be written by the SCS instrution, we'll need a temporary.
825 if (scs_mask
!= unsigned(dst
.writemask
)) {
826 tmp
= get_temp(glsl_type::vec4_type
);
829 for (unsigned i
= 0; i
< 4; i
++) {
830 unsigned this_mask
= (1U << i
);
831 st_src_reg src0
= src
;
833 if ((done_mask
& this_mask
) != 0)
836 /* The source swizzle specified which component of the source generates
837 * sine / cosine for the current component in the destination. The SCS
838 * instruction requires that this value be swizzle to the X component.
839 * Replace the current swizzle with a swizzle that puts the source in
842 unsigned src0_swiz
= GET_SWZ(src
.swizzle
, i
);
844 src0
.swizzle
= MAKE_SWIZZLE4(src0_swiz
, src0_swiz
,
845 src0_swiz
, src0_swiz
);
846 for (unsigned j
= i
+ 1; j
< 4; j
++) {
847 /* If there is another enabled component in the destination that is
848 * derived from the same inputs, generate its value on this pass as
851 if (!(done_mask
& (1 << j
)) &&
852 GET_SWZ(src0
.swizzle
, j
) == src0_swiz
) {
853 this_mask
|= (1 << j
);
857 if (this_mask
!= scs_mask
) {
858 glsl_to_tgsi_instruction
*inst
;
859 st_dst_reg tmp_dst
= st_dst_reg(tmp
);
861 /* Emit the SCS instruction.
863 inst
= emit(ir
, TGSI_OPCODE_SCS
, tmp_dst
, src0
);
864 inst
->dst
.writemask
= scs_mask
;
866 /* Move the result of the SCS instruction to the desired location in
869 tmp
.swizzle
= MAKE_SWIZZLE4(component
, component
,
870 component
, component
);
871 inst
= emit(ir
, TGSI_OPCODE_SCS
, dst
, tmp
);
872 inst
->dst
.writemask
= this_mask
;
874 /* Emit the SCS instruction to write directly to the destination.
876 glsl_to_tgsi_instruction
*inst
= emit(ir
, TGSI_OPCODE_SCS
, dst
, src0
);
877 inst
->dst
.writemask
= scs_mask
;
880 done_mask
|= this_mask
;
885 glsl_to_tgsi_visitor::add_constant(gl_register_file file
,
886 gl_constant_value values
[4], int size
, int datatype
,
889 if (file
== PROGRAM_CONSTANT
) {
890 return _mesa_add_typed_unnamed_constant(this->prog
->Parameters
, values
,
891 size
, datatype
, swizzle_out
);
894 immediate_storage
*entry
;
895 assert(file
== PROGRAM_IMMEDIATE
);
897 /* Search immediate storage to see if we already have an identical
898 * immediate that we can use instead of adding a duplicate entry.
900 foreach_iter(exec_list_iterator
, iter
, this->immediates
) {
901 entry
= (immediate_storage
*)iter
.get();
903 if (entry
->size
== size
&&
904 entry
->type
== datatype
&&
905 !memcmp(entry
->values
, values
, size
* sizeof(gl_constant_value
))) {
911 /* Add this immediate to the list. */
912 entry
= new(mem_ctx
) immediate_storage(values
, size
, datatype
);
913 this->immediates
.push_tail(entry
);
914 this->num_immediates
++;
920 glsl_to_tgsi_visitor::st_src_reg_for_float(float val
)
922 st_src_reg
src(PROGRAM_IMMEDIATE
, -1, GLSL_TYPE_FLOAT
);
923 union gl_constant_value uval
;
926 src
.index
= add_constant(src
.file
, &uval
, 1, GL_FLOAT
, &src
.swizzle
);
932 glsl_to_tgsi_visitor::st_src_reg_for_int(int val
)
934 st_src_reg
src(PROGRAM_IMMEDIATE
, -1, GLSL_TYPE_INT
);
935 union gl_constant_value uval
;
937 assert(native_integers
);
940 src
.index
= add_constant(src
.file
, &uval
, 1, GL_INT
, &src
.swizzle
);
946 glsl_to_tgsi_visitor::st_src_reg_for_type(int type
, int val
)
949 return type
== GLSL_TYPE_FLOAT
? st_src_reg_for_float(val
) :
950 st_src_reg_for_int(val
);
952 return st_src_reg_for_float(val
);
956 type_size(const struct glsl_type
*type
)
961 switch (type
->base_type
) {
964 case GLSL_TYPE_FLOAT
:
966 if (type
->is_matrix()) {
967 return type
->matrix_columns
;
969 /* Regardless of size of vector, it gets a vec4. This is bad
970 * packing for things like floats, but otherwise arrays become a
971 * mess. Hopefully a later pass over the code can pack scalars
972 * down if appropriate.
976 case GLSL_TYPE_ARRAY
:
977 assert(type
->length
> 0);
978 return type_size(type
->fields
.array
) * type
->length
;
979 case GLSL_TYPE_STRUCT
:
981 for (i
= 0; i
< type
->length
; i
++) {
982 size
+= type_size(type
->fields
.structure
[i
].type
);
985 case GLSL_TYPE_SAMPLER
:
986 /* Samplers take up one slot in UNIFORMS[], but they're baked in
990 case GLSL_TYPE_ATOMIC_UINT
:
991 case GLSL_TYPE_INTERFACE
:
993 case GLSL_TYPE_ERROR
:
994 assert(!"Invalid type in type_size");
1001 * In the initial pass of codegen, we assign temporary numbers to
1002 * intermediate results. (not SSA -- variable assignments will reuse
1006 glsl_to_tgsi_visitor::get_temp(const glsl_type
*type
)
1010 src
.type
= native_integers
? type
->base_type
: GLSL_TYPE_FLOAT
;
1014 if (!options
->EmitNoIndirectTemp
&&
1015 (type
->is_array() || type
->is_matrix())) {
1017 src
.file
= PROGRAM_ARRAY
;
1018 src
.index
= next_array
<< 16 | 0x8000;
1019 array_sizes
[next_array
] = type_size(type
);
1023 src
.file
= PROGRAM_TEMPORARY
;
1024 src
.index
= next_temp
;
1025 next_temp
+= type_size(type
);
1028 if (type
->is_array() || type
->is_record()) {
1029 src
.swizzle
= SWIZZLE_NOOP
;
1031 src
.swizzle
= swizzle_for_size(type
->vector_elements
);
1038 glsl_to_tgsi_visitor::find_variable_storage(ir_variable
*var
)
1041 variable_storage
*entry
;
1043 foreach_iter(exec_list_iterator
, iter
, this->variables
) {
1044 entry
= (variable_storage
*)iter
.get();
1046 if (entry
->var
== var
)
1054 glsl_to_tgsi_visitor::visit(ir_variable
*ir
)
1056 if (strcmp(ir
->name
, "gl_FragCoord") == 0) {
1057 struct gl_fragment_program
*fp
= (struct gl_fragment_program
*)this->prog
;
1059 fp
->OriginUpperLeft
= ir
->origin_upper_left
;
1060 fp
->PixelCenterInteger
= ir
->pixel_center_integer
;
1063 if (ir
->mode
== ir_var_uniform
&& strncmp(ir
->name
, "gl_", 3) == 0) {
1065 const ir_state_slot
*const slots
= ir
->state_slots
;
1066 assert(ir
->state_slots
!= NULL
);
1068 /* Check if this statevar's setup in the STATE file exactly
1069 * matches how we'll want to reference it as a
1070 * struct/array/whatever. If not, then we need to move it into
1071 * temporary storage and hope that it'll get copy-propagated
1074 for (i
= 0; i
< ir
->num_state_slots
; i
++) {
1075 if (slots
[i
].swizzle
!= SWIZZLE_XYZW
) {
1080 variable_storage
*storage
;
1082 if (i
== ir
->num_state_slots
) {
1083 /* We'll set the index later. */
1084 storage
= new(mem_ctx
) variable_storage(ir
, PROGRAM_STATE_VAR
, -1);
1085 this->variables
.push_tail(storage
);
1089 /* The variable_storage constructor allocates slots based on the size
1090 * of the type. However, this had better match the number of state
1091 * elements that we're going to copy into the new temporary.
1093 assert((int) ir
->num_state_slots
== type_size(ir
->type
));
1095 dst
= st_dst_reg(get_temp(ir
->type
));
1097 storage
= new(mem_ctx
) variable_storage(ir
, dst
.file
, dst
.index
);
1099 this->variables
.push_tail(storage
);
1103 for (unsigned int i
= 0; i
< ir
->num_state_slots
; i
++) {
1104 int index
= _mesa_add_state_reference(this->prog
->Parameters
,
1105 (gl_state_index
*)slots
[i
].tokens
);
1107 if (storage
->file
== PROGRAM_STATE_VAR
) {
1108 if (storage
->index
== -1) {
1109 storage
->index
= index
;
1111 assert(index
== storage
->index
+ (int)i
);
1114 /* We use GLSL_TYPE_FLOAT here regardless of the actual type of
1115 * the data being moved since MOV does not care about the type of
1116 * data it is moving, and we don't want to declare registers with
1117 * array or struct types.
1119 st_src_reg
src(PROGRAM_STATE_VAR
, index
, GLSL_TYPE_FLOAT
);
1120 src
.swizzle
= slots
[i
].swizzle
;
1121 emit(ir
, TGSI_OPCODE_MOV
, dst
, src
);
1122 /* even a float takes up a whole vec4 reg in a struct/array. */
1127 if (storage
->file
== PROGRAM_TEMPORARY
&&
1128 dst
.index
!= storage
->index
+ (int) ir
->num_state_slots
) {
1129 fail_link(this->shader_program
,
1130 "failed to load builtin uniform `%s' (%d/%d regs loaded)\n",
1131 ir
->name
, dst
.index
- storage
->index
,
1132 type_size(ir
->type
));
1138 glsl_to_tgsi_visitor::visit(ir_loop
*ir
)
1140 emit(NULL
, TGSI_OPCODE_BGNLOOP
);
1142 visit_exec_list(&ir
->body_instructions
, this);
1144 emit(NULL
, TGSI_OPCODE_ENDLOOP
);
1148 glsl_to_tgsi_visitor::visit(ir_loop_jump
*ir
)
1151 case ir_loop_jump::jump_break
:
1152 emit(NULL
, TGSI_OPCODE_BRK
);
1154 case ir_loop_jump::jump_continue
:
1155 emit(NULL
, TGSI_OPCODE_CONT
);
1162 glsl_to_tgsi_visitor::visit(ir_function_signature
*ir
)
1169 glsl_to_tgsi_visitor::visit(ir_function
*ir
)
1171 /* Ignore function bodies other than main() -- we shouldn't see calls to
1172 * them since they should all be inlined before we get to glsl_to_tgsi.
1174 if (strcmp(ir
->name
, "main") == 0) {
1175 const ir_function_signature
*sig
;
1178 sig
= ir
->matching_signature(NULL
, &empty
);
1182 foreach_iter(exec_list_iterator
, iter
, sig
->body
) {
1183 ir_instruction
*ir
= (ir_instruction
*)iter
.get();
1191 glsl_to_tgsi_visitor::try_emit_mad(ir_expression
*ir
, int mul_operand
)
1193 int nonmul_operand
= 1 - mul_operand
;
1195 st_dst_reg result_dst
;
1197 ir_expression
*expr
= ir
->operands
[mul_operand
]->as_expression();
1198 if (!expr
|| expr
->operation
!= ir_binop_mul
)
1201 expr
->operands
[0]->accept(this);
1203 expr
->operands
[1]->accept(this);
1205 ir
->operands
[nonmul_operand
]->accept(this);
1208 this->result
= get_temp(ir
->type
);
1209 result_dst
= st_dst_reg(this->result
);
1210 result_dst
.writemask
= (1 << ir
->type
->vector_elements
) - 1;
1211 emit(ir
, TGSI_OPCODE_MAD
, result_dst
, a
, b
, c
);
1217 * Emit MAD(a, -b, a) instead of AND(a, NOT(b))
1219 * The logic values are 1.0 for true and 0.0 for false. Logical-and is
1220 * implemented using multiplication, and logical-or is implemented using
1221 * addition. Logical-not can be implemented as (true - x), or (1.0 - x).
1222 * As result, the logical expression (a & !b) can be rewritten as:
1226 * - (a * 1) - (a * b)
1230 * This final expression can be implemented as a single MAD(a, -b, a)
1234 glsl_to_tgsi_visitor::try_emit_mad_for_and_not(ir_expression
*ir
, int try_operand
)
1236 const int other_operand
= 1 - try_operand
;
1239 ir_expression
*expr
= ir
->operands
[try_operand
]->as_expression();
1240 if (!expr
|| expr
->operation
!= ir_unop_logic_not
)
1243 ir
->operands
[other_operand
]->accept(this);
1245 expr
->operands
[0]->accept(this);
1248 b
.negate
= ~b
.negate
;
1250 this->result
= get_temp(ir
->type
);
1251 emit(ir
, TGSI_OPCODE_MAD
, st_dst_reg(this->result
), a
, b
, a
);
1257 glsl_to_tgsi_visitor::try_emit_sat(ir_expression
*ir
)
1259 /* Emit saturates in the vertex shader only if SM 3.0 is supported.
1261 if (this->prog
->Target
== GL_VERTEX_PROGRAM_ARB
&&
1262 !st_context(this->ctx
)->has_shader_model3
) {
1266 ir_rvalue
*sat_src
= ir
->as_rvalue_to_saturate();
1270 sat_src
->accept(this);
1271 st_src_reg src
= this->result
;
1273 /* If we generated an expression instruction into a temporary in
1274 * processing the saturate's operand, apply the saturate to that
1275 * instruction. Otherwise, generate a MOV to do the saturate.
1277 * Note that we have to be careful to only do this optimization if
1278 * the instruction in question was what generated src->result. For
1279 * example, ir_dereference_array might generate a MUL instruction
1280 * to create the reladdr, and return us a src reg using that
1281 * reladdr. That MUL result is not the value we're trying to
1284 ir_expression
*sat_src_expr
= sat_src
->as_expression();
1285 if (sat_src_expr
&& (sat_src_expr
->operation
== ir_binop_mul
||
1286 sat_src_expr
->operation
== ir_binop_add
||
1287 sat_src_expr
->operation
== ir_binop_dot
)) {
1288 glsl_to_tgsi_instruction
*new_inst
;
1289 new_inst
= (glsl_to_tgsi_instruction
*)this->instructions
.get_tail();
1290 new_inst
->saturate
= true;
1292 this->result
= get_temp(ir
->type
);
1293 st_dst_reg result_dst
= st_dst_reg(this->result
);
1294 result_dst
.writemask
= (1 << ir
->type
->vector_elements
) - 1;
1295 glsl_to_tgsi_instruction
*inst
;
1296 inst
= emit(ir
, TGSI_OPCODE_MOV
, result_dst
, src
);
1297 inst
->saturate
= true;
1304 glsl_to_tgsi_visitor::reladdr_to_temp(ir_instruction
*ir
,
1305 st_src_reg
*reg
, int *num_reladdr
)
1307 if (!reg
->reladdr
&& !reg
->reladdr2
)
1310 if (reg
->reladdr
) emit_arl(ir
, address_reg
, *reg
->reladdr
);
1311 if (reg
->reladdr2
) emit_arl(ir
, address_reg2
, *reg
->reladdr2
);
1313 if (*num_reladdr
!= 1) {
1314 st_src_reg temp
= get_temp(glsl_type::vec4_type
);
1316 emit(ir
, TGSI_OPCODE_MOV
, st_dst_reg(temp
), *reg
);
1324 glsl_to_tgsi_visitor::visit(ir_expression
*ir
)
1326 unsigned int operand
;
1327 st_src_reg op
[Elements(ir
->operands
)];
1328 st_src_reg result_src
;
1329 st_dst_reg result_dst
;
1331 /* Quick peephole: Emit MAD(a, b, c) instead of ADD(MUL(a, b), c)
1333 if (ir
->operation
== ir_binop_add
) {
1334 if (try_emit_mad(ir
, 1))
1336 if (try_emit_mad(ir
, 0))
1340 /* Quick peephole: Emit OPCODE_MAD(-a, -b, a) instead of AND(a, NOT(b))
1342 if (ir
->operation
== ir_binop_logic_and
) {
1343 if (try_emit_mad_for_and_not(ir
, 1))
1345 if (try_emit_mad_for_and_not(ir
, 0))
1349 if (try_emit_sat(ir
))
1352 if (ir
->operation
== ir_quadop_vector
)
1353 assert(!"ir_quadop_vector should have been lowered");
1355 for (operand
= 0; operand
< ir
->get_num_operands(); operand
++) {
1356 this->result
.file
= PROGRAM_UNDEFINED
;
1357 ir
->operands
[operand
]->accept(this);
1358 if (this->result
.file
== PROGRAM_UNDEFINED
) {
1359 printf("Failed to get tree for expression operand:\n");
1360 ir
->operands
[operand
]->print();
1364 op
[operand
] = this->result
;
1366 /* Matrix expression operands should have been broken down to vector
1367 * operations already.
1369 assert(!ir
->operands
[operand
]->type
->is_matrix());
1372 int vector_elements
= ir
->operands
[0]->type
->vector_elements
;
1373 if (ir
->operands
[1]) {
1374 vector_elements
= MAX2(vector_elements
,
1375 ir
->operands
[1]->type
->vector_elements
);
1378 this->result
.file
= PROGRAM_UNDEFINED
;
1380 /* Storage for our result. Ideally for an assignment we'd be using
1381 * the actual storage for the result here, instead.
1383 result_src
= get_temp(ir
->type
);
1384 /* convenience for the emit functions below. */
1385 result_dst
= st_dst_reg(result_src
);
1386 /* Limit writes to the channels that will be used by result_src later.
1387 * This does limit this temp's use as a temporary for multi-instruction
1390 result_dst
.writemask
= (1 << ir
->type
->vector_elements
) - 1;
1392 switch (ir
->operation
) {
1393 case ir_unop_logic_not
:
1394 if (result_dst
.type
!= GLSL_TYPE_FLOAT
)
1395 emit(ir
, TGSI_OPCODE_NOT
, result_dst
, op
[0]);
1397 /* Previously 'SEQ dst, src, 0.0' was used for this. However, many
1398 * older GPUs implement SEQ using multiple instructions (i915 uses two
1399 * SGE instructions and a MUL instruction). Since our logic values are
1400 * 0.0 and 1.0, 1-x also implements !x.
1402 op
[0].negate
= ~op
[0].negate
;
1403 emit(ir
, TGSI_OPCODE_ADD
, result_dst
, op
[0], st_src_reg_for_float(1.0));
1407 if (result_dst
.type
== GLSL_TYPE_INT
|| result_dst
.type
== GLSL_TYPE_UINT
)
1408 emit(ir
, TGSI_OPCODE_INEG
, result_dst
, op
[0]);
1410 op
[0].negate
= ~op
[0].negate
;
1415 emit(ir
, TGSI_OPCODE_ABS
, result_dst
, op
[0]);
1418 emit(ir
, TGSI_OPCODE_SSG
, result_dst
, op
[0]);
1421 emit_scalar(ir
, TGSI_OPCODE_RCP
, result_dst
, op
[0]);
1425 emit_scalar(ir
, TGSI_OPCODE_EX2
, result_dst
, op
[0]);
1429 assert(!"not reached: should be handled by ir_explog_to_explog2");
1432 emit_scalar(ir
, TGSI_OPCODE_LG2
, result_dst
, op
[0]);
1435 emit_scalar(ir
, TGSI_OPCODE_SIN
, result_dst
, op
[0]);
1438 emit_scalar(ir
, TGSI_OPCODE_COS
, result_dst
, op
[0]);
1440 case ir_unop_sin_reduced
:
1441 emit_scs(ir
, TGSI_OPCODE_SIN
, result_dst
, op
[0]);
1443 case ir_unop_cos_reduced
:
1444 emit_scs(ir
, TGSI_OPCODE_COS
, result_dst
, op
[0]);
1448 emit(ir
, TGSI_OPCODE_DDX
, result_dst
, op
[0]);
1452 /* The X component contains 1 or -1 depending on whether the framebuffer
1453 * is a FBO or the window system buffer, respectively.
1454 * It is then multiplied with the source operand of DDY.
1456 static const gl_state_index transform_y_state
[STATE_LENGTH
]
1457 = { STATE_INTERNAL
, STATE_FB_WPOS_Y_TRANSFORM
};
1459 unsigned transform_y_index
=
1460 _mesa_add_state_reference(this->prog
->Parameters
,
1463 st_src_reg transform_y
= st_src_reg(PROGRAM_STATE_VAR
,
1465 glsl_type::vec4_type
);
1466 transform_y
.swizzle
= SWIZZLE_XXXX
;
1468 st_src_reg temp
= get_temp(glsl_type::vec4_type
);
1470 emit(ir
, TGSI_OPCODE_MUL
, st_dst_reg(temp
), transform_y
, op
[0]);
1471 emit(ir
, TGSI_OPCODE_DDY
, result_dst
, temp
);
1475 case ir_unop_noise
: {
1476 /* At some point, a motivated person could add a better
1477 * implementation of noise. Currently not even the nvidia
1478 * binary drivers do anything more than this. In any case, the
1479 * place to do this is in the GL state tracker, not the poor
1482 emit(ir
, TGSI_OPCODE_MOV
, result_dst
, st_src_reg_for_float(0.5));
1487 emit(ir
, TGSI_OPCODE_ADD
, result_dst
, op
[0], op
[1]);
1490 emit(ir
, TGSI_OPCODE_SUB
, result_dst
, op
[0], op
[1]);
1494 emit(ir
, TGSI_OPCODE_MUL
, result_dst
, op
[0], op
[1]);
1497 if (result_dst
.type
== GLSL_TYPE_FLOAT
)
1498 assert(!"not reached: should be handled by ir_div_to_mul_rcp");
1500 emit(ir
, TGSI_OPCODE_DIV
, result_dst
, op
[0], op
[1]);
1503 if (result_dst
.type
== GLSL_TYPE_FLOAT
)
1504 assert(!"ir_binop_mod should have been converted to b * fract(a/b)");
1506 emit(ir
, TGSI_OPCODE_MOD
, result_dst
, op
[0], op
[1]);
1510 emit(ir
, TGSI_OPCODE_SLT
, result_dst
, op
[0], op
[1]);
1512 case ir_binop_greater
:
1513 emit(ir
, TGSI_OPCODE_SLT
, result_dst
, op
[1], op
[0]);
1515 case ir_binop_lequal
:
1516 emit(ir
, TGSI_OPCODE_SGE
, result_dst
, op
[1], op
[0]);
1518 case ir_binop_gequal
:
1519 emit(ir
, TGSI_OPCODE_SGE
, result_dst
, op
[0], op
[1]);
1521 case ir_binop_equal
:
1522 emit(ir
, TGSI_OPCODE_SEQ
, result_dst
, op
[0], op
[1]);
1524 case ir_binop_nequal
:
1525 emit(ir
, TGSI_OPCODE_SNE
, result_dst
, op
[0], op
[1]);
1527 case ir_binop_all_equal
:
1528 /* "==" operator producing a scalar boolean. */
1529 if (ir
->operands
[0]->type
->is_vector() ||
1530 ir
->operands
[1]->type
->is_vector()) {
1531 st_src_reg temp
= get_temp(native_integers
?
1532 glsl_type::get_instance(ir
->operands
[0]->type
->base_type
, 4, 1) :
1533 glsl_type::vec4_type
);
1535 if (native_integers
) {
1536 st_dst_reg temp_dst
= st_dst_reg(temp
);
1537 st_src_reg temp1
= st_src_reg(temp
), temp2
= st_src_reg(temp
);
1539 emit(ir
, TGSI_OPCODE_SEQ
, st_dst_reg(temp
), op
[0], op
[1]);
1541 /* Emit 1-3 AND operations to combine the SEQ results. */
1542 switch (ir
->operands
[0]->type
->vector_elements
) {
1546 temp_dst
.writemask
= WRITEMASK_Y
;
1547 temp1
.swizzle
= SWIZZLE_YYYY
;
1548 temp2
.swizzle
= SWIZZLE_ZZZZ
;
1549 emit(ir
, TGSI_OPCODE_AND
, temp_dst
, temp1
, temp2
);
1552 temp_dst
.writemask
= WRITEMASK_X
;
1553 temp1
.swizzle
= SWIZZLE_XXXX
;
1554 temp2
.swizzle
= SWIZZLE_YYYY
;
1555 emit(ir
, TGSI_OPCODE_AND
, temp_dst
, temp1
, temp2
);
1556 temp_dst
.writemask
= WRITEMASK_Y
;
1557 temp1
.swizzle
= SWIZZLE_ZZZZ
;
1558 temp2
.swizzle
= SWIZZLE_WWWW
;
1559 emit(ir
, TGSI_OPCODE_AND
, temp_dst
, temp1
, temp2
);
1562 temp1
.swizzle
= SWIZZLE_XXXX
;
1563 temp2
.swizzle
= SWIZZLE_YYYY
;
1564 emit(ir
, TGSI_OPCODE_AND
, result_dst
, temp1
, temp2
);
1566 emit(ir
, TGSI_OPCODE_SNE
, st_dst_reg(temp
), op
[0], op
[1]);
1568 /* After the dot-product, the value will be an integer on the
1569 * range [0,4]. Zero becomes 1.0, and positive values become zero.
1571 emit_dp(ir
, result_dst
, temp
, temp
, vector_elements
);
1573 /* Negating the result of the dot-product gives values on the range
1574 * [-4, 0]. Zero becomes 1.0, and negative values become zero.
1575 * This is achieved using SGE.
1577 st_src_reg sge_src
= result_src
;
1578 sge_src
.negate
= ~sge_src
.negate
;
1579 emit(ir
, TGSI_OPCODE_SGE
, result_dst
, sge_src
, st_src_reg_for_float(0.0));
1582 emit(ir
, TGSI_OPCODE_SEQ
, result_dst
, op
[0], op
[1]);
1585 case ir_binop_any_nequal
:
1586 /* "!=" operator producing a scalar boolean. */
1587 if (ir
->operands
[0]->type
->is_vector() ||
1588 ir
->operands
[1]->type
->is_vector()) {
1589 st_src_reg temp
= get_temp(native_integers
?
1590 glsl_type::get_instance(ir
->operands
[0]->type
->base_type
, 4, 1) :
1591 glsl_type::vec4_type
);
1592 emit(ir
, TGSI_OPCODE_SNE
, st_dst_reg(temp
), op
[0], op
[1]);
1594 if (native_integers
) {
1595 st_dst_reg temp_dst
= st_dst_reg(temp
);
1596 st_src_reg temp1
= st_src_reg(temp
), temp2
= st_src_reg(temp
);
1598 /* Emit 1-3 OR operations to combine the SNE results. */
1599 switch (ir
->operands
[0]->type
->vector_elements
) {
1603 temp_dst
.writemask
= WRITEMASK_Y
;
1604 temp1
.swizzle
= SWIZZLE_YYYY
;
1605 temp2
.swizzle
= SWIZZLE_ZZZZ
;
1606 emit(ir
, TGSI_OPCODE_OR
, temp_dst
, temp1
, temp2
);
1609 temp_dst
.writemask
= WRITEMASK_X
;
1610 temp1
.swizzle
= SWIZZLE_XXXX
;
1611 temp2
.swizzle
= SWIZZLE_YYYY
;
1612 emit(ir
, TGSI_OPCODE_OR
, temp_dst
, temp1
, temp2
);
1613 temp_dst
.writemask
= WRITEMASK_Y
;
1614 temp1
.swizzle
= SWIZZLE_ZZZZ
;
1615 temp2
.swizzle
= SWIZZLE_WWWW
;
1616 emit(ir
, TGSI_OPCODE_OR
, temp_dst
, temp1
, temp2
);
1619 temp1
.swizzle
= SWIZZLE_XXXX
;
1620 temp2
.swizzle
= SWIZZLE_YYYY
;
1621 emit(ir
, TGSI_OPCODE_OR
, result_dst
, temp1
, temp2
);
1623 /* After the dot-product, the value will be an integer on the
1624 * range [0,4]. Zero stays zero, and positive values become 1.0.
1626 glsl_to_tgsi_instruction
*const dp
=
1627 emit_dp(ir
, result_dst
, temp
, temp
, vector_elements
);
1628 if (this->prog
->Target
== GL_FRAGMENT_PROGRAM_ARB
) {
1629 /* The clamping to [0,1] can be done for free in the fragment
1630 * shader with a saturate.
1632 dp
->saturate
= true;
1634 /* Negating the result of the dot-product gives values on the range
1635 * [-4, 0]. Zero stays zero, and negative values become 1.0. This
1636 * achieved using SLT.
1638 st_src_reg slt_src
= result_src
;
1639 slt_src
.negate
= ~slt_src
.negate
;
1640 emit(ir
, TGSI_OPCODE_SLT
, result_dst
, slt_src
, st_src_reg_for_float(0.0));
1644 emit(ir
, TGSI_OPCODE_SNE
, result_dst
, op
[0], op
[1]);
1649 assert(ir
->operands
[0]->type
->is_vector());
1651 /* After the dot-product, the value will be an integer on the
1652 * range [0,4]. Zero stays zero, and positive values become 1.0.
1654 glsl_to_tgsi_instruction
*const dp
=
1655 emit_dp(ir
, result_dst
, op
[0], op
[0],
1656 ir
->operands
[0]->type
->vector_elements
);
1657 if (this->prog
->Target
== GL_FRAGMENT_PROGRAM_ARB
&&
1658 result_dst
.type
== GLSL_TYPE_FLOAT
) {
1659 /* The clamping to [0,1] can be done for free in the fragment
1660 * shader with a saturate.
1662 dp
->saturate
= true;
1663 } else if (result_dst
.type
== GLSL_TYPE_FLOAT
) {
1664 /* Negating the result of the dot-product gives values on the range
1665 * [-4, 0]. Zero stays zero, and negative values become 1.0. This
1666 * is achieved using SLT.
1668 st_src_reg slt_src
= result_src
;
1669 slt_src
.negate
= ~slt_src
.negate
;
1670 emit(ir
, TGSI_OPCODE_SLT
, result_dst
, slt_src
, st_src_reg_for_float(0.0));
1673 /* Use SNE 0 if integers are being used as boolean values. */
1674 emit(ir
, TGSI_OPCODE_SNE
, result_dst
, result_src
, st_src_reg_for_int(0));
1679 case ir_binop_logic_xor
:
1680 if (native_integers
)
1681 emit(ir
, TGSI_OPCODE_XOR
, result_dst
, op
[0], op
[1]);
1683 emit(ir
, TGSI_OPCODE_SNE
, result_dst
, op
[0], op
[1]);
1686 case ir_binop_logic_or
: {
1687 if (native_integers
) {
1688 /* If integers are used as booleans, we can use an actual "or"
1691 assert(native_integers
);
1692 emit(ir
, TGSI_OPCODE_OR
, result_dst
, op
[0], op
[1]);
1694 /* After the addition, the value will be an integer on the
1695 * range [0,2]. Zero stays zero, and positive values become 1.0.
1697 glsl_to_tgsi_instruction
*add
=
1698 emit(ir
, TGSI_OPCODE_ADD
, result_dst
, op
[0], op
[1]);
1699 if (this->prog
->Target
== GL_FRAGMENT_PROGRAM_ARB
) {
1700 /* The clamping to [0,1] can be done for free in the fragment
1701 * shader with a saturate if floats are being used as boolean values.
1703 add
->saturate
= true;
1705 /* Negating the result of the addition gives values on the range
1706 * [-2, 0]. Zero stays zero, and negative values become 1.0. This
1707 * is achieved using SLT.
1709 st_src_reg slt_src
= result_src
;
1710 slt_src
.negate
= ~slt_src
.negate
;
1711 emit(ir
, TGSI_OPCODE_SLT
, result_dst
, slt_src
, st_src_reg_for_float(0.0));
1717 case ir_binop_logic_and
:
1718 /* If native integers are disabled, the bool args are stored as float 0.0
1719 * or 1.0, so "mul" gives us "and". If they're enabled, just use the
1720 * actual AND opcode.
1722 if (native_integers
)
1723 emit(ir
, TGSI_OPCODE_AND
, result_dst
, op
[0], op
[1]);
1725 emit(ir
, TGSI_OPCODE_MUL
, result_dst
, op
[0], op
[1]);
1729 assert(ir
->operands
[0]->type
->is_vector());
1730 assert(ir
->operands
[0]->type
== ir
->operands
[1]->type
);
1731 emit_dp(ir
, result_dst
, op
[0], op
[1],
1732 ir
->operands
[0]->type
->vector_elements
);
1737 emit_scalar(ir
, TGSI_OPCODE_SQRT
, result_dst
, op
[0]);
1740 /* sqrt(x) = x * rsq(x). */
1741 emit_scalar(ir
, TGSI_OPCODE_RSQ
, result_dst
, op
[0]);
1742 emit(ir
, TGSI_OPCODE_MUL
, result_dst
, result_src
, op
[0]);
1743 /* For incoming channels <= 0, set the result to 0. */
1744 op
[0].negate
= ~op
[0].negate
;
1745 emit(ir
, TGSI_OPCODE_CMP
, result_dst
,
1746 op
[0], result_src
, st_src_reg_for_float(0.0));
1750 emit_scalar(ir
, TGSI_OPCODE_RSQ
, result_dst
, op
[0]);
1753 if (native_integers
) {
1754 emit(ir
, TGSI_OPCODE_I2F
, result_dst
, op
[0]);
1757 /* fallthrough to next case otherwise */
1759 if (native_integers
) {
1760 emit(ir
, TGSI_OPCODE_AND
, result_dst
, op
[0], st_src_reg_for_float(1.0));
1763 /* fallthrough to next case otherwise */
1766 /* Converting between signed and unsigned integers is a no-op. */
1770 if (native_integers
) {
1771 /* Booleans are stored as integers using ~0 for true and 0 for false.
1772 * GLSL requires that int(bool) return 1 for true and 0 for false.
1773 * This conversion is done with AND, but it could be done with NEG.
1775 emit(ir
, TGSI_OPCODE_AND
, result_dst
, op
[0], st_src_reg_for_int(1));
1777 /* Booleans and integers are both stored as floats when native
1778 * integers are disabled.
1784 if (native_integers
)
1785 emit(ir
, TGSI_OPCODE_F2I
, result_dst
, op
[0]);
1787 emit(ir
, TGSI_OPCODE_TRUNC
, result_dst
, op
[0]);
1790 if (native_integers
)
1791 emit(ir
, TGSI_OPCODE_F2U
, result_dst
, op
[0]);
1793 emit(ir
, TGSI_OPCODE_TRUNC
, result_dst
, op
[0]);
1795 case ir_unop_bitcast_f2i
:
1797 result_src
.type
= GLSL_TYPE_INT
;
1799 case ir_unop_bitcast_f2u
:
1801 result_src
.type
= GLSL_TYPE_UINT
;
1803 case ir_unop_bitcast_i2f
:
1804 case ir_unop_bitcast_u2f
:
1806 result_src
.type
= GLSL_TYPE_FLOAT
;
1809 emit(ir
, TGSI_OPCODE_SNE
, result_dst
, op
[0], st_src_reg_for_float(0.0));
1812 if (native_integers
)
1813 emit(ir
, TGSI_OPCODE_INEG
, result_dst
, op
[0]);
1815 emit(ir
, TGSI_OPCODE_SNE
, result_dst
, op
[0], st_src_reg_for_float(0.0));
1818 emit(ir
, TGSI_OPCODE_TRUNC
, result_dst
, op
[0]);
1821 emit(ir
, TGSI_OPCODE_CEIL
, result_dst
, op
[0]);
1824 emit(ir
, TGSI_OPCODE_FLR
, result_dst
, op
[0]);
1826 case ir_unop_round_even
:
1827 emit(ir
, TGSI_OPCODE_ROUND
, result_dst
, op
[0]);
1830 emit(ir
, TGSI_OPCODE_FRC
, result_dst
, op
[0]);
1834 emit(ir
, TGSI_OPCODE_MIN
, result_dst
, op
[0], op
[1]);
1837 emit(ir
, TGSI_OPCODE_MAX
, result_dst
, op
[0], op
[1]);
1840 emit_scalar(ir
, TGSI_OPCODE_POW
, result_dst
, op
[0], op
[1]);
1843 case ir_unop_bit_not
:
1844 if (native_integers
) {
1845 emit(ir
, TGSI_OPCODE_NOT
, result_dst
, op
[0]);
1849 if (native_integers
) {
1850 emit(ir
, TGSI_OPCODE_U2F
, result_dst
, op
[0]);
1853 case ir_binop_lshift
:
1854 if (native_integers
) {
1855 emit(ir
, TGSI_OPCODE_SHL
, result_dst
, op
[0], op
[1]);
1858 case ir_binop_rshift
:
1859 if (native_integers
) {
1860 emit(ir
, TGSI_OPCODE_ISHR
, result_dst
, op
[0], op
[1]);
1863 case ir_binop_bit_and
:
1864 if (native_integers
) {
1865 emit(ir
, TGSI_OPCODE_AND
, result_dst
, op
[0], op
[1]);
1868 case ir_binop_bit_xor
:
1869 if (native_integers
) {
1870 emit(ir
, TGSI_OPCODE_XOR
, result_dst
, op
[0], op
[1]);
1873 case ir_binop_bit_or
:
1874 if (native_integers
) {
1875 emit(ir
, TGSI_OPCODE_OR
, result_dst
, op
[0], op
[1]);
1879 assert(!"GLSL 1.30 features unsupported");
1882 case ir_binop_ubo_load
: {
1883 ir_constant
*uniform_block
= ir
->operands
[0]->as_constant();
1884 ir_constant
*const_offset_ir
= ir
->operands
[1]->as_constant();
1885 unsigned const_offset
= const_offset_ir
? const_offset_ir
->value
.u
[0] : 0;
1886 st_src_reg index_reg
= get_temp(glsl_type::uint_type
);
1889 cbuf
.type
= glsl_type::vec4_type
->base_type
;
1890 cbuf
.file
= PROGRAM_CONSTANT
;
1892 cbuf
.index2D
= uniform_block
->value
.u
[0] + 1;
1893 cbuf
.reladdr
= NULL
;
1896 assert(ir
->type
->is_vector() || ir
->type
->is_scalar());
1898 if (const_offset_ir
) {
1899 index_reg
= st_src_reg_for_int(const_offset
/ 16);
1901 emit(ir
, TGSI_OPCODE_USHR
, st_dst_reg(index_reg
), op
[1], st_src_reg_for_int(4));
1904 cbuf
.swizzle
= swizzle_for_size(ir
->type
->vector_elements
);
1905 cbuf
.swizzle
+= MAKE_SWIZZLE4(const_offset
% 16 / 4,
1906 const_offset
% 16 / 4,
1907 const_offset
% 16 / 4,
1908 const_offset
% 16 / 4);
1910 cbuf
.reladdr
= ralloc(mem_ctx
, st_src_reg
);
1911 memcpy(cbuf
.reladdr
, &index_reg
, sizeof(index_reg
));
1913 if (ir
->type
->base_type
== GLSL_TYPE_BOOL
) {
1914 emit(ir
, TGSI_OPCODE_USNE
, result_dst
, cbuf
, st_src_reg_for_int(0));
1916 emit(ir
, TGSI_OPCODE_MOV
, result_dst
, cbuf
);
1921 /* note: we have to reorder the three args here */
1922 emit(ir
, TGSI_OPCODE_LRP
, result_dst
, op
[2], op
[1], op
[0]);
1925 if (this->ctx
->Const
.NativeIntegers
)
1926 emit(ir
, TGSI_OPCODE_UCMP
, result_dst
, op
[0], op
[1], op
[2]);
1928 op
[0].negate
= ~op
[0].negate
;
1929 emit(ir
, TGSI_OPCODE_CMP
, result_dst
, op
[0], op
[1], op
[2]);
1932 case ir_unop_pack_snorm_2x16
:
1933 case ir_unop_pack_unorm_2x16
:
1934 case ir_unop_pack_half_2x16
:
1935 case ir_unop_pack_snorm_4x8
:
1936 case ir_unop_pack_unorm_4x8
:
1937 case ir_unop_unpack_snorm_2x16
:
1938 case ir_unop_unpack_unorm_2x16
:
1939 case ir_unop_unpack_half_2x16
:
1940 case ir_unop_unpack_half_2x16_split_x
:
1941 case ir_unop_unpack_half_2x16_split_y
:
1942 case ir_unop_unpack_snorm_4x8
:
1943 case ir_unop_unpack_unorm_4x8
:
1944 case ir_binop_pack_half_2x16_split
:
1945 case ir_unop_bitfield_reverse
:
1946 case ir_unop_bit_count
:
1947 case ir_unop_find_msb
:
1948 case ir_unop_find_lsb
:
1952 case ir_triop_bitfield_extract
:
1953 case ir_quadop_bitfield_insert
:
1954 case ir_quadop_vector
:
1955 case ir_binop_vector_extract
:
1956 case ir_triop_vector_insert
:
1957 case ir_binop_ldexp
:
1958 case ir_binop_carry
:
1959 case ir_binop_borrow
:
1960 case ir_binop_imul_high
:
1961 /* This operation is not supported, or should have already been handled.
1963 assert(!"Invalid ir opcode in glsl_to_tgsi_visitor::visit()");
1967 this->result
= result_src
;
1972 glsl_to_tgsi_visitor::visit(ir_swizzle
*ir
)
1978 /* Note that this is only swizzles in expressions, not those on the left
1979 * hand side of an assignment, which do write masking. See ir_assignment
1983 ir
->val
->accept(this);
1985 assert(src
.file
!= PROGRAM_UNDEFINED
);
1987 for (i
= 0; i
< 4; i
++) {
1988 if (i
< ir
->type
->vector_elements
) {
1991 swizzle
[i
] = GET_SWZ(src
.swizzle
, ir
->mask
.x
);
1994 swizzle
[i
] = GET_SWZ(src
.swizzle
, ir
->mask
.y
);
1997 swizzle
[i
] = GET_SWZ(src
.swizzle
, ir
->mask
.z
);
2000 swizzle
[i
] = GET_SWZ(src
.swizzle
, ir
->mask
.w
);
2004 /* If the type is smaller than a vec4, replicate the last
2007 swizzle
[i
] = swizzle
[ir
->type
->vector_elements
- 1];
2011 src
.swizzle
= MAKE_SWIZZLE4(swizzle
[0], swizzle
[1], swizzle
[2], swizzle
[3]);
2017 glsl_to_tgsi_visitor::visit(ir_dereference_variable
*ir
)
2019 variable_storage
*entry
= find_variable_storage(ir
->var
);
2020 ir_variable
*var
= ir
->var
;
2023 switch (var
->mode
) {
2024 case ir_var_uniform
:
2025 entry
= new(mem_ctx
) variable_storage(var
, PROGRAM_UNIFORM
,
2027 this->variables
.push_tail(entry
);
2029 case ir_var_shader_in
:
2030 /* The linker assigns locations for varyings and attributes,
2031 * including deprecated builtins (like gl_Color), user-assign
2032 * generic attributes (glBindVertexLocation), and
2033 * user-defined varyings.
2035 assert(var
->location
!= -1);
2036 entry
= new(mem_ctx
) variable_storage(var
,
2040 case ir_var_shader_out
:
2041 assert(var
->location
!= -1);
2042 entry
= new(mem_ctx
) variable_storage(var
,
2044 var
->location
+ var
->index
);
2046 case ir_var_system_value
:
2047 entry
= new(mem_ctx
) variable_storage(var
,
2048 PROGRAM_SYSTEM_VALUE
,
2052 case ir_var_temporary
:
2053 st_src_reg src
= get_temp(var
->type
);
2055 entry
= new(mem_ctx
) variable_storage(var
, src
.file
, src
.index
);
2056 this->variables
.push_tail(entry
);
2062 printf("Failed to make storage for %s\n", var
->name
);
2067 this->result
= st_src_reg(entry
->file
, entry
->index
, var
->type
);
2068 if (!native_integers
)
2069 this->result
.type
= GLSL_TYPE_FLOAT
;
2073 glsl_to_tgsi_visitor::visit(ir_dereference_array
*ir
)
2077 int element_size
= type_size(ir
->type
);
2080 index
= ir
->array_index
->constant_expression_value();
2082 ir
->array
->accept(this);
2085 is_2D_input
= this->prog
->Target
== GL_GEOMETRY_PROGRAM_NV
&&
2086 src
.file
== PROGRAM_INPUT
&&
2087 ir
->array
->ir_type
!= ir_type_dereference_array
;
2094 src
.index2D
= index
->value
.i
[0];
2095 src
.has_index2
= true;
2097 src
.index
+= index
->value
.i
[0] * element_size
;
2099 /* Variable index array dereference. It eats the "vec4" of the
2100 * base of the array and an index that offsets the TGSI register
2103 ir
->array_index
->accept(this);
2105 st_src_reg index_reg
;
2107 if (element_size
== 1) {
2108 index_reg
= this->result
;
2110 index_reg
= get_temp(native_integers
?
2111 glsl_type::int_type
: glsl_type::float_type
);
2113 emit(ir
, TGSI_OPCODE_MUL
, st_dst_reg(index_reg
),
2114 this->result
, st_src_reg_for_type(index_reg
.type
, element_size
));
2117 /* If there was already a relative address register involved, add the
2118 * new and the old together to get the new offset.
2120 if (!is_2D_input
&& src
.reladdr
!= NULL
) {
2121 st_src_reg accum_reg
= get_temp(native_integers
?
2122 glsl_type::int_type
: glsl_type::float_type
);
2124 emit(ir
, TGSI_OPCODE_ADD
, st_dst_reg(accum_reg
),
2125 index_reg
, *src
.reladdr
);
2127 index_reg
= accum_reg
;
2131 src
.reladdr2
= ralloc(mem_ctx
, st_src_reg
);
2132 memcpy(src
.reladdr2
, &index_reg
, sizeof(index_reg
));
2134 src
.has_index2
= true;
2136 src
.reladdr
= ralloc(mem_ctx
, st_src_reg
);
2137 memcpy(src
.reladdr
, &index_reg
, sizeof(index_reg
));
2141 /* If the type is smaller than a vec4, replicate the last channel out. */
2142 if (ir
->type
->is_scalar() || ir
->type
->is_vector())
2143 src
.swizzle
= swizzle_for_size(ir
->type
->vector_elements
);
2145 src
.swizzle
= SWIZZLE_NOOP
;
2147 /* Change the register type to the element type of the array. */
2148 src
.type
= ir
->type
->base_type
;
2154 glsl_to_tgsi_visitor::visit(ir_dereference_record
*ir
)
2157 const glsl_type
*struct_type
= ir
->record
->type
;
2160 ir
->record
->accept(this);
2162 for (i
= 0; i
< struct_type
->length
; i
++) {
2163 if (strcmp(struct_type
->fields
.structure
[i
].name
, ir
->field
) == 0)
2165 offset
+= type_size(struct_type
->fields
.structure
[i
].type
);
2168 /* If the type is smaller than a vec4, replicate the last channel out. */
2169 if (ir
->type
->is_scalar() || ir
->type
->is_vector())
2170 this->result
.swizzle
= swizzle_for_size(ir
->type
->vector_elements
);
2172 this->result
.swizzle
= SWIZZLE_NOOP
;
2174 this->result
.index
+= offset
;
2175 this->result
.type
= ir
->type
->base_type
;
2179 * We want to be careful in assignment setup to hit the actual storage
2180 * instead of potentially using a temporary like we might with the
2181 * ir_dereference handler.
2184 get_assignment_lhs(ir_dereference
*ir
, glsl_to_tgsi_visitor
*v
)
2186 /* The LHS must be a dereference. If the LHS is a variable indexed array
2187 * access of a vector, it must be separated into a series conditional moves
2188 * before reaching this point (see ir_vec_index_to_cond_assign).
2190 assert(ir
->as_dereference());
2191 ir_dereference_array
*deref_array
= ir
->as_dereference_array();
2193 assert(!deref_array
->array
->type
->is_vector());
2196 /* Use the rvalue deref handler for the most part. We'll ignore
2197 * swizzles in it and write swizzles using writemask, though.
2200 return st_dst_reg(v
->result
);
2204 * Process the condition of a conditional assignment
2206 * Examines the condition of a conditional assignment to generate the optimal
2207 * first operand of a \c CMP instruction. If the condition is a relational
2208 * operator with 0 (e.g., \c ir_binop_less), the value being compared will be
2209 * used as the source for the \c CMP instruction. Otherwise the comparison
2210 * is processed to a boolean result, and the boolean result is used as the
2211 * operand to the CMP instruction.
2214 glsl_to_tgsi_visitor::process_move_condition(ir_rvalue
*ir
)
2216 ir_rvalue
*src_ir
= ir
;
2218 bool switch_order
= false;
2220 ir_expression
*const expr
= ir
->as_expression();
2221 if ((expr
!= NULL
) && (expr
->get_num_operands() == 2)) {
2222 bool zero_on_left
= false;
2224 if (expr
->operands
[0]->is_zero()) {
2225 src_ir
= expr
->operands
[1];
2226 zero_on_left
= true;
2227 } else if (expr
->operands
[1]->is_zero()) {
2228 src_ir
= expr
->operands
[0];
2229 zero_on_left
= false;
2233 * (a < 0) T F F ( a < 0) T F F
2234 * (0 < a) F F T (-a < 0) F F T
2235 * (a <= 0) T T F (-a < 0) F F T (swap order of other operands)
2236 * (0 <= a) F T T ( a < 0) T F F (swap order of other operands)
2237 * (a > 0) F F T (-a < 0) F F T
2238 * (0 > a) T F F ( a < 0) T F F
2239 * (a >= 0) F T T ( a < 0) T F F (swap order of other operands)
2240 * (0 >= a) T T F (-a < 0) F F T (swap order of other operands)
2242 * Note that exchanging the order of 0 and 'a' in the comparison simply
2243 * means that the value of 'a' should be negated.
2246 switch (expr
->operation
) {
2248 switch_order
= false;
2249 negate
= zero_on_left
;
2252 case ir_binop_greater
:
2253 switch_order
= false;
2254 negate
= !zero_on_left
;
2257 case ir_binop_lequal
:
2258 switch_order
= true;
2259 negate
= !zero_on_left
;
2262 case ir_binop_gequal
:
2263 switch_order
= true;
2264 negate
= zero_on_left
;
2268 /* This isn't the right kind of comparison afterall, so make sure
2269 * the whole condition is visited.
2277 src_ir
->accept(this);
2279 /* We use the TGSI_OPCODE_CMP (a < 0 ? b : c) for conditional moves, and the
2280 * condition we produced is 0.0 or 1.0. By flipping the sign, we can
2281 * choose which value TGSI_OPCODE_CMP produces without an extra instruction
2282 * computing the condition.
2285 this->result
.negate
= ~this->result
.negate
;
2287 return switch_order
;
2291 glsl_to_tgsi_visitor::emit_block_mov(ir_assignment
*ir
, const struct glsl_type
*type
,
2292 st_dst_reg
*l
, st_src_reg
*r
)
2294 if (type
->base_type
== GLSL_TYPE_STRUCT
) {
2295 for (unsigned int i
= 0; i
< type
->length
; i
++) {
2296 emit_block_mov(ir
, type
->fields
.structure
[i
].type
, l
, r
);
2301 if (type
->is_array()) {
2302 for (unsigned int i
= 0; i
< type
->length
; i
++) {
2303 emit_block_mov(ir
, type
->fields
.array
, l
, r
);
2308 if (type
->is_matrix()) {
2309 const struct glsl_type
*vec_type
;
2311 vec_type
= glsl_type::get_instance(GLSL_TYPE_FLOAT
,
2312 type
->vector_elements
, 1);
2314 for (int i
= 0; i
< type
->matrix_columns
; i
++) {
2315 emit_block_mov(ir
, vec_type
, l
, r
);
2320 assert(type
->is_scalar() || type
->is_vector());
2322 r
->type
= type
->base_type
;
2323 emit(ir
, TGSI_OPCODE_MOV
, *l
, *r
);
2329 glsl_to_tgsi_visitor::visit(ir_assignment
*ir
)
2335 ir
->rhs
->accept(this);
2338 l
= get_assignment_lhs(ir
->lhs
, this);
2340 /* FINISHME: This should really set to the correct maximal writemask for each
2341 * FINISHME: component written (in the loops below). This case can only
2342 * FINISHME: occur for matrices, arrays, and structures.
2344 if (ir
->write_mask
== 0) {
2345 assert(!ir
->lhs
->type
->is_scalar() && !ir
->lhs
->type
->is_vector());
2346 l
.writemask
= WRITEMASK_XYZW
;
2347 } else if (ir
->lhs
->type
->is_scalar() &&
2348 ir
->lhs
->variable_referenced()->mode
== ir_var_shader_out
) {
2349 /* FINISHME: This hack makes writing to gl_FragDepth, which lives in the
2350 * FINISHME: W component of fragment shader output zero, work correctly.
2352 l
.writemask
= WRITEMASK_XYZW
;
2355 int first_enabled_chan
= 0;
2358 l
.writemask
= ir
->write_mask
;
2360 for (int i
= 0; i
< 4; i
++) {
2361 if (l
.writemask
& (1 << i
)) {
2362 first_enabled_chan
= GET_SWZ(r
.swizzle
, i
);
2367 /* Swizzle a small RHS vector into the channels being written.
2369 * glsl ir treats write_mask as dictating how many channels are
2370 * present on the RHS while TGSI treats write_mask as just
2371 * showing which channels of the vec4 RHS get written.
2373 for (int i
= 0; i
< 4; i
++) {
2374 if (l
.writemask
& (1 << i
))
2375 swizzles
[i
] = GET_SWZ(r
.swizzle
, rhs_chan
++);
2377 swizzles
[i
] = first_enabled_chan
;
2379 r
.swizzle
= MAKE_SWIZZLE4(swizzles
[0], swizzles
[1],
2380 swizzles
[2], swizzles
[3]);
2383 assert(l
.file
!= PROGRAM_UNDEFINED
);
2384 assert(r
.file
!= PROGRAM_UNDEFINED
);
2386 if (ir
->condition
) {
2387 const bool switch_order
= this->process_move_condition(ir
->condition
);
2388 st_src_reg condition
= this->result
;
2390 for (i
= 0; i
< type_size(ir
->lhs
->type
); i
++) {
2391 st_src_reg l_src
= st_src_reg(l
);
2392 st_src_reg condition_temp
= condition
;
2393 l_src
.swizzle
= swizzle_for_size(ir
->lhs
->type
->vector_elements
);
2395 if (native_integers
) {
2396 /* This is necessary because TGSI's CMP instruction expects the
2397 * condition to be a float, and we store booleans as integers.
2398 * TODO: really want to avoid i2f path and use UCMP. Requires
2399 * changes to process_move_condition though too.
2401 condition_temp
= get_temp(glsl_type::vec4_type
);
2402 condition
.negate
= 0;
2403 emit(ir
, TGSI_OPCODE_I2F
, st_dst_reg(condition_temp
), condition
);
2404 condition_temp
.swizzle
= condition
.swizzle
;
2408 emit(ir
, TGSI_OPCODE_CMP
, l
, condition_temp
, l_src
, r
);
2410 emit(ir
, TGSI_OPCODE_CMP
, l
, condition_temp
, r
, l_src
);
2416 } else if (ir
->rhs
->as_expression() &&
2417 this->instructions
.get_tail() &&
2418 ir
->rhs
== ((glsl_to_tgsi_instruction
*)this->instructions
.get_tail())->ir
&&
2419 type_size(ir
->lhs
->type
) == 1 &&
2420 l
.writemask
== ((glsl_to_tgsi_instruction
*)this->instructions
.get_tail())->dst
.writemask
) {
2421 /* To avoid emitting an extra MOV when assigning an expression to a
2422 * variable, emit the last instruction of the expression again, but
2423 * replace the destination register with the target of the assignment.
2424 * Dead code elimination will remove the original instruction.
2426 glsl_to_tgsi_instruction
*inst
, *new_inst
;
2427 inst
= (glsl_to_tgsi_instruction
*)this->instructions
.get_tail();
2428 new_inst
= emit(ir
, inst
->op
, l
, inst
->src
[0], inst
->src
[1], inst
->src
[2]);
2429 new_inst
->saturate
= inst
->saturate
;
2430 inst
->dead_mask
= inst
->dst
.writemask
;
2432 emit_block_mov(ir
, ir
->rhs
->type
, &l
, &r
);
2438 glsl_to_tgsi_visitor::visit(ir_constant
*ir
)
2441 GLfloat stack_vals
[4] = { 0 };
2442 gl_constant_value
*values
= (gl_constant_value
*) stack_vals
;
2443 GLenum gl_type
= GL_NONE
;
2445 static int in_array
= 0;
2446 gl_register_file file
= in_array
? PROGRAM_CONSTANT
: PROGRAM_IMMEDIATE
;
2448 /* Unfortunately, 4 floats is all we can get into
2449 * _mesa_add_typed_unnamed_constant. So, make a temp to store an
2450 * aggregate constant and move each constant value into it. If we
2451 * get lucky, copy propagation will eliminate the extra moves.
2453 if (ir
->type
->base_type
== GLSL_TYPE_STRUCT
) {
2454 st_src_reg temp_base
= get_temp(ir
->type
);
2455 st_dst_reg temp
= st_dst_reg(temp_base
);
2457 foreach_iter(exec_list_iterator
, iter
, ir
->components
) {
2458 ir_constant
*field_value
= (ir_constant
*)iter
.get();
2459 int size
= type_size(field_value
->type
);
2463 field_value
->accept(this);
2466 for (i
= 0; i
< (unsigned int)size
; i
++) {
2467 emit(ir
, TGSI_OPCODE_MOV
, temp
, src
);
2473 this->result
= temp_base
;
2477 if (ir
->type
->is_array()) {
2478 st_src_reg temp_base
= get_temp(ir
->type
);
2479 st_dst_reg temp
= st_dst_reg(temp_base
);
2480 int size
= type_size(ir
->type
->fields
.array
);
2485 for (i
= 0; i
< ir
->type
->length
; i
++) {
2486 ir
->array_elements
[i
]->accept(this);
2488 for (int j
= 0; j
< size
; j
++) {
2489 emit(ir
, TGSI_OPCODE_MOV
, temp
, src
);
2495 this->result
= temp_base
;
2500 if (ir
->type
->is_matrix()) {
2501 st_src_reg mat
= get_temp(ir
->type
);
2502 st_dst_reg mat_column
= st_dst_reg(mat
);
2504 for (i
= 0; i
< ir
->type
->matrix_columns
; i
++) {
2505 assert(ir
->type
->base_type
== GLSL_TYPE_FLOAT
);
2506 values
= (gl_constant_value
*) &ir
->value
.f
[i
* ir
->type
->vector_elements
];
2508 src
= st_src_reg(file
, -1, ir
->type
->base_type
);
2509 src
.index
= add_constant(file
,
2511 ir
->type
->vector_elements
,
2514 emit(ir
, TGSI_OPCODE_MOV
, mat_column
, src
);
2523 switch (ir
->type
->base_type
) {
2524 case GLSL_TYPE_FLOAT
:
2526 for (i
= 0; i
< ir
->type
->vector_elements
; i
++) {
2527 values
[i
].f
= ir
->value
.f
[i
];
2530 case GLSL_TYPE_UINT
:
2531 gl_type
= native_integers
? GL_UNSIGNED_INT
: GL_FLOAT
;
2532 for (i
= 0; i
< ir
->type
->vector_elements
; i
++) {
2533 if (native_integers
)
2534 values
[i
].u
= ir
->value
.u
[i
];
2536 values
[i
].f
= ir
->value
.u
[i
];
2540 gl_type
= native_integers
? GL_INT
: GL_FLOAT
;
2541 for (i
= 0; i
< ir
->type
->vector_elements
; i
++) {
2542 if (native_integers
)
2543 values
[i
].i
= ir
->value
.i
[i
];
2545 values
[i
].f
= ir
->value
.i
[i
];
2548 case GLSL_TYPE_BOOL
:
2549 gl_type
= native_integers
? GL_BOOL
: GL_FLOAT
;
2550 for (i
= 0; i
< ir
->type
->vector_elements
; i
++) {
2551 if (native_integers
)
2552 values
[i
].u
= ir
->value
.b
[i
] ? ~0 : 0;
2554 values
[i
].f
= ir
->value
.b
[i
];
2558 assert(!"Non-float/uint/int/bool constant");
2561 this->result
= st_src_reg(file
, -1, ir
->type
);
2562 this->result
.index
= add_constant(file
,
2564 ir
->type
->vector_elements
,
2566 &this->result
.swizzle
);
2570 glsl_to_tgsi_visitor::get_function_signature(ir_function_signature
*sig
)
2572 function_entry
*entry
;
2574 foreach_iter(exec_list_iterator
, iter
, this->function_signatures
) {
2575 entry
= (function_entry
*)iter
.get();
2577 if (entry
->sig
== sig
)
2581 entry
= ralloc(mem_ctx
, function_entry
);
2583 entry
->sig_id
= this->next_signature_id
++;
2584 entry
->bgn_inst
= NULL
;
2586 /* Allocate storage for all the parameters. */
2587 foreach_iter(exec_list_iterator
, iter
, sig
->parameters
) {
2588 ir_variable
*param
= (ir_variable
*)iter
.get();
2589 variable_storage
*storage
;
2591 storage
= find_variable_storage(param
);
2594 st_src_reg src
= get_temp(param
->type
);
2596 storage
= new(mem_ctx
) variable_storage(param
, src
.file
, src
.index
);
2597 this->variables
.push_tail(storage
);
2600 if (!sig
->return_type
->is_void()) {
2601 entry
->return_reg
= get_temp(sig
->return_type
);
2603 entry
->return_reg
= undef_src
;
2606 this->function_signatures
.push_tail(entry
);
2611 glsl_to_tgsi_visitor::visit(ir_call
*ir
)
2613 glsl_to_tgsi_instruction
*call_inst
;
2614 ir_function_signature
*sig
= ir
->callee
;
2615 function_entry
*entry
= get_function_signature(sig
);
2618 /* Process in parameters. */
2619 exec_list_iterator sig_iter
= sig
->parameters
.iterator();
2620 foreach_iter(exec_list_iterator
, iter
, *ir
) {
2621 ir_rvalue
*param_rval
= (ir_rvalue
*)iter
.get();
2622 ir_variable
*param
= (ir_variable
*)sig_iter
.get();
2624 if (param
->mode
== ir_var_function_in
||
2625 param
->mode
== ir_var_function_inout
) {
2626 variable_storage
*storage
= find_variable_storage(param
);
2629 param_rval
->accept(this);
2630 st_src_reg r
= this->result
;
2633 l
.file
= storage
->file
;
2634 l
.index
= storage
->index
;
2636 l
.writemask
= WRITEMASK_XYZW
;
2637 l
.cond_mask
= COND_TR
;
2639 for (i
= 0; i
< type_size(param
->type
); i
++) {
2640 emit(ir
, TGSI_OPCODE_MOV
, l
, r
);
2648 assert(!sig_iter
.has_next());
2650 /* Emit call instruction */
2651 call_inst
= emit(ir
, TGSI_OPCODE_CAL
);
2652 call_inst
->function
= entry
;
2654 /* Process out parameters. */
2655 sig_iter
= sig
->parameters
.iterator();
2656 foreach_iter(exec_list_iterator
, iter
, *ir
) {
2657 ir_rvalue
*param_rval
= (ir_rvalue
*)iter
.get();
2658 ir_variable
*param
= (ir_variable
*)sig_iter
.get();
2660 if (param
->mode
== ir_var_function_out
||
2661 param
->mode
== ir_var_function_inout
) {
2662 variable_storage
*storage
= find_variable_storage(param
);
2666 r
.file
= storage
->file
;
2667 r
.index
= storage
->index
;
2669 r
.swizzle
= SWIZZLE_NOOP
;
2672 param_rval
->accept(this);
2673 st_dst_reg l
= st_dst_reg(this->result
);
2675 for (i
= 0; i
< type_size(param
->type
); i
++) {
2676 emit(ir
, TGSI_OPCODE_MOV
, l
, r
);
2684 assert(!sig_iter
.has_next());
2686 /* Process return value. */
2687 this->result
= entry
->return_reg
;
2691 glsl_to_tgsi_visitor::visit(ir_texture
*ir
)
2693 st_src_reg result_src
, coord
, cube_sc
, lod_info
, projector
, dx
, dy
, offset
, sample_index
;
2694 st_dst_reg result_dst
, coord_dst
, cube_sc_dst
;
2695 glsl_to_tgsi_instruction
*inst
= NULL
;
2696 unsigned opcode
= TGSI_OPCODE_NOP
;
2697 const glsl_type
*sampler_type
= ir
->sampler
->type
;
2698 bool is_cube_array
= false;
2700 /* if we are a cube array sampler */
2701 if ((sampler_type
->sampler_dimensionality
== GLSL_SAMPLER_DIM_CUBE
&&
2702 sampler_type
->sampler_array
)) {
2703 is_cube_array
= true;
2706 if (ir
->coordinate
) {
2707 ir
->coordinate
->accept(this);
2709 /* Put our coords in a temp. We'll need to modify them for shadow,
2710 * projection, or LOD, so the only case we'd use it as is is if
2711 * we're doing plain old texturing. The optimization passes on
2712 * glsl_to_tgsi_visitor should handle cleaning up our mess in that case.
2714 coord
= get_temp(glsl_type::vec4_type
);
2715 coord_dst
= st_dst_reg(coord
);
2716 coord_dst
.writemask
= (1 << ir
->coordinate
->type
->vector_elements
) - 1;
2717 emit(ir
, TGSI_OPCODE_MOV
, coord_dst
, this->result
);
2720 if (ir
->projector
) {
2721 ir
->projector
->accept(this);
2722 projector
= this->result
;
2725 /* Storage for our result. Ideally for an assignment we'd be using
2726 * the actual storage for the result here, instead.
2728 result_src
= get_temp(ir
->type
);
2729 result_dst
= st_dst_reg(result_src
);
2733 opcode
= (is_cube_array
&& ir
->shadow_comparitor
) ? TGSI_OPCODE_TEX2
: TGSI_OPCODE_TEX
;
2735 ir
->offset
->accept(this);
2736 offset
= this->result
;
2740 opcode
= is_cube_array
? TGSI_OPCODE_TXB2
: TGSI_OPCODE_TXB
;
2741 ir
->lod_info
.bias
->accept(this);
2742 lod_info
= this->result
;
2744 ir
->offset
->accept(this);
2745 offset
= this->result
;
2749 opcode
= is_cube_array
? TGSI_OPCODE_TXL2
: TGSI_OPCODE_TXL
;
2750 ir
->lod_info
.lod
->accept(this);
2751 lod_info
= this->result
;
2753 ir
->offset
->accept(this);
2754 offset
= this->result
;
2758 opcode
= TGSI_OPCODE_TXD
;
2759 ir
->lod_info
.grad
.dPdx
->accept(this);
2761 ir
->lod_info
.grad
.dPdy
->accept(this);
2764 ir
->offset
->accept(this);
2765 offset
= this->result
;
2769 opcode
= TGSI_OPCODE_TXQ
;
2770 ir
->lod_info
.lod
->accept(this);
2771 lod_info
= this->result
;
2774 opcode
= TGSI_OPCODE_TXF
;
2775 ir
->lod_info
.lod
->accept(this);
2776 lod_info
= this->result
;
2778 ir
->offset
->accept(this);
2779 offset
= this->result
;
2783 opcode
= TGSI_OPCODE_TXF
;
2784 ir
->lod_info
.sample_index
->accept(this);
2785 sample_index
= this->result
;
2788 assert(!"Unexpected ir_lod opcode");
2791 assert(!"Unexpected ir_tg4 opcode");
2793 case ir_query_levels
:
2794 assert(!"Unexpected ir_query_levels opcode");
2798 if (ir
->projector
) {
2799 if (opcode
== TGSI_OPCODE_TEX
) {
2800 /* Slot the projector in as the last component of the coord. */
2801 coord_dst
.writemask
= WRITEMASK_W
;
2802 emit(ir
, TGSI_OPCODE_MOV
, coord_dst
, projector
);
2803 coord_dst
.writemask
= WRITEMASK_XYZW
;
2804 opcode
= TGSI_OPCODE_TXP
;
2806 st_src_reg coord_w
= coord
;
2807 coord_w
.swizzle
= SWIZZLE_WWWW
;
2809 /* For the other TEX opcodes there's no projective version
2810 * since the last slot is taken up by LOD info. Do the
2811 * projective divide now.
2813 coord_dst
.writemask
= WRITEMASK_W
;
2814 emit(ir
, TGSI_OPCODE_RCP
, coord_dst
, projector
);
2816 /* In the case where we have to project the coordinates "by hand,"
2817 * the shadow comparator value must also be projected.
2819 st_src_reg tmp_src
= coord
;
2820 if (ir
->shadow_comparitor
) {
2821 /* Slot the shadow value in as the second to last component of the
2824 ir
->shadow_comparitor
->accept(this);
2826 tmp_src
= get_temp(glsl_type::vec4_type
);
2827 st_dst_reg tmp_dst
= st_dst_reg(tmp_src
);
2829 /* Projective division not allowed for array samplers. */
2830 assert(!sampler_type
->sampler_array
);
2832 tmp_dst
.writemask
= WRITEMASK_Z
;
2833 emit(ir
, TGSI_OPCODE_MOV
, tmp_dst
, this->result
);
2835 tmp_dst
.writemask
= WRITEMASK_XY
;
2836 emit(ir
, TGSI_OPCODE_MOV
, tmp_dst
, coord
);
2839 coord_dst
.writemask
= WRITEMASK_XYZ
;
2840 emit(ir
, TGSI_OPCODE_MUL
, coord_dst
, tmp_src
, coord_w
);
2842 coord_dst
.writemask
= WRITEMASK_XYZW
;
2843 coord
.swizzle
= SWIZZLE_XYZW
;
2847 /* If projection is done and the opcode is not TGSI_OPCODE_TXP, then the shadow
2848 * comparator was put in the correct place (and projected) by the code,
2849 * above, that handles by-hand projection.
2851 if (ir
->shadow_comparitor
&& (!ir
->projector
|| opcode
== TGSI_OPCODE_TXP
)) {
2852 /* Slot the shadow value in as the second to last component of the
2855 ir
->shadow_comparitor
->accept(this);
2857 if (is_cube_array
) {
2858 cube_sc
= get_temp(glsl_type::float_type
);
2859 cube_sc_dst
= st_dst_reg(cube_sc
);
2860 cube_sc_dst
.writemask
= WRITEMASK_X
;
2861 emit(ir
, TGSI_OPCODE_MOV
, cube_sc_dst
, this->result
);
2862 cube_sc_dst
.writemask
= WRITEMASK_X
;
2865 if ((sampler_type
->sampler_dimensionality
== GLSL_SAMPLER_DIM_2D
&&
2866 sampler_type
->sampler_array
) ||
2867 sampler_type
->sampler_dimensionality
== GLSL_SAMPLER_DIM_CUBE
) {
2868 coord_dst
.writemask
= WRITEMASK_W
;
2870 coord_dst
.writemask
= WRITEMASK_Z
;
2873 emit(ir
, TGSI_OPCODE_MOV
, coord_dst
, this->result
);
2874 coord_dst
.writemask
= WRITEMASK_XYZW
;
2878 if (ir
->op
== ir_txf_ms
) {
2879 coord_dst
.writemask
= WRITEMASK_W
;
2880 emit(ir
, TGSI_OPCODE_MOV
, coord_dst
, sample_index
);
2881 coord_dst
.writemask
= WRITEMASK_XYZW
;
2882 } else if (opcode
== TGSI_OPCODE_TXL
|| opcode
== TGSI_OPCODE_TXB
||
2883 opcode
== TGSI_OPCODE_TXF
) {
2884 /* TGSI stores LOD or LOD bias in the last channel of the coords. */
2885 coord_dst
.writemask
= WRITEMASK_W
;
2886 emit(ir
, TGSI_OPCODE_MOV
, coord_dst
, lod_info
);
2887 coord_dst
.writemask
= WRITEMASK_XYZW
;
2890 if (opcode
== TGSI_OPCODE_TXD
)
2891 inst
= emit(ir
, opcode
, result_dst
, coord
, dx
, dy
);
2892 else if (opcode
== TGSI_OPCODE_TXQ
)
2893 inst
= emit(ir
, opcode
, result_dst
, lod_info
);
2894 else if (opcode
== TGSI_OPCODE_TXF
) {
2895 inst
= emit(ir
, opcode
, result_dst
, coord
);
2896 } else if (opcode
== TGSI_OPCODE_TXL2
|| opcode
== TGSI_OPCODE_TXB2
) {
2897 inst
= emit(ir
, opcode
, result_dst
, coord
, lod_info
);
2898 } else if (opcode
== TGSI_OPCODE_TEX2
) {
2899 inst
= emit(ir
, opcode
, result_dst
, coord
, cube_sc
);
2901 inst
= emit(ir
, opcode
, result_dst
, coord
);
2903 if (ir
->shadow_comparitor
)
2904 inst
->tex_shadow
= GL_TRUE
;
2906 inst
->sampler
= _mesa_get_sampler_uniform_value(ir
->sampler
,
2907 this->shader_program
,
2911 inst
->tex_offset_num_offset
= 1;
2912 inst
->tex_offsets
[0].Index
= offset
.index
;
2913 inst
->tex_offsets
[0].File
= offset
.file
;
2914 inst
->tex_offsets
[0].SwizzleX
= GET_SWZ(offset
.swizzle
, 0);
2915 inst
->tex_offsets
[0].SwizzleY
= GET_SWZ(offset
.swizzle
, 1);
2916 inst
->tex_offsets
[0].SwizzleZ
= GET_SWZ(offset
.swizzle
, 2);
2919 switch (sampler_type
->sampler_dimensionality
) {
2920 case GLSL_SAMPLER_DIM_1D
:
2921 inst
->tex_target
= (sampler_type
->sampler_array
)
2922 ? TEXTURE_1D_ARRAY_INDEX
: TEXTURE_1D_INDEX
;
2924 case GLSL_SAMPLER_DIM_2D
:
2925 inst
->tex_target
= (sampler_type
->sampler_array
)
2926 ? TEXTURE_2D_ARRAY_INDEX
: TEXTURE_2D_INDEX
;
2928 case GLSL_SAMPLER_DIM_3D
:
2929 inst
->tex_target
= TEXTURE_3D_INDEX
;
2931 case GLSL_SAMPLER_DIM_CUBE
:
2932 inst
->tex_target
= (sampler_type
->sampler_array
)
2933 ? TEXTURE_CUBE_ARRAY_INDEX
: TEXTURE_CUBE_INDEX
;
2935 case GLSL_SAMPLER_DIM_RECT
:
2936 inst
->tex_target
= TEXTURE_RECT_INDEX
;
2938 case GLSL_SAMPLER_DIM_BUF
:
2939 inst
->tex_target
= TEXTURE_BUFFER_INDEX
;
2941 case GLSL_SAMPLER_DIM_EXTERNAL
:
2942 inst
->tex_target
= TEXTURE_EXTERNAL_INDEX
;
2944 case GLSL_SAMPLER_DIM_MS
:
2945 inst
->tex_target
= (sampler_type
->sampler_array
)
2946 ? TEXTURE_2D_MULTISAMPLE_ARRAY_INDEX
: TEXTURE_2D_MULTISAMPLE_INDEX
;
2949 assert(!"Should not get here.");
2952 this->result
= result_src
;
2956 glsl_to_tgsi_visitor::visit(ir_return
*ir
)
2958 if (ir
->get_value()) {
2962 assert(current_function
);
2964 ir
->get_value()->accept(this);
2965 st_src_reg r
= this->result
;
2967 l
= st_dst_reg(current_function
->return_reg
);
2969 for (i
= 0; i
< type_size(current_function
->sig
->return_type
); i
++) {
2970 emit(ir
, TGSI_OPCODE_MOV
, l
, r
);
2976 emit(ir
, TGSI_OPCODE_RET
);
2980 glsl_to_tgsi_visitor::visit(ir_discard
*ir
)
2982 if (ir
->condition
) {
2983 ir
->condition
->accept(this);
2984 this->result
.negate
= ~this->result
.negate
;
2985 emit(ir
, TGSI_OPCODE_KILL_IF
, undef_dst
, this->result
);
2987 /* unconditional kil */
2988 emit(ir
, TGSI_OPCODE_KILL
);
2993 glsl_to_tgsi_visitor::visit(ir_if
*ir
)
2996 glsl_to_tgsi_instruction
*if_inst
;
2998 ir
->condition
->accept(this);
2999 assert(this->result
.file
!= PROGRAM_UNDEFINED
);
3001 if_opcode
= native_integers
? TGSI_OPCODE_UIF
: TGSI_OPCODE_IF
;
3003 if_inst
= emit(ir
->condition
, if_opcode
, undef_dst
, this->result
);
3005 this->instructions
.push_tail(if_inst
);
3007 visit_exec_list(&ir
->then_instructions
, this);
3009 if (!ir
->else_instructions
.is_empty()) {
3010 emit(ir
->condition
, TGSI_OPCODE_ELSE
);
3011 visit_exec_list(&ir
->else_instructions
, this);
3014 if_inst
= emit(ir
->condition
, TGSI_OPCODE_ENDIF
);
3019 glsl_to_tgsi_visitor::visit(ir_emit_vertex
*ir
)
3021 assert(this->prog
->Target
== GL_GEOMETRY_PROGRAM_NV
);
3022 emit(ir
, TGSI_OPCODE_EMIT
);
3026 glsl_to_tgsi_visitor::visit(ir_end_primitive
*ir
)
3028 assert(this->prog
->Target
== GL_GEOMETRY_PROGRAM_NV
);
3029 emit(ir
, TGSI_OPCODE_ENDPRIM
);
3032 glsl_to_tgsi_visitor::glsl_to_tgsi_visitor()
3034 result
.file
= PROGRAM_UNDEFINED
;
3037 next_signature_id
= 1;
3039 current_function
= NULL
;
3040 num_address_regs
= 0;
3042 indirect_addr_consts
= false;
3044 native_integers
= false;
3045 mem_ctx
= ralloc_context(NULL
);
3048 shader_program
= NULL
;
3052 glsl_to_tgsi_visitor::~glsl_to_tgsi_visitor()
3054 ralloc_free(mem_ctx
);
3057 extern "C" void free_glsl_to_tgsi_visitor(glsl_to_tgsi_visitor
*v
)
3064 * Count resources used by the given gpu program (number of texture
3068 count_resources(glsl_to_tgsi_visitor
*v
, gl_program
*prog
)
3070 v
->samplers_used
= 0;
3072 foreach_iter(exec_list_iterator
, iter
, v
->instructions
) {
3073 glsl_to_tgsi_instruction
*inst
= (glsl_to_tgsi_instruction
*)iter
.get();
3075 if (is_tex_instruction(inst
->op
)) {
3076 v
->samplers_used
|= 1 << inst
->sampler
;
3078 if (inst
->tex_shadow
) {
3079 prog
->ShadowSamplers
|= 1 << inst
->sampler
;
3084 prog
->SamplersUsed
= v
->samplers_used
;
3086 if (v
->shader_program
!= NULL
)
3087 _mesa_update_shader_textures_used(v
->shader_program
, prog
);
3091 set_uniform_initializer(struct gl_context
*ctx
, void *mem_ctx
,
3092 struct gl_shader_program
*shader_program
,
3093 const char *name
, const glsl_type
*type
,
3096 if (type
->is_record()) {
3097 ir_constant
*field_constant
;
3099 field_constant
= (ir_constant
*)val
->components
.get_head();
3101 for (unsigned int i
= 0; i
< type
->length
; i
++) {
3102 const glsl_type
*field_type
= type
->fields
.structure
[i
].type
;
3103 const char *field_name
= ralloc_asprintf(mem_ctx
, "%s.%s", name
,
3104 type
->fields
.structure
[i
].name
);
3105 set_uniform_initializer(ctx
, mem_ctx
, shader_program
, field_name
,
3106 field_type
, field_constant
);
3107 field_constant
= (ir_constant
*)field_constant
->next
;
3113 unsigned index
= _mesa_get_uniform_location(ctx
, shader_program
, name
,
3115 if (offset
== GL_INVALID_INDEX
) {
3116 fail_link(shader_program
,
3117 "Couldn't find uniform for initializer %s\n", name
);
3120 int loc
= _mesa_uniform_merge_location_offset(shader_program
, index
, offset
);
3122 for (unsigned int i
= 0; i
< (type
->is_array() ? type
->length
: 1); i
++) {
3123 ir_constant
*element
;
3124 const glsl_type
*element_type
;
3125 if (type
->is_array()) {
3126 element
= val
->array_elements
[i
];
3127 element_type
= type
->fields
.array
;
3130 element_type
= type
;
3135 if (element_type
->base_type
== GLSL_TYPE_BOOL
) {
3136 int *conv
= ralloc_array(mem_ctx
, int, element_type
->components());
3137 for (unsigned int j
= 0; j
< element_type
->components(); j
++) {
3138 conv
[j
] = element
->value
.b
[j
];
3140 values
= (void *)conv
;
3141 element_type
= glsl_type::get_instance(GLSL_TYPE_INT
,
3142 element_type
->vector_elements
,
3145 values
= &element
->value
;
3148 if (element_type
->is_matrix()) {
3149 _mesa_uniform_matrix(ctx
, shader_program
,
3150 element_type
->matrix_columns
,
3151 element_type
->vector_elements
,
3152 loc
, 1, GL_FALSE
, (GLfloat
*)values
);
3154 _mesa_uniform(ctx
, shader_program
, loc
, element_type
->matrix_columns
,
3155 values
, element_type
->gl_type
);
3163 * Returns the mask of channels (bitmask of WRITEMASK_X,Y,Z,W) which
3164 * are read from the given src in this instruction
3167 get_src_arg_mask(st_dst_reg dst
, st_src_reg src
)
3169 int read_mask
= 0, comp
;
3171 /* Now, given the src swizzle and the written channels, find which
3172 * components are actually read
3174 for (comp
= 0; comp
< 4; ++comp
) {
3175 const unsigned coord
= GET_SWZ(src
.swizzle
, comp
);
3177 if (dst
.writemask
& (1 << comp
) && coord
<= SWIZZLE_W
)
3178 read_mask
|= 1 << coord
;
3185 * This pass replaces CMP T0, T1 T2 T0 with MOV T0, T2 when the CMP
3186 * instruction is the first instruction to write to register T0. There are
3187 * several lowering passes done in GLSL IR (e.g. branches and
3188 * relative addressing) that create a large number of conditional assignments
3189 * that ir_to_mesa converts to CMP instructions like the one mentioned above.
3191 * Here is why this conversion is safe:
3192 * CMP T0, T1 T2 T0 can be expanded to:
3198 * If (T1 < 0.0) evaluates to true then our replacement MOV T0, T2 is the same
3199 * as the original program. If (T1 < 0.0) evaluates to false, executing
3200 * MOV T0, T0 will store a garbage value in T0 since T0 is uninitialized.
3201 * Therefore, it doesn't matter that we are replacing MOV T0, T0 with MOV T0, T2
3202 * because any instruction that was going to read from T0 after this was going
3203 * to read a garbage value anyway.
3206 glsl_to_tgsi_visitor::simplify_cmp(void)
3208 unsigned *tempWrites
;
3209 unsigned outputWrites
[MAX_PROGRAM_OUTPUTS
];
3211 tempWrites
= new unsigned[MAX_TEMPS
];
3215 memset(tempWrites
, 0, sizeof(unsigned) * MAX_TEMPS
);
3216 memset(outputWrites
, 0, sizeof(outputWrites
));
3218 foreach_iter(exec_list_iterator
, iter
, this->instructions
) {
3219 glsl_to_tgsi_instruction
*inst
= (glsl_to_tgsi_instruction
*)iter
.get();
3220 unsigned prevWriteMask
= 0;
3222 /* Give up if we encounter relative addressing or flow control. */
3223 if (inst
->dst
.reladdr
||
3224 tgsi_get_opcode_info(inst
->op
)->is_branch
||
3225 inst
->op
== TGSI_OPCODE_BGNSUB
||
3226 inst
->op
== TGSI_OPCODE_CONT
||
3227 inst
->op
== TGSI_OPCODE_END
||
3228 inst
->op
== TGSI_OPCODE_ENDSUB
||
3229 inst
->op
== TGSI_OPCODE_RET
) {
3233 if (inst
->dst
.file
== PROGRAM_OUTPUT
) {
3234 assert(inst
->dst
.index
< MAX_PROGRAM_OUTPUTS
);
3235 prevWriteMask
= outputWrites
[inst
->dst
.index
];
3236 outputWrites
[inst
->dst
.index
] |= inst
->dst
.writemask
;
3237 } else if (inst
->dst
.file
== PROGRAM_TEMPORARY
) {
3238 assert(inst
->dst
.index
< MAX_TEMPS
);
3239 prevWriteMask
= tempWrites
[inst
->dst
.index
];
3240 tempWrites
[inst
->dst
.index
] |= inst
->dst
.writemask
;
3244 /* For a CMP to be considered a conditional write, the destination
3245 * register and source register two must be the same. */
3246 if (inst
->op
== TGSI_OPCODE_CMP
3247 && !(inst
->dst
.writemask
& prevWriteMask
)
3248 && inst
->src
[2].file
== inst
->dst
.file
3249 && inst
->src
[2].index
== inst
->dst
.index
3250 && inst
->dst
.writemask
== get_src_arg_mask(inst
->dst
, inst
->src
[2])) {
3252 inst
->op
= TGSI_OPCODE_MOV
;
3253 inst
->src
[0] = inst
->src
[1];
3257 delete [] tempWrites
;
3260 /* Replaces all references to a temporary register index with another index. */
3262 glsl_to_tgsi_visitor::rename_temp_register(int index
, int new_index
)
3264 foreach_iter(exec_list_iterator
, iter
, this->instructions
) {
3265 glsl_to_tgsi_instruction
*inst
= (glsl_to_tgsi_instruction
*)iter
.get();
3268 for (j
=0; j
< num_inst_src_regs(inst
->op
); j
++) {
3269 if (inst
->src
[j
].file
== PROGRAM_TEMPORARY
&&
3270 inst
->src
[j
].index
== index
) {
3271 inst
->src
[j
].index
= new_index
;
3275 if (inst
->dst
.file
== PROGRAM_TEMPORARY
&& inst
->dst
.index
== index
) {
3276 inst
->dst
.index
= new_index
;
3282 glsl_to_tgsi_visitor::get_first_temp_read(int index
)
3284 int depth
= 0; /* loop depth */
3285 int loop_start
= -1; /* index of the first active BGNLOOP (if any) */
3288 foreach_iter(exec_list_iterator
, iter
, this->instructions
) {
3289 glsl_to_tgsi_instruction
*inst
= (glsl_to_tgsi_instruction
*)iter
.get();
3291 for (j
=0; j
< num_inst_src_regs(inst
->op
); j
++) {
3292 if (inst
->src
[j
].file
== PROGRAM_TEMPORARY
&&
3293 inst
->src
[j
].index
== index
) {
3294 return (depth
== 0) ? i
: loop_start
;
3298 if (inst
->op
== TGSI_OPCODE_BGNLOOP
) {
3301 } else if (inst
->op
== TGSI_OPCODE_ENDLOOP
) {
3314 glsl_to_tgsi_visitor::get_first_temp_write(int index
)
3316 int depth
= 0; /* loop depth */
3317 int loop_start
= -1; /* index of the first active BGNLOOP (if any) */
3320 foreach_iter(exec_list_iterator
, iter
, this->instructions
) {
3321 glsl_to_tgsi_instruction
*inst
= (glsl_to_tgsi_instruction
*)iter
.get();
3323 if (inst
->dst
.file
== PROGRAM_TEMPORARY
&& inst
->dst
.index
== index
) {
3324 return (depth
== 0) ? i
: loop_start
;
3327 if (inst
->op
== TGSI_OPCODE_BGNLOOP
) {
3330 } else if (inst
->op
== TGSI_OPCODE_ENDLOOP
) {
3343 glsl_to_tgsi_visitor::get_last_temp_read(int index
)
3345 int depth
= 0; /* loop depth */
3346 int last
= -1; /* index of last instruction that reads the temporary */
3349 foreach_iter(exec_list_iterator
, iter
, this->instructions
) {
3350 glsl_to_tgsi_instruction
*inst
= (glsl_to_tgsi_instruction
*)iter
.get();
3352 for (j
=0; j
< num_inst_src_regs(inst
->op
); j
++) {
3353 if (inst
->src
[j
].file
== PROGRAM_TEMPORARY
&&
3354 inst
->src
[j
].index
== index
) {
3355 last
= (depth
== 0) ? i
: -2;
3359 if (inst
->op
== TGSI_OPCODE_BGNLOOP
)
3361 else if (inst
->op
== TGSI_OPCODE_ENDLOOP
)
3362 if (--depth
== 0 && last
== -2)
3374 glsl_to_tgsi_visitor::get_last_temp_write(int index
)
3376 int depth
= 0; /* loop depth */
3377 int last
= -1; /* index of last instruction that writes to the temporary */
3380 foreach_iter(exec_list_iterator
, iter
, this->instructions
) {
3381 glsl_to_tgsi_instruction
*inst
= (glsl_to_tgsi_instruction
*)iter
.get();
3383 if (inst
->dst
.file
== PROGRAM_TEMPORARY
&& inst
->dst
.index
== index
)
3384 last
= (depth
== 0) ? i
: -2;
3386 if (inst
->op
== TGSI_OPCODE_BGNLOOP
)
3388 else if (inst
->op
== TGSI_OPCODE_ENDLOOP
)
3389 if (--depth
== 0 && last
== -2)
3401 * On a basic block basis, tracks available PROGRAM_TEMPORARY register
3402 * channels for copy propagation and updates following instructions to
3403 * use the original versions.
3405 * The glsl_to_tgsi_visitor lazily produces code assuming that this pass
3406 * will occur. As an example, a TXP production before this pass:
3408 * 0: MOV TEMP[1], INPUT[4].xyyy;
3409 * 1: MOV TEMP[1].w, INPUT[4].wwww;
3410 * 2: TXP TEMP[2], TEMP[1], texture[0], 2D;
3414 * 0: MOV TEMP[1], INPUT[4].xyyy;
3415 * 1: MOV TEMP[1].w, INPUT[4].wwww;
3416 * 2: TXP TEMP[2], INPUT[4].xyyw, texture[0], 2D;
3418 * which allows for dead code elimination on TEMP[1]'s writes.
3421 glsl_to_tgsi_visitor::copy_propagate(void)
3423 glsl_to_tgsi_instruction
**acp
= rzalloc_array(mem_ctx
,
3424 glsl_to_tgsi_instruction
*,
3425 this->next_temp
* 4);
3426 int *acp_level
= rzalloc_array(mem_ctx
, int, this->next_temp
* 4);
3429 foreach_iter(exec_list_iterator
, iter
, this->instructions
) {
3430 glsl_to_tgsi_instruction
*inst
= (glsl_to_tgsi_instruction
*)iter
.get();
3432 assert(inst
->dst
.file
!= PROGRAM_TEMPORARY
3433 || inst
->dst
.index
< this->next_temp
);
3435 /* First, do any copy propagation possible into the src regs. */
3436 for (int r
= 0; r
< 3; r
++) {
3437 glsl_to_tgsi_instruction
*first
= NULL
;
3439 int acp_base
= inst
->src
[r
].index
* 4;
3441 if (inst
->src
[r
].file
!= PROGRAM_TEMPORARY
||
3442 inst
->src
[r
].reladdr
||
3443 inst
->src
[r
].reladdr2
)
3446 /* See if we can find entries in the ACP consisting of MOVs
3447 * from the same src register for all the swizzled channels
3448 * of this src register reference.
3450 for (int i
= 0; i
< 4; i
++) {
3451 int src_chan
= GET_SWZ(inst
->src
[r
].swizzle
, i
);
3452 glsl_to_tgsi_instruction
*copy_chan
= acp
[acp_base
+ src_chan
];
3459 assert(acp_level
[acp_base
+ src_chan
] <= level
);
3464 if (first
->src
[0].file
!= copy_chan
->src
[0].file
||
3465 first
->src
[0].index
!= copy_chan
->src
[0].index
) {
3473 /* We've now validated that we can copy-propagate to
3474 * replace this src register reference. Do it.
3476 inst
->src
[r
].file
= first
->src
[0].file
;
3477 inst
->src
[r
].index
= first
->src
[0].index
;
3478 inst
->src
[r
].index2D
= first
->src
[0].index2D
;
3479 inst
->src
[r
].has_index2
= first
->src
[0].has_index2
;
3482 for (int i
= 0; i
< 4; i
++) {
3483 int src_chan
= GET_SWZ(inst
->src
[r
].swizzle
, i
);
3484 glsl_to_tgsi_instruction
*copy_inst
= acp
[acp_base
+ src_chan
];
3485 swizzle
|= (GET_SWZ(copy_inst
->src
[0].swizzle
, src_chan
) <<
3488 inst
->src
[r
].swizzle
= swizzle
;
3493 case TGSI_OPCODE_BGNLOOP
:
3494 case TGSI_OPCODE_ENDLOOP
:
3495 /* End of a basic block, clear the ACP entirely. */
3496 memset(acp
, 0, sizeof(*acp
) * this->next_temp
* 4);
3499 case TGSI_OPCODE_IF
:
3500 case TGSI_OPCODE_UIF
:
3504 case TGSI_OPCODE_ENDIF
:
3505 case TGSI_OPCODE_ELSE
:
3506 /* Clear all channels written inside the block from the ACP, but
3507 * leaving those that were not touched.
3509 for (int r
= 0; r
< this->next_temp
; r
++) {
3510 for (int c
= 0; c
< 4; c
++) {
3511 if (!acp
[4 * r
+ c
])
3514 if (acp_level
[4 * r
+ c
] >= level
)
3515 acp
[4 * r
+ c
] = NULL
;
3518 if (inst
->op
== TGSI_OPCODE_ENDIF
)
3523 /* Continuing the block, clear any written channels from
3526 if (inst
->dst
.file
== PROGRAM_TEMPORARY
&& inst
->dst
.reladdr
) {
3527 /* Any temporary might be written, so no copy propagation
3528 * across this instruction.
3530 memset(acp
, 0, sizeof(*acp
) * this->next_temp
* 4);
3531 } else if (inst
->dst
.file
== PROGRAM_OUTPUT
&&
3532 inst
->dst
.reladdr
) {
3533 /* Any output might be written, so no copy propagation
3534 * from outputs across this instruction.
3536 for (int r
= 0; r
< this->next_temp
; r
++) {
3537 for (int c
= 0; c
< 4; c
++) {
3538 if (!acp
[4 * r
+ c
])
3541 if (acp
[4 * r
+ c
]->src
[0].file
== PROGRAM_OUTPUT
)
3542 acp
[4 * r
+ c
] = NULL
;
3545 } else if (inst
->dst
.file
== PROGRAM_TEMPORARY
||
3546 inst
->dst
.file
== PROGRAM_OUTPUT
) {
3547 /* Clear where it's used as dst. */
3548 if (inst
->dst
.file
== PROGRAM_TEMPORARY
) {
3549 for (int c
= 0; c
< 4; c
++) {
3550 if (inst
->dst
.writemask
& (1 << c
)) {
3551 acp
[4 * inst
->dst
.index
+ c
] = NULL
;
3556 /* Clear where it's used as src. */
3557 for (int r
= 0; r
< this->next_temp
; r
++) {
3558 for (int c
= 0; c
< 4; c
++) {
3559 if (!acp
[4 * r
+ c
])
3562 int src_chan
= GET_SWZ(acp
[4 * r
+ c
]->src
[0].swizzle
, c
);
3564 if (acp
[4 * r
+ c
]->src
[0].file
== inst
->dst
.file
&&
3565 acp
[4 * r
+ c
]->src
[0].index
== inst
->dst
.index
&&
3566 inst
->dst
.writemask
& (1 << src_chan
))
3568 acp
[4 * r
+ c
] = NULL
;
3576 /* If this is a copy, add it to the ACP. */
3577 if (inst
->op
== TGSI_OPCODE_MOV
&&
3578 inst
->dst
.file
== PROGRAM_TEMPORARY
&&
3579 !(inst
->dst
.file
== inst
->src
[0].file
&&
3580 inst
->dst
.index
== inst
->src
[0].index
) &&
3581 !inst
->dst
.reladdr
&&
3583 !inst
->src
[0].reladdr
&&
3584 !inst
->src
[0].reladdr2
&&
3585 !inst
->src
[0].negate
) {
3586 for (int i
= 0; i
< 4; i
++) {
3587 if (inst
->dst
.writemask
& (1 << i
)) {
3588 acp
[4 * inst
->dst
.index
+ i
] = inst
;
3589 acp_level
[4 * inst
->dst
.index
+ i
] = level
;
3595 ralloc_free(acp_level
);
3600 * Tracks available PROGRAM_TEMPORARY registers for dead code elimination.
3602 * The glsl_to_tgsi_visitor lazily produces code assuming that this pass
3603 * will occur. As an example, a TXP production after copy propagation but
3606 * 0: MOV TEMP[1], INPUT[4].xyyy;
3607 * 1: MOV TEMP[1].w, INPUT[4].wwww;
3608 * 2: TXP TEMP[2], INPUT[4].xyyw, texture[0], 2D;
3610 * and after this pass:
3612 * 0: TXP TEMP[2], INPUT[4].xyyw, texture[0], 2D;
3614 * FIXME: assumes that all functions are inlined (no support for BGNSUB/ENDSUB)
3615 * FIXME: doesn't eliminate all dead code inside of loops; it steps around them
3618 glsl_to_tgsi_visitor::eliminate_dead_code(void)
3622 for (i
=0; i
< this->next_temp
; i
++) {
3623 int last_read
= get_last_temp_read(i
);
3626 foreach_iter(exec_list_iterator
, iter
, this->instructions
) {
3627 glsl_to_tgsi_instruction
*inst
= (glsl_to_tgsi_instruction
*)iter
.get();
3629 if (inst
->dst
.file
== PROGRAM_TEMPORARY
&& inst
->dst
.index
== i
&&
3642 * On a basic block basis, tracks available PROGRAM_TEMPORARY registers for dead
3643 * code elimination. This is less primitive than eliminate_dead_code(), as it
3644 * is per-channel and can detect consecutive writes without a read between them
3645 * as dead code. However, there is some dead code that can be eliminated by
3646 * eliminate_dead_code() but not this function - for example, this function
3647 * cannot eliminate an instruction writing to a register that is never read and
3648 * is the only instruction writing to that register.
3650 * The glsl_to_tgsi_visitor lazily produces code assuming that this pass
3654 glsl_to_tgsi_visitor::eliminate_dead_code_advanced(void)
3656 glsl_to_tgsi_instruction
**writes
= rzalloc_array(mem_ctx
,
3657 glsl_to_tgsi_instruction
*,
3658 this->next_temp
* 4);
3659 int *write_level
= rzalloc_array(mem_ctx
, int, this->next_temp
* 4);
3663 foreach_iter(exec_list_iterator
, iter
, this->instructions
) {
3664 glsl_to_tgsi_instruction
*inst
= (glsl_to_tgsi_instruction
*)iter
.get();
3666 assert(inst
->dst
.file
!= PROGRAM_TEMPORARY
3667 || inst
->dst
.index
< this->next_temp
);
3670 case TGSI_OPCODE_BGNLOOP
:
3671 case TGSI_OPCODE_ENDLOOP
:
3672 case TGSI_OPCODE_CONT
:
3673 case TGSI_OPCODE_BRK
:
3674 /* End of a basic block, clear the write array entirely.
3676 * This keeps us from killing dead code when the writes are
3677 * on either side of a loop, even when the register isn't touched
3678 * inside the loop. However, glsl_to_tgsi_visitor doesn't seem to emit
3679 * dead code of this type, so it shouldn't make a difference as long as
3680 * the dead code elimination pass in the GLSL compiler does its job.
3682 memset(writes
, 0, sizeof(*writes
) * this->next_temp
* 4);
3685 case TGSI_OPCODE_ENDIF
:
3686 case TGSI_OPCODE_ELSE
:
3687 /* Promote the recorded level of all channels written inside the
3688 * preceding if or else block to the level above the if/else block.
3690 for (int r
= 0; r
< this->next_temp
; r
++) {
3691 for (int c
= 0; c
< 4; c
++) {
3692 if (!writes
[4 * r
+ c
])
3695 if (write_level
[4 * r
+ c
] == level
)
3696 write_level
[4 * r
+ c
] = level
-1;
3700 if(inst
->op
== TGSI_OPCODE_ENDIF
)
3705 case TGSI_OPCODE_IF
:
3706 case TGSI_OPCODE_UIF
:
3708 /* fallthrough to default case to mark the condition as read */
3711 /* Continuing the block, clear any channels from the write array that
3712 * are read by this instruction.
3714 for (unsigned i
= 0; i
< Elements(inst
->src
); i
++) {
3715 if (inst
->src
[i
].file
== PROGRAM_TEMPORARY
&& inst
->src
[i
].reladdr
){
3716 /* Any temporary might be read, so no dead code elimination
3717 * across this instruction.
3719 memset(writes
, 0, sizeof(*writes
) * this->next_temp
* 4);
3720 } else if (inst
->src
[i
].file
== PROGRAM_TEMPORARY
) {
3721 /* Clear where it's used as src. */
3722 int src_chans
= 1 << GET_SWZ(inst
->src
[i
].swizzle
, 0);
3723 src_chans
|= 1 << GET_SWZ(inst
->src
[i
].swizzle
, 1);
3724 src_chans
|= 1 << GET_SWZ(inst
->src
[i
].swizzle
, 2);
3725 src_chans
|= 1 << GET_SWZ(inst
->src
[i
].swizzle
, 3);
3727 for (int c
= 0; c
< 4; c
++) {
3728 if (src_chans
& (1 << c
)) {
3729 writes
[4 * inst
->src
[i
].index
+ c
] = NULL
;
3737 /* If this instruction writes to a temporary, add it to the write array.
3738 * If there is already an instruction in the write array for one or more
3739 * of the channels, flag that channel write as dead.
3741 if (inst
->dst
.file
== PROGRAM_TEMPORARY
&&
3742 !inst
->dst
.reladdr
&&
3744 for (int c
= 0; c
< 4; c
++) {
3745 if (inst
->dst
.writemask
& (1 << c
)) {
3746 if (writes
[4 * inst
->dst
.index
+ c
]) {
3747 if (write_level
[4 * inst
->dst
.index
+ c
] < level
)
3750 writes
[4 * inst
->dst
.index
+ c
]->dead_mask
|= (1 << c
);
3752 writes
[4 * inst
->dst
.index
+ c
] = inst
;
3753 write_level
[4 * inst
->dst
.index
+ c
] = level
;
3759 /* Anything still in the write array at this point is dead code. */
3760 for (int r
= 0; r
< this->next_temp
; r
++) {
3761 for (int c
= 0; c
< 4; c
++) {
3762 glsl_to_tgsi_instruction
*inst
= writes
[4 * r
+ c
];
3764 inst
->dead_mask
|= (1 << c
);
3768 /* Now actually remove the instructions that are completely dead and update
3769 * the writemask of other instructions with dead channels.
3771 foreach_iter(exec_list_iterator
, iter
, this->instructions
) {
3772 glsl_to_tgsi_instruction
*inst
= (glsl_to_tgsi_instruction
*)iter
.get();
3774 if (!inst
->dead_mask
|| !inst
->dst
.writemask
)
3776 else if ((inst
->dst
.writemask
& ~inst
->dead_mask
) == 0) {
3781 inst
->dst
.writemask
&= ~(inst
->dead_mask
);
3784 ralloc_free(write_level
);
3785 ralloc_free(writes
);
3790 /* Merges temporary registers together where possible to reduce the number of
3791 * registers needed to run a program.
3793 * Produces optimal code only after copy propagation and dead code elimination
3796 glsl_to_tgsi_visitor::merge_registers(void)
3798 int *last_reads
= rzalloc_array(mem_ctx
, int, this->next_temp
);
3799 int *first_writes
= rzalloc_array(mem_ctx
, int, this->next_temp
);
3802 /* Read the indices of the last read and first write to each temp register
3803 * into an array so that we don't have to traverse the instruction list as
3805 for (i
=0; i
< this->next_temp
; i
++) {
3806 last_reads
[i
] = get_last_temp_read(i
);
3807 first_writes
[i
] = get_first_temp_write(i
);
3810 /* Start looking for registers with non-overlapping usages that can be
3811 * merged together. */
3812 for (i
=0; i
< this->next_temp
; i
++) {
3813 /* Don't touch unused registers. */
3814 if (last_reads
[i
] < 0 || first_writes
[i
] < 0) continue;
3816 for (j
=0; j
< this->next_temp
; j
++) {
3817 /* Don't touch unused registers. */
3818 if (last_reads
[j
] < 0 || first_writes
[j
] < 0) continue;
3820 /* We can merge the two registers if the first write to j is after or
3821 * in the same instruction as the last read from i. Note that the
3822 * register at index i will always be used earlier or at the same time
3823 * as the register at index j. */
3824 if (first_writes
[i
] <= first_writes
[j
] &&
3825 last_reads
[i
] <= first_writes
[j
])
3827 rename_temp_register(j
, i
); /* Replace all references to j with i.*/
3829 /* Update the first_writes and last_reads arrays with the new
3830 * values for the merged register index, and mark the newly unused
3831 * register index as such. */
3832 last_reads
[i
] = last_reads
[j
];
3833 first_writes
[j
] = -1;
3839 ralloc_free(last_reads
);
3840 ralloc_free(first_writes
);
3843 /* Reassign indices to temporary registers by reusing unused indices created
3844 * by optimization passes. */
3846 glsl_to_tgsi_visitor::renumber_registers(void)
3851 for (i
=0; i
< this->next_temp
; i
++) {
3852 if (get_first_temp_read(i
) < 0) continue;
3854 rename_temp_register(i
, new_index
);
3858 this->next_temp
= new_index
;
3862 * Returns a fragment program which implements the current pixel transfer ops.
3863 * Based on get_pixel_transfer_program in st_atom_pixeltransfer.c.
3866 get_pixel_transfer_visitor(struct st_fragment_program
*fp
,
3867 glsl_to_tgsi_visitor
*original
,
3868 int scale_and_bias
, int pixel_maps
)
3870 glsl_to_tgsi_visitor
*v
= new glsl_to_tgsi_visitor();
3871 struct st_context
*st
= st_context(original
->ctx
);
3872 struct gl_program
*prog
= &fp
->Base
.Base
;
3873 struct gl_program_parameter_list
*params
= _mesa_new_parameter_list();
3874 st_src_reg coord
, src0
;
3876 glsl_to_tgsi_instruction
*inst
;
3878 /* Copy attributes of the glsl_to_tgsi_visitor in the original shader. */
3879 v
->ctx
= original
->ctx
;
3881 v
->shader_program
= NULL
;
3882 v
->glsl_version
= original
->glsl_version
;
3883 v
->native_integers
= original
->native_integers
;
3884 v
->options
= original
->options
;
3885 v
->next_temp
= original
->next_temp
;
3886 v
->num_address_regs
= original
->num_address_regs
;
3887 v
->samplers_used
= prog
->SamplersUsed
= original
->samplers_used
;
3888 v
->indirect_addr_consts
= original
->indirect_addr_consts
;
3889 memcpy(&v
->immediates
, &original
->immediates
, sizeof(v
->immediates
));
3890 v
->num_immediates
= original
->num_immediates
;
3893 * Get initial pixel color from the texture.
3894 * TEX colorTemp, fragment.texcoord[0], texture[0], 2D;
3896 coord
= st_src_reg(PROGRAM_INPUT
, VARYING_SLOT_TEX0
, glsl_type::vec2_type
);
3897 src0
= v
->get_temp(glsl_type::vec4_type
);
3898 dst0
= st_dst_reg(src0
);
3899 inst
= v
->emit(NULL
, TGSI_OPCODE_TEX
, dst0
, coord
);
3901 inst
->tex_target
= TEXTURE_2D_INDEX
;
3903 prog
->InputsRead
|= VARYING_BIT_TEX0
;
3904 prog
->SamplersUsed
|= (1 << 0); /* mark sampler 0 as used */
3905 v
->samplers_used
|= (1 << 0);
3907 if (scale_and_bias
) {
3908 static const gl_state_index scale_state
[STATE_LENGTH
] =
3909 { STATE_INTERNAL
, STATE_PT_SCALE
,
3910 (gl_state_index
) 0, (gl_state_index
) 0, (gl_state_index
) 0 };
3911 static const gl_state_index bias_state
[STATE_LENGTH
] =
3912 { STATE_INTERNAL
, STATE_PT_BIAS
,
3913 (gl_state_index
) 0, (gl_state_index
) 0, (gl_state_index
) 0 };
3914 GLint scale_p
, bias_p
;
3915 st_src_reg scale
, bias
;
3917 scale_p
= _mesa_add_state_reference(params
, scale_state
);
3918 bias_p
= _mesa_add_state_reference(params
, bias_state
);
3920 /* MAD colorTemp, colorTemp, scale, bias; */
3921 scale
= st_src_reg(PROGRAM_STATE_VAR
, scale_p
, GLSL_TYPE_FLOAT
);
3922 bias
= st_src_reg(PROGRAM_STATE_VAR
, bias_p
, GLSL_TYPE_FLOAT
);
3923 inst
= v
->emit(NULL
, TGSI_OPCODE_MAD
, dst0
, src0
, scale
, bias
);
3927 st_src_reg temp
= v
->get_temp(glsl_type::vec4_type
);
3928 st_dst_reg temp_dst
= st_dst_reg(temp
);
3930 assert(st
->pixel_xfer
.pixelmap_texture
);
3932 /* With a little effort, we can do four pixel map look-ups with
3933 * two TEX instructions:
3936 /* TEX temp.rg, colorTemp.rgba, texture[1], 2D; */
3937 temp_dst
.writemask
= WRITEMASK_XY
; /* write R,G */
3938 inst
= v
->emit(NULL
, TGSI_OPCODE_TEX
, temp_dst
, src0
);
3940 inst
->tex_target
= TEXTURE_2D_INDEX
;
3942 /* TEX temp.ba, colorTemp.baba, texture[1], 2D; */
3943 src0
.swizzle
= MAKE_SWIZZLE4(SWIZZLE_Z
, SWIZZLE_W
, SWIZZLE_Z
, SWIZZLE_W
);
3944 temp_dst
.writemask
= WRITEMASK_ZW
; /* write B,A */
3945 inst
= v
->emit(NULL
, TGSI_OPCODE_TEX
, temp_dst
, src0
);
3947 inst
->tex_target
= TEXTURE_2D_INDEX
;
3949 prog
->SamplersUsed
|= (1 << 1); /* mark sampler 1 as used */
3950 v
->samplers_used
|= (1 << 1);
3952 /* MOV colorTemp, temp; */
3953 inst
= v
->emit(NULL
, TGSI_OPCODE_MOV
, dst0
, temp
);
3956 /* Now copy the instructions from the original glsl_to_tgsi_visitor into the
3958 foreach_iter(exec_list_iterator
, iter
, original
->instructions
) {
3959 glsl_to_tgsi_instruction
*inst
= (glsl_to_tgsi_instruction
*)iter
.get();
3960 glsl_to_tgsi_instruction
*newinst
;
3961 st_src_reg src_regs
[3];
3963 if (inst
->dst
.file
== PROGRAM_OUTPUT
)
3964 prog
->OutputsWritten
|= BITFIELD64_BIT(inst
->dst
.index
);
3966 for (int i
=0; i
<3; i
++) {
3967 src_regs
[i
] = inst
->src
[i
];
3968 if (src_regs
[i
].file
== PROGRAM_INPUT
&&
3969 src_regs
[i
].index
== VARYING_SLOT_COL0
)
3971 src_regs
[i
].file
= PROGRAM_TEMPORARY
;
3972 src_regs
[i
].index
= src0
.index
;
3974 else if (src_regs
[i
].file
== PROGRAM_INPUT
)
3975 prog
->InputsRead
|= BITFIELD64_BIT(src_regs
[i
].index
);
3978 newinst
= v
->emit(NULL
, inst
->op
, inst
->dst
, src_regs
[0], src_regs
[1], src_regs
[2]);
3979 newinst
->tex_target
= inst
->tex_target
;
3982 /* Make modifications to fragment program info. */
3983 prog
->Parameters
= _mesa_combine_parameter_lists(params
,
3984 original
->prog
->Parameters
);
3985 _mesa_free_parameter_list(params
);
3986 count_resources(v
, prog
);
3987 fp
->glsl_to_tgsi
= v
;
3991 * Make fragment program for glBitmap:
3992 * Sample the texture and kill the fragment if the bit is 0.
3993 * This program will be combined with the user's fragment program.
3995 * Based on make_bitmap_fragment_program in st_cb_bitmap.c.
3998 get_bitmap_visitor(struct st_fragment_program
*fp
,
3999 glsl_to_tgsi_visitor
*original
, int samplerIndex
)
4001 glsl_to_tgsi_visitor
*v
= new glsl_to_tgsi_visitor();
4002 struct st_context
*st
= st_context(original
->ctx
);
4003 struct gl_program
*prog
= &fp
->Base
.Base
;
4004 st_src_reg coord
, src0
;
4006 glsl_to_tgsi_instruction
*inst
;
4008 /* Copy attributes of the glsl_to_tgsi_visitor in the original shader. */
4009 v
->ctx
= original
->ctx
;
4011 v
->shader_program
= NULL
;
4012 v
->glsl_version
= original
->glsl_version
;
4013 v
->native_integers
= original
->native_integers
;
4014 v
->options
= original
->options
;
4015 v
->next_temp
= original
->next_temp
;
4016 v
->num_address_regs
= original
->num_address_regs
;
4017 v
->samplers_used
= prog
->SamplersUsed
= original
->samplers_used
;
4018 v
->indirect_addr_consts
= original
->indirect_addr_consts
;
4019 memcpy(&v
->immediates
, &original
->immediates
, sizeof(v
->immediates
));
4020 v
->num_immediates
= original
->num_immediates
;
4022 /* TEX tmp0, fragment.texcoord[0], texture[0], 2D; */
4023 coord
= st_src_reg(PROGRAM_INPUT
, VARYING_SLOT_TEX0
, glsl_type::vec2_type
);
4024 src0
= v
->get_temp(glsl_type::vec4_type
);
4025 dst0
= st_dst_reg(src0
);
4026 inst
= v
->emit(NULL
, TGSI_OPCODE_TEX
, dst0
, coord
);
4027 inst
->sampler
= samplerIndex
;
4028 inst
->tex_target
= TEXTURE_2D_INDEX
;
4030 prog
->InputsRead
|= VARYING_BIT_TEX0
;
4031 prog
->SamplersUsed
|= (1 << samplerIndex
); /* mark sampler as used */
4032 v
->samplers_used
|= (1 << samplerIndex
);
4034 /* KIL if -tmp0 < 0 # texel=0 -> keep / texel=0 -> discard */
4035 src0
.negate
= NEGATE_XYZW
;
4036 if (st
->bitmap
.tex_format
== PIPE_FORMAT_L8_UNORM
)
4037 src0
.swizzle
= SWIZZLE_XXXX
;
4038 inst
= v
->emit(NULL
, TGSI_OPCODE_KILL_IF
, undef_dst
, src0
);
4040 /* Now copy the instructions from the original glsl_to_tgsi_visitor into the
4042 foreach_iter(exec_list_iterator
, iter
, original
->instructions
) {
4043 glsl_to_tgsi_instruction
*inst
= (glsl_to_tgsi_instruction
*)iter
.get();
4044 glsl_to_tgsi_instruction
*newinst
;
4045 st_src_reg src_regs
[3];
4047 if (inst
->dst
.file
== PROGRAM_OUTPUT
)
4048 prog
->OutputsWritten
|= BITFIELD64_BIT(inst
->dst
.index
);
4050 for (int i
=0; i
<3; i
++) {
4051 src_regs
[i
] = inst
->src
[i
];
4052 if (src_regs
[i
].file
== PROGRAM_INPUT
)
4053 prog
->InputsRead
|= BITFIELD64_BIT(src_regs
[i
].index
);
4056 newinst
= v
->emit(NULL
, inst
->op
, inst
->dst
, src_regs
[0], src_regs
[1], src_regs
[2]);
4057 newinst
->tex_target
= inst
->tex_target
;
4060 /* Make modifications to fragment program info. */
4061 prog
->Parameters
= _mesa_clone_parameter_list(original
->prog
->Parameters
);
4062 count_resources(v
, prog
);
4063 fp
->glsl_to_tgsi
= v
;
4066 /* ------------------------- TGSI conversion stuff -------------------------- */
4068 unsigned branch_target
;
4073 * Intermediate state used during shader translation.
4075 struct st_translate
{
4076 struct ureg_program
*ureg
;
4078 struct ureg_dst temps
[MAX_TEMPS
];
4079 struct ureg_dst arrays
[MAX_ARRAYS
];
4080 struct ureg_src
*constants
;
4081 struct ureg_src
*immediates
;
4082 struct ureg_dst outputs
[PIPE_MAX_SHADER_OUTPUTS
];
4083 struct ureg_src inputs
[PIPE_MAX_SHADER_INPUTS
];
4084 struct ureg_dst address
[2];
4085 struct ureg_src samplers
[PIPE_MAX_SAMPLERS
];
4086 struct ureg_src systemValues
[SYSTEM_VALUE_MAX
];
4088 unsigned array_sizes
[MAX_ARRAYS
];
4090 const GLuint
*inputMapping
;
4091 const GLuint
*outputMapping
;
4093 /* For every instruction that contains a label (eg CALL), keep
4094 * details so that we can go back afterwards and emit the correct
4095 * tgsi instruction number for each label.
4097 struct label
*labels
;
4098 unsigned labels_size
;
4099 unsigned labels_count
;
4101 /* Keep a record of the tgsi instruction number that each mesa
4102 * instruction starts at, will be used to fix up labels after
4107 unsigned insn_count
;
4109 unsigned procType
; /**< TGSI_PROCESSOR_VERTEX/FRAGMENT */
4114 /** Map Mesa's SYSTEM_VALUE_x to TGSI_SEMANTIC_x */
4115 static unsigned mesa_sysval_to_semantic
[SYSTEM_VALUE_MAX
] = {
4117 TGSI_SEMANTIC_VERTEXID
,
4118 TGSI_SEMANTIC_INSTANCEID
4122 * Make note of a branch to a label in the TGSI code.
4123 * After we've emitted all instructions, we'll go over the list
4124 * of labels built here and patch the TGSI code with the actual
4125 * location of each label.
4127 static unsigned *get_label(struct st_translate
*t
, unsigned branch_target
)
4131 if (t
->labels_count
+ 1 >= t
->labels_size
) {
4132 t
->labels_size
= 1 << (util_logbase2(t
->labels_size
) + 1);
4133 t
->labels
= (struct label
*)realloc(t
->labels
,
4134 t
->labels_size
* sizeof(struct label
));
4135 if (t
->labels
== NULL
) {
4136 static unsigned dummy
;
4142 i
= t
->labels_count
++;
4143 t
->labels
[i
].branch_target
= branch_target
;
4144 return &t
->labels
[i
].token
;
4148 * Called prior to emitting the TGSI code for each instruction.
4149 * Allocate additional space for instructions if needed.
4150 * Update the insn[] array so the next glsl_to_tgsi_instruction points to
4151 * the next TGSI instruction.
4153 static void set_insn_start(struct st_translate
*t
, unsigned start
)
4155 if (t
->insn_count
+ 1 >= t
->insn_size
) {
4156 t
->insn_size
= 1 << (util_logbase2(t
->insn_size
) + 1);
4157 t
->insn
= (unsigned *)realloc(t
->insn
, t
->insn_size
* sizeof(t
->insn
[0]));
4158 if (t
->insn
== NULL
) {
4164 t
->insn
[t
->insn_count
++] = start
;
4168 * Map a glsl_to_tgsi constant/immediate to a TGSI immediate.
4170 static struct ureg_src
4171 emit_immediate(struct st_translate
*t
,
4172 gl_constant_value values
[4],
4175 struct ureg_program
*ureg
= t
->ureg
;
4180 return ureg_DECL_immediate(ureg
, &values
[0].f
, size
);
4182 return ureg_DECL_immediate_int(ureg
, &values
[0].i
, size
);
4183 case GL_UNSIGNED_INT
:
4185 return ureg_DECL_immediate_uint(ureg
, &values
[0].u
, size
);
4187 assert(!"should not get here - type must be float, int, uint, or bool");
4188 return ureg_src_undef();
4193 * Map a glsl_to_tgsi dst register to a TGSI ureg_dst register.
4195 static struct ureg_dst
4196 dst_register(struct st_translate
*t
,
4197 gl_register_file file
,
4203 case PROGRAM_UNDEFINED
:
4204 return ureg_dst_undef();
4206 case PROGRAM_TEMPORARY
:
4208 assert(index
< (int) Elements(t
->temps
));
4210 if (ureg_dst_is_undef(t
->temps
[index
]))
4211 t
->temps
[index
] = ureg_DECL_local_temporary(t
->ureg
);
4213 return t
->temps
[index
];
4216 array
= index
>> 16;
4219 assert(array
< (int) Elements(t
->arrays
));
4221 if (ureg_dst_is_undef(t
->arrays
[array
]))
4222 t
->arrays
[array
] = ureg_DECL_array_temporary(
4223 t
->ureg
, t
->array_sizes
[array
], TRUE
);
4225 return ureg_dst_array_offset(t
->arrays
[array
],
4226 (int)(index
& 0xFFFF) - 0x8000);
4228 case PROGRAM_OUTPUT
:
4229 if (t
->procType
== TGSI_PROCESSOR_VERTEX
)
4230 assert(index
< VARYING_SLOT_MAX
);
4231 else if (t
->procType
== TGSI_PROCESSOR_FRAGMENT
)
4232 assert(index
< FRAG_RESULT_MAX
);
4234 assert(index
< VARYING_SLOT_MAX
);
4236 assert(t
->outputMapping
[index
] < Elements(t
->outputs
));
4238 return t
->outputs
[t
->outputMapping
[index
]];
4240 case PROGRAM_ADDRESS
:
4241 return t
->address
[index
];
4244 assert(!"unknown dst register file");
4245 return ureg_dst_undef();
4250 * Map a glsl_to_tgsi src register to a TGSI ureg_src register.
4252 static struct ureg_src
4253 src_register(struct st_translate
*t
,
4254 gl_register_file file
,
4255 GLint index
, GLint index2D
)
4258 case PROGRAM_UNDEFINED
:
4259 return ureg_src_undef();
4261 case PROGRAM_TEMPORARY
:
4263 return ureg_src(dst_register(t
, file
, index
));
4265 case PROGRAM_UNIFORM
:
4267 return t
->constants
[index
];
4268 case PROGRAM_STATE_VAR
:
4269 case PROGRAM_CONSTANT
: /* ie, immediate */
4271 struct ureg_src src
;
4272 src
= ureg_src_register(TGSI_FILE_CONSTANT
, 0);
4274 src
.DimensionIndex
= index2D
;
4276 } else if (index
< 0)
4277 return ureg_DECL_constant(t
->ureg
, 0);
4279 return t
->constants
[index
];
4281 case PROGRAM_IMMEDIATE
:
4282 return t
->immediates
[index
];
4285 assert(t
->inputMapping
[index
] < Elements(t
->inputs
));
4286 return t
->inputs
[t
->inputMapping
[index
]];
4288 case PROGRAM_OUTPUT
:
4289 assert(t
->outputMapping
[index
] < Elements(t
->outputs
));
4290 return ureg_src(t
->outputs
[t
->outputMapping
[index
]]); /* not needed? */
4292 case PROGRAM_ADDRESS
:
4293 return ureg_src(t
->address
[index
]);
4295 case PROGRAM_SYSTEM_VALUE
:
4296 assert(index
< (int) Elements(t
->systemValues
));
4297 return t
->systemValues
[index
];
4300 assert(!"unknown src register file");
4301 return ureg_src_undef();
4306 * Create a TGSI ureg_dst register from an st_dst_reg.
4308 static struct ureg_dst
4309 translate_dst(struct st_translate
*t
,
4310 const st_dst_reg
*dst_reg
,
4311 bool saturate
, bool clamp_color
)
4313 struct ureg_dst dst
= dst_register(t
,
4317 dst
= ureg_writemask(dst
, dst_reg
->writemask
);
4320 dst
= ureg_saturate(dst
);
4321 else if (clamp_color
&& dst_reg
->file
== PROGRAM_OUTPUT
) {
4322 /* Clamp colors for ARB_color_buffer_float. */
4323 switch (t
->procType
) {
4324 case TGSI_PROCESSOR_VERTEX
:
4325 /* XXX if the geometry shader is present, this must be done there
4326 * instead of here. */
4327 if (dst_reg
->index
== VARYING_SLOT_COL0
||
4328 dst_reg
->index
== VARYING_SLOT_COL1
||
4329 dst_reg
->index
== VARYING_SLOT_BFC0
||
4330 dst_reg
->index
== VARYING_SLOT_BFC1
) {
4331 dst
= ureg_saturate(dst
);
4335 case TGSI_PROCESSOR_FRAGMENT
:
4336 if (dst_reg
->index
>= FRAG_RESULT_COLOR
) {
4337 dst
= ureg_saturate(dst
);
4343 if (dst_reg
->reladdr
!= NULL
) {
4344 assert(dst_reg
->file
!= PROGRAM_TEMPORARY
);
4345 dst
= ureg_dst_indirect(dst
, ureg_src(t
->address
[0]));
4352 * Create a TGSI ureg_src register from an st_src_reg.
4354 static struct ureg_src
4355 translate_src(struct st_translate
*t
, const st_src_reg
*src_reg
)
4357 struct ureg_src src
= src_register(t
, src_reg
->file
, src_reg
->index
, src_reg
->index2D
);
4359 if (t
->procType
== TGSI_PROCESSOR_GEOMETRY
&& src_reg
->has_index2
) {
4360 src
= src_register(t
, src_reg
->file
, src_reg
->index
, src_reg
->index2D
);
4361 if (src_reg
->reladdr2
)
4362 src
= ureg_src_dimension_indirect(src
, ureg_src(t
->address
[1]),
4365 src
= ureg_src_dimension(src
, src_reg
->index2D
);
4368 src
= ureg_swizzle(src
,
4369 GET_SWZ(src_reg
->swizzle
, 0) & 0x3,
4370 GET_SWZ(src_reg
->swizzle
, 1) & 0x3,
4371 GET_SWZ(src_reg
->swizzle
, 2) & 0x3,
4372 GET_SWZ(src_reg
->swizzle
, 3) & 0x3);
4374 if ((src_reg
->negate
& 0xf) == NEGATE_XYZW
)
4375 src
= ureg_negate(src
);
4377 if (src_reg
->reladdr
!= NULL
) {
4378 assert(src_reg
->file
!= PROGRAM_TEMPORARY
);
4379 src
= ureg_src_indirect(src
, ureg_src(t
->address
[0]));
4385 static struct tgsi_texture_offset
4386 translate_tex_offset(struct st_translate
*t
,
4387 const struct tgsi_texture_offset
*in_offset
)
4389 struct tgsi_texture_offset offset
;
4390 struct ureg_src imm_src
;
4392 assert(in_offset
->File
== PROGRAM_IMMEDIATE
);
4393 imm_src
= t
->immediates
[in_offset
->Index
];
4395 offset
.File
= imm_src
.File
;
4396 offset
.Index
= imm_src
.Index
;
4397 offset
.SwizzleX
= imm_src
.SwizzleX
;
4398 offset
.SwizzleY
= imm_src
.SwizzleY
;
4399 offset
.SwizzleZ
= imm_src
.SwizzleZ
;
4400 offset
.File
= TGSI_FILE_IMMEDIATE
;
4407 compile_tgsi_instruction(struct st_translate
*t
,
4408 const glsl_to_tgsi_instruction
*inst
,
4409 bool clamp_dst_color_output
)
4411 struct ureg_program
*ureg
= t
->ureg
;
4413 struct ureg_dst dst
[1];
4414 struct ureg_src src
[4];
4415 struct tgsi_texture_offset texoffsets
[MAX_GLSL_TEXTURE_OFFSET
];
4419 unsigned tex_target
;
4421 num_dst
= num_inst_dst_regs(inst
->op
);
4422 num_src
= num_inst_src_regs(inst
->op
);
4425 dst
[0] = translate_dst(t
,
4428 clamp_dst_color_output
);
4430 for (i
= 0; i
< num_src
; i
++)
4431 src
[i
] = translate_src(t
, &inst
->src
[i
]);
4434 case TGSI_OPCODE_BGNLOOP
:
4435 case TGSI_OPCODE_CAL
:
4436 case TGSI_OPCODE_ELSE
:
4437 case TGSI_OPCODE_ENDLOOP
:
4438 case TGSI_OPCODE_IF
:
4439 case TGSI_OPCODE_UIF
:
4440 assert(num_dst
== 0);
4441 ureg_label_insn(ureg
,
4445 inst
->op
== TGSI_OPCODE_CAL
? inst
->function
->sig_id
: 0));
4448 case TGSI_OPCODE_TEX
:
4449 case TGSI_OPCODE_TXB
:
4450 case TGSI_OPCODE_TXD
:
4451 case TGSI_OPCODE_TXL
:
4452 case TGSI_OPCODE_TXP
:
4453 case TGSI_OPCODE_TXQ
:
4454 case TGSI_OPCODE_TXF
:
4455 case TGSI_OPCODE_TEX2
:
4456 case TGSI_OPCODE_TXB2
:
4457 case TGSI_OPCODE_TXL2
:
4458 src
[num_src
++] = t
->samplers
[inst
->sampler
];
4459 for (i
= 0; i
< inst
->tex_offset_num_offset
; i
++) {
4460 texoffsets
[i
] = translate_tex_offset(t
, &inst
->tex_offsets
[i
]);
4462 tex_target
= st_translate_texture_target(inst
->tex_target
, inst
->tex_shadow
);
4468 texoffsets
, inst
->tex_offset_num_offset
,
4472 case TGSI_OPCODE_SCS
:
4473 dst
[0] = ureg_writemask(dst
[0], TGSI_WRITEMASK_XY
);
4474 ureg_insn(ureg
, inst
->op
, dst
, num_dst
, src
, num_src
);
4487 * Emit the TGSI instructions for inverting and adjusting WPOS.
4488 * This code is unavoidable because it also depends on whether
4489 * a FBO is bound (STATE_FB_WPOS_Y_TRANSFORM).
4492 emit_wpos_adjustment( struct st_translate
*t
,
4493 const struct gl_program
*program
,
4495 GLfloat adjX
, GLfloat adjY
[2])
4497 struct ureg_program
*ureg
= t
->ureg
;
4499 /* Fragment program uses fragment position input.
4500 * Need to replace instances of INPUT[WPOS] with temp T
4501 * where T = INPUT[WPOS] by y is inverted.
4503 static const gl_state_index wposTransformState
[STATE_LENGTH
]
4504 = { STATE_INTERNAL
, STATE_FB_WPOS_Y_TRANSFORM
,
4505 (gl_state_index
)0, (gl_state_index
)0, (gl_state_index
)0 };
4507 /* XXX: note we are modifying the incoming shader here! Need to
4508 * do this before emitting the constant decls below, or this
4511 unsigned wposTransConst
= _mesa_add_state_reference(program
->Parameters
,
4512 wposTransformState
);
4514 struct ureg_src wpostrans
= ureg_DECL_constant( ureg
, wposTransConst
);
4515 struct ureg_dst wpos_temp
= ureg_DECL_temporary( ureg
);
4516 struct ureg_src wpos_input
= t
->inputs
[t
->inputMapping
[VARYING_SLOT_POS
]];
4518 /* First, apply the coordinate shift: */
4519 if (adjX
|| adjY
[0] || adjY
[1]) {
4520 if (adjY
[0] != adjY
[1]) {
4521 /* Adjust the y coordinate by adjY[1] or adjY[0] respectively
4522 * depending on whether inversion is actually going to be applied
4523 * or not, which is determined by testing against the inversion
4524 * state variable used below, which will be either +1 or -1.
4526 struct ureg_dst adj_temp
= ureg_DECL_local_temporary(ureg
);
4528 ureg_CMP(ureg
, adj_temp
,
4529 ureg_scalar(wpostrans
, invert
? 2 : 0),
4530 ureg_imm4f(ureg
, adjX
, adjY
[0], 0.0f
, 0.0f
),
4531 ureg_imm4f(ureg
, adjX
, adjY
[1], 0.0f
, 0.0f
));
4532 ureg_ADD(ureg
, wpos_temp
, wpos_input
, ureg_src(adj_temp
));
4534 ureg_ADD(ureg
, wpos_temp
, wpos_input
,
4535 ureg_imm4f(ureg
, adjX
, adjY
[0], 0.0f
, 0.0f
));
4537 wpos_input
= ureg_src(wpos_temp
);
4539 /* MOV wpos_temp, input[wpos]
4541 ureg_MOV( ureg
, wpos_temp
, wpos_input
);
4544 /* Now the conditional y flip: STATE_FB_WPOS_Y_TRANSFORM.xy/zw will be
4545 * inversion/identity, or the other way around if we're drawing to an FBO.
4548 /* MAD wpos_temp.y, wpos_input, wpostrans.xxxx, wpostrans.yyyy
4551 ureg_writemask(wpos_temp
, TGSI_WRITEMASK_Y
),
4553 ureg_scalar(wpostrans
, 0),
4554 ureg_scalar(wpostrans
, 1));
4556 /* MAD wpos_temp.y, wpos_input, wpostrans.zzzz, wpostrans.wwww
4559 ureg_writemask(wpos_temp
, TGSI_WRITEMASK_Y
),
4561 ureg_scalar(wpostrans
, 2),
4562 ureg_scalar(wpostrans
, 3));
4565 /* Use wpos_temp as position input from here on:
4567 t
->inputs
[t
->inputMapping
[VARYING_SLOT_POS
]] = ureg_src(wpos_temp
);
4572 * Emit fragment position/ooordinate code.
4575 emit_wpos(struct st_context
*st
,
4576 struct st_translate
*t
,
4577 const struct gl_program
*program
,
4578 struct ureg_program
*ureg
)
4580 const struct gl_fragment_program
*fp
=
4581 (const struct gl_fragment_program
*) program
;
4582 struct pipe_screen
*pscreen
= st
->pipe
->screen
;
4583 GLfloat adjX
= 0.0f
;
4584 GLfloat adjY
[2] = { 0.0f
, 0.0f
};
4585 boolean invert
= FALSE
;
4587 /* Query the pixel center conventions supported by the pipe driver and set
4588 * adjX, adjY to help out if it cannot handle the requested one internally.
4590 * The bias of the y-coordinate depends on whether y-inversion takes place
4591 * (adjY[1]) or not (adjY[0]), which is in turn dependent on whether we are
4592 * drawing to an FBO (causes additional inversion), and whether the the pipe
4593 * driver origin and the requested origin differ (the latter condition is
4594 * stored in the 'invert' variable).
4596 * For height = 100 (i = integer, h = half-integer, l = lower, u = upper):
4598 * center shift only:
4603 * l,i -> u,i: ( 0.0 + 1.0) * -1 + 100 = 99
4604 * l,h -> u,h: ( 0.5 + 0.0) * -1 + 100 = 99.5
4605 * u,i -> l,i: (99.0 + 1.0) * -1 + 100 = 0
4606 * u,h -> l,h: (99.5 + 0.0) * -1 + 100 = 0.5
4608 * inversion and center shift:
4609 * l,i -> u,h: ( 0.0 + 0.5) * -1 + 100 = 99.5
4610 * l,h -> u,i: ( 0.5 + 0.5) * -1 + 100 = 99
4611 * u,i -> l,h: (99.0 + 0.5) * -1 + 100 = 0.5
4612 * u,h -> l,i: (99.5 + 0.5) * -1 + 100 = 0
4614 if (fp
->OriginUpperLeft
) {
4615 /* Fragment shader wants origin in upper-left */
4616 if (pscreen
->get_param(pscreen
, PIPE_CAP_TGSI_FS_COORD_ORIGIN_UPPER_LEFT
)) {
4617 /* the driver supports upper-left origin */
4619 else if (pscreen
->get_param(pscreen
, PIPE_CAP_TGSI_FS_COORD_ORIGIN_LOWER_LEFT
)) {
4620 /* the driver supports lower-left origin, need to invert Y */
4621 ureg_property_fs_coord_origin(ureg
, TGSI_FS_COORD_ORIGIN_LOWER_LEFT
);
4628 /* Fragment shader wants origin in lower-left */
4629 if (pscreen
->get_param(pscreen
, PIPE_CAP_TGSI_FS_COORD_ORIGIN_LOWER_LEFT
))
4630 /* the driver supports lower-left origin */
4631 ureg_property_fs_coord_origin(ureg
, TGSI_FS_COORD_ORIGIN_LOWER_LEFT
);
4632 else if (pscreen
->get_param(pscreen
, PIPE_CAP_TGSI_FS_COORD_ORIGIN_UPPER_LEFT
))
4633 /* the driver supports upper-left origin, need to invert Y */
4639 if (fp
->PixelCenterInteger
) {
4640 /* Fragment shader wants pixel center integer */
4641 if (pscreen
->get_param(pscreen
, PIPE_CAP_TGSI_FS_COORD_PIXEL_CENTER_INTEGER
)) {
4642 /* the driver supports pixel center integer */
4644 ureg_property_fs_coord_pixel_center(ureg
, TGSI_FS_COORD_PIXEL_CENTER_INTEGER
);
4646 else if (pscreen
->get_param(pscreen
, PIPE_CAP_TGSI_FS_COORD_PIXEL_CENTER_HALF_INTEGER
)) {
4647 /* the driver supports pixel center half integer, need to bias X,Y */
4656 /* Fragment shader wants pixel center half integer */
4657 if (pscreen
->get_param(pscreen
, PIPE_CAP_TGSI_FS_COORD_PIXEL_CENTER_HALF_INTEGER
)) {
4658 /* the driver supports pixel center half integer */
4660 else if (pscreen
->get_param(pscreen
, PIPE_CAP_TGSI_FS_COORD_PIXEL_CENTER_INTEGER
)) {
4661 /* the driver supports pixel center integer, need to bias X,Y */
4662 adjX
= adjY
[0] = adjY
[1] = 0.5f
;
4663 ureg_property_fs_coord_pixel_center(ureg
, TGSI_FS_COORD_PIXEL_CENTER_INTEGER
);
4669 /* we invert after adjustment so that we avoid the MOV to temporary,
4670 * and reuse the adjustment ADD instead */
4671 emit_wpos_adjustment(t
, program
, invert
, adjX
, adjY
);
4675 * OpenGL's fragment gl_FrontFace input is 1 for front-facing, 0 for back.
4676 * TGSI uses +1 for front, -1 for back.
4677 * This function converts the TGSI value to the GL value. Simply clamping/
4678 * saturating the value to [0,1] does the job.
4681 emit_face_var(struct st_translate
*t
)
4683 struct ureg_program
*ureg
= t
->ureg
;
4684 struct ureg_dst face_temp
= ureg_DECL_temporary(ureg
);
4685 struct ureg_src face_input
= t
->inputs
[t
->inputMapping
[VARYING_SLOT_FACE
]];
4687 /* MOV_SAT face_temp, input[face] */
4688 face_temp
= ureg_saturate(face_temp
);
4689 ureg_MOV(ureg
, face_temp
, face_input
);
4691 /* Use face_temp as face input from here on: */
4692 t
->inputs
[t
->inputMapping
[VARYING_SLOT_FACE
]] = ureg_src(face_temp
);
4696 emit_edgeflags(struct st_translate
*t
)
4698 struct ureg_program
*ureg
= t
->ureg
;
4699 struct ureg_dst edge_dst
= t
->outputs
[t
->outputMapping
[VARYING_SLOT_EDGE
]];
4700 struct ureg_src edge_src
= t
->inputs
[t
->inputMapping
[VERT_ATTRIB_EDGEFLAG
]];
4702 ureg_MOV(ureg
, edge_dst
, edge_src
);
4706 * Translate intermediate IR (glsl_to_tgsi_instruction) to TGSI format.
4707 * \param program the program to translate
4708 * \param numInputs number of input registers used
4709 * \param inputMapping maps Mesa fragment program inputs to TGSI generic
4711 * \param inputSemanticName the TGSI_SEMANTIC flag for each input
4712 * \param inputSemanticIndex the semantic index (ex: which texcoord) for
4714 * \param interpMode the TGSI_INTERPOLATE_LINEAR/PERSP mode for each input
4715 * \param numOutputs number of output registers used
4716 * \param outputMapping maps Mesa fragment program outputs to TGSI
4718 * \param outputSemanticName the TGSI_SEMANTIC flag for each output
4719 * \param outputSemanticIndex the semantic index (ex: which texcoord) for
4722 * \return PIPE_OK or PIPE_ERROR_OUT_OF_MEMORY
4724 extern "C" enum pipe_error
4725 st_translate_program(
4726 struct gl_context
*ctx
,
4728 struct ureg_program
*ureg
,
4729 glsl_to_tgsi_visitor
*program
,
4730 const struct gl_program
*proginfo
,
4732 const GLuint inputMapping
[],
4733 const ubyte inputSemanticName
[],
4734 const ubyte inputSemanticIndex
[],
4735 const GLuint interpMode
[],
4736 const GLboolean is_centroid
[],
4738 const GLuint outputMapping
[],
4739 const ubyte outputSemanticName
[],
4740 const ubyte outputSemanticIndex
[],
4741 boolean passthrough_edgeflags
,
4742 boolean clamp_color
)
4744 struct st_translate
*t
;
4746 enum pipe_error ret
= PIPE_OK
;
4748 assert(numInputs
<= Elements(t
->inputs
));
4749 assert(numOutputs
<= Elements(t
->outputs
));
4751 t
= CALLOC_STRUCT(st_translate
);
4753 ret
= PIPE_ERROR_OUT_OF_MEMORY
;
4757 memset(t
, 0, sizeof *t
);
4759 t
->procType
= procType
;
4760 t
->inputMapping
= inputMapping
;
4761 t
->outputMapping
= outputMapping
;
4764 if (program
->shader_program
) {
4765 for (i
= 0; i
< program
->shader_program
->NumUserUniformStorage
; i
++) {
4766 struct gl_uniform_storage
*const storage
=
4767 &program
->shader_program
->UniformStorage
[i
];
4769 _mesa_uniform_detach_all_driver_storage(storage
);
4774 * Declare input attributes.
4776 if (procType
== TGSI_PROCESSOR_FRAGMENT
) {
4777 for (i
= 0; i
< numInputs
; i
++) {
4778 t
->inputs
[i
] = ureg_DECL_fs_input_cyl_centroid(ureg
,
4779 inputSemanticName
[i
],
4780 inputSemanticIndex
[i
],
4785 if (proginfo
->InputsRead
& VARYING_BIT_POS
) {
4786 /* Must do this after setting up t->inputs, and before
4787 * emitting constant references, below:
4789 emit_wpos(st_context(ctx
), t
, proginfo
, ureg
);
4792 if (proginfo
->InputsRead
& VARYING_BIT_FACE
)
4796 * Declare output attributes.
4798 for (i
= 0; i
< numOutputs
; i
++) {
4799 switch (outputSemanticName
[i
]) {
4800 case TGSI_SEMANTIC_POSITION
:
4801 t
->outputs
[i
] = ureg_DECL_output(ureg
,
4802 TGSI_SEMANTIC_POSITION
, /* Z/Depth */
4803 outputSemanticIndex
[i
]);
4804 t
->outputs
[i
] = ureg_writemask(t
->outputs
[i
], TGSI_WRITEMASK_Z
);
4806 case TGSI_SEMANTIC_STENCIL
:
4807 t
->outputs
[i
] = ureg_DECL_output(ureg
,
4808 TGSI_SEMANTIC_STENCIL
, /* Stencil */
4809 outputSemanticIndex
[i
]);
4810 t
->outputs
[i
] = ureg_writemask(t
->outputs
[i
], TGSI_WRITEMASK_Y
);
4812 case TGSI_SEMANTIC_COLOR
:
4813 t
->outputs
[i
] = ureg_DECL_output(ureg
,
4814 TGSI_SEMANTIC_COLOR
,
4815 outputSemanticIndex
[i
]);
4818 assert(!"fragment shader outputs must be POSITION/STENCIL/COLOR");
4819 ret
= PIPE_ERROR_BAD_INPUT
;
4824 else if (procType
== TGSI_PROCESSOR_GEOMETRY
) {
4825 for (i
= 0; i
< numInputs
; i
++) {
4826 t
->inputs
[i
] = ureg_DECL_gs_input(ureg
,
4828 inputSemanticName
[i
],
4829 inputSemanticIndex
[i
]);
4832 for (i
= 0; i
< numOutputs
; i
++) {
4833 t
->outputs
[i
] = ureg_DECL_output(ureg
,
4834 outputSemanticName
[i
],
4835 outputSemanticIndex
[i
]);
4839 assert(procType
== TGSI_PROCESSOR_VERTEX
);
4841 for (i
= 0; i
< numInputs
; i
++) {
4842 t
->inputs
[i
] = ureg_DECL_vs_input(ureg
, i
);
4845 for (i
= 0; i
< numOutputs
; i
++) {
4846 t
->outputs
[i
] = ureg_DECL_output(ureg
,
4847 outputSemanticName
[i
],
4848 outputSemanticIndex
[i
]);
4849 if (outputSemanticName
[i
] == TGSI_SEMANTIC_FOG
) {
4850 /* force register to contain a fog coordinate in the form (F, 0, 0, 1). */
4852 ureg_writemask(t
->outputs
[i
], TGSI_WRITEMASK_YZW
),
4853 ureg_imm4f(ureg
, 0.0f
, 0.0f
, 0.0f
, 1.0f
));
4854 t
->outputs
[i
] = ureg_writemask(t
->outputs
[i
], TGSI_WRITEMASK_X
);
4857 if (passthrough_edgeflags
)
4861 /* Declare address register.
4863 if (program
->num_address_regs
> 0) {
4864 assert(program
->num_address_regs
<= 2);
4865 t
->address
[0] = ureg_DECL_address(ureg
);
4866 if (program
->num_address_regs
== 2)
4867 t
->address
[1] = ureg_DECL_address(ureg
);
4870 /* Declare misc input registers
4873 GLbitfield sysInputs
= proginfo
->SystemValuesRead
;
4874 unsigned numSys
= 0;
4875 for (i
= 0; sysInputs
; i
++) {
4876 if (sysInputs
& (1 << i
)) {
4877 unsigned semName
= mesa_sysval_to_semantic
[i
];
4878 t
->systemValues
[i
] = ureg_DECL_system_value(ureg
, numSys
, semName
, 0);
4879 if (semName
== TGSI_SEMANTIC_INSTANCEID
||
4880 semName
== TGSI_SEMANTIC_VERTEXID
) {
4881 /* From Gallium perspective, these system values are always
4882 * integer, and require native integer support. However, if
4883 * native integer is supported on the vertex stage but not the
4884 * pixel stage (e.g, i915g + draw), Mesa will generate IR that
4885 * assumes these system values are floats. To resolve the
4886 * inconsistency, we insert a U2F.
4888 struct st_context
*st
= st_context(ctx
);
4889 struct pipe_screen
*pscreen
= st
->pipe
->screen
;
4890 assert(procType
== TGSI_PROCESSOR_VERTEX
);
4891 assert(pscreen
->get_shader_param(pscreen
, PIPE_SHADER_VERTEX
, PIPE_SHADER_CAP_INTEGERS
));
4892 if (!ctx
->Const
.NativeIntegers
) {
4893 struct ureg_dst temp
= ureg_DECL_local_temporary(t
->ureg
);
4894 ureg_U2F( t
->ureg
, ureg_writemask(temp
, TGSI_WRITEMASK_X
), t
->systemValues
[i
]);
4895 t
->systemValues
[i
] = ureg_scalar(ureg_src(temp
), 0);
4899 sysInputs
&= ~(1 << i
);
4904 /* Copy over array sizes
4906 memcpy(t
->array_sizes
, program
->array_sizes
, sizeof(unsigned) * program
->next_array
);
4908 /* Emit constants and uniforms. TGSI uses a single index space for these,
4909 * so we put all the translated regs in t->constants.
4911 if (proginfo
->Parameters
) {
4912 t
->constants
= (struct ureg_src
*)
4913 calloc(proginfo
->Parameters
->NumParameters
, sizeof(t
->constants
[0]));
4914 if (t
->constants
== NULL
) {
4915 ret
= PIPE_ERROR_OUT_OF_MEMORY
;
4919 for (i
= 0; i
< proginfo
->Parameters
->NumParameters
; i
++) {
4920 switch (proginfo
->Parameters
->Parameters
[i
].Type
) {
4921 case PROGRAM_STATE_VAR
:
4922 case PROGRAM_UNIFORM
:
4923 t
->constants
[i
] = ureg_DECL_constant(ureg
, i
);
4926 /* Emit immediates for PROGRAM_CONSTANT only when there's no indirect
4927 * addressing of the const buffer.
4928 * FIXME: Be smarter and recognize param arrays:
4929 * indirect addressing is only valid within the referenced
4932 case PROGRAM_CONSTANT
:
4933 if (program
->indirect_addr_consts
)
4934 t
->constants
[i
] = ureg_DECL_constant(ureg
, i
);
4936 t
->constants
[i
] = emit_immediate(t
,
4937 proginfo
->Parameters
->ParameterValues
[i
],
4938 proginfo
->Parameters
->Parameters
[i
].DataType
,
4947 if (program
->shader_program
) {
4948 unsigned num_ubos
= program
->shader_program
->NumUniformBlocks
;
4950 for (i
= 0; i
< num_ubos
; i
++) {
4951 ureg_DECL_constant2D(t
->ureg
, 0, program
->shader_program
->UniformBlocks
[i
].UniformBufferSize
/ 4, i
+ 1);
4955 /* Emit immediate values.
4957 t
->immediates
= (struct ureg_src
*)
4958 calloc(program
->num_immediates
, sizeof(struct ureg_src
));
4959 if (t
->immediates
== NULL
) {
4960 ret
= PIPE_ERROR_OUT_OF_MEMORY
;
4964 foreach_iter(exec_list_iterator
, iter
, program
->immediates
) {
4965 immediate_storage
*imm
= (immediate_storage
*)iter
.get();
4966 assert(i
< program
->num_immediates
);
4967 t
->immediates
[i
++] = emit_immediate(t
, imm
->values
, imm
->type
, imm
->size
);
4969 assert(i
== program
->num_immediates
);
4971 /* texture samplers */
4972 for (i
= 0; i
< ctx
->Const
.FragmentProgram
.MaxTextureImageUnits
; i
++) {
4973 if (program
->samplers_used
& (1 << i
)) {
4974 t
->samplers
[i
] = ureg_DECL_sampler(ureg
, i
);
4978 /* Emit each instruction in turn:
4980 foreach_iter(exec_list_iterator
, iter
, program
->instructions
) {
4981 set_insn_start(t
, ureg_get_instruction_number(ureg
));
4982 compile_tgsi_instruction(t
, (glsl_to_tgsi_instruction
*)iter
.get(),
4986 /* Fix up all emitted labels:
4988 for (i
= 0; i
< t
->labels_count
; i
++) {
4989 ureg_fixup_label(ureg
, t
->labels
[i
].token
,
4990 t
->insn
[t
->labels
[i
].branch_target
]);
4993 if (program
->shader_program
) {
4994 /* This has to be done last. Any operation the can cause
4995 * prog->ParameterValues to get reallocated (e.g., anything that adds a
4996 * program constant) has to happen before creating this linkage.
4998 for (unsigned i
= 0; i
< MESA_SHADER_TYPES
; i
++) {
4999 if (program
->shader_program
->_LinkedShaders
[i
] == NULL
)
5002 _mesa_associate_uniform_storage(ctx
, program
->shader_program
,
5003 program
->shader_program
->_LinkedShaders
[i
]->Program
->Parameters
);
5012 free(t
->immediates
);
5015 debug_printf("%s: translate error flag set\n", __FUNCTION__
);
5023 /* ----------------------------- End TGSI code ------------------------------ */
5026 * Convert a shader's GLSL IR into a Mesa gl_program, although without
5027 * generating Mesa IR.
5029 static struct gl_program
*
5030 get_mesa_program(struct gl_context
*ctx
,
5031 struct gl_shader_program
*shader_program
,
5032 struct gl_shader
*shader
)
5034 glsl_to_tgsi_visitor
* v
;
5035 struct gl_program
*prog
;
5038 struct gl_shader_compiler_options
*options
=
5039 &ctx
->ShaderCompilerOptions
[_mesa_shader_type_to_index(shader
->Type
)];
5040 struct pipe_screen
*pscreen
= ctx
->st
->pipe
->screen
;
5043 switch (shader
->Type
) {
5044 case GL_VERTEX_SHADER
:
5045 target
= GL_VERTEX_PROGRAM_ARB
;
5046 ptarget
= PIPE_SHADER_VERTEX
;
5048 case GL_FRAGMENT_SHADER
:
5049 target
= GL_FRAGMENT_PROGRAM_ARB
;
5050 ptarget
= PIPE_SHADER_FRAGMENT
;
5052 case GL_GEOMETRY_SHADER
:
5053 target
= GL_GEOMETRY_PROGRAM_NV
;
5054 ptarget
= PIPE_SHADER_GEOMETRY
;
5057 assert(!"should not be reached");
5061 validate_ir_tree(shader
->ir
);
5063 prog
= ctx
->Driver
.NewProgram(ctx
, target
, shader_program
->Name
);
5066 prog
->Parameters
= _mesa_new_parameter_list();
5067 v
= new glsl_to_tgsi_visitor();
5070 v
->shader_program
= shader_program
;
5071 v
->options
= options
;
5072 v
->glsl_version
= ctx
->Const
.GLSLVersion
;
5073 v
->native_integers
= ctx
->Const
.NativeIntegers
;
5075 v
->have_sqrt
= pscreen
->get_shader_param(pscreen
, ptarget
,
5076 PIPE_SHADER_CAP_TGSI_SQRT_SUPPORTED
);
5078 _mesa_generate_parameters_list_for_uniforms(shader_program
, shader
,
5081 /* Remove reads from output registers. */
5082 lower_output_reads(shader
->ir
);
5084 /* Emit intermediate IR for main(). */
5085 visit_exec_list(shader
->ir
, v
);
5087 /* Now emit bodies for any functions that were used. */
5089 progress
= GL_FALSE
;
5091 foreach_iter(exec_list_iterator
, iter
, v
->function_signatures
) {
5092 function_entry
*entry
= (function_entry
*)iter
.get();
5094 if (!entry
->bgn_inst
) {
5095 v
->current_function
= entry
;
5097 entry
->bgn_inst
= v
->emit(NULL
, TGSI_OPCODE_BGNSUB
);
5098 entry
->bgn_inst
->function
= entry
;
5100 visit_exec_list(&entry
->sig
->body
, v
);
5102 glsl_to_tgsi_instruction
*last
;
5103 last
= (glsl_to_tgsi_instruction
*)v
->instructions
.get_tail();
5104 if (last
->op
!= TGSI_OPCODE_RET
)
5105 v
->emit(NULL
, TGSI_OPCODE_RET
);
5107 glsl_to_tgsi_instruction
*end
;
5108 end
= v
->emit(NULL
, TGSI_OPCODE_ENDSUB
);
5109 end
->function
= entry
;
5117 /* Print out some information (for debugging purposes) used by the
5118 * optimization passes. */
5119 for (i
=0; i
< v
->next_temp
; i
++) {
5120 int fr
= v
->get_first_temp_read(i
);
5121 int fw
= v
->get_first_temp_write(i
);
5122 int lr
= v
->get_last_temp_read(i
);
5123 int lw
= v
->get_last_temp_write(i
);
5125 printf("Temp %d: FR=%3d FW=%3d LR=%3d LW=%3d\n", i
, fr
, fw
, lr
, lw
);
5130 /* Perform optimizations on the instructions in the glsl_to_tgsi_visitor. */
5132 v
->copy_propagate();
5133 while (v
->eliminate_dead_code_advanced());
5135 v
->eliminate_dead_code();
5136 v
->merge_registers();
5137 v
->renumber_registers();
5139 /* Write the END instruction. */
5140 v
->emit(NULL
, TGSI_OPCODE_END
);
5142 if (ctx
->Shader
.Flags
& GLSL_DUMP
) {
5144 printf("GLSL IR for linked %s program %d:\n",
5145 _mesa_glsl_shader_target_name(shader
->Type
),
5146 shader_program
->Name
);
5147 _mesa_print_ir(shader
->ir
, NULL
);
5153 prog
->Instructions
= NULL
;
5154 prog
->NumInstructions
= 0;
5156 do_set_program_inouts(shader
->ir
, prog
, shader
->Type
);
5157 count_resources(v
, prog
);
5159 _mesa_reference_program(ctx
, &shader
->Program
, prog
);
5161 /* This has to be done last. Any operation the can cause
5162 * prog->ParameterValues to get reallocated (e.g., anything that adds a
5163 * program constant) has to happen before creating this linkage.
5165 _mesa_associate_uniform_storage(ctx
, shader_program
, prog
->Parameters
);
5166 if (!shader_program
->LinkStatus
) {
5170 struct st_vertex_program
*stvp
;
5171 struct st_fragment_program
*stfp
;
5172 struct st_geometry_program
*stgp
;
5174 switch (shader
->Type
) {
5175 case GL_VERTEX_SHADER
:
5176 stvp
= (struct st_vertex_program
*)prog
;
5177 stvp
->glsl_to_tgsi
= v
;
5179 case GL_FRAGMENT_SHADER
:
5180 stfp
= (struct st_fragment_program
*)prog
;
5181 stfp
->glsl_to_tgsi
= v
;
5183 case GL_GEOMETRY_SHADER
:
5184 stgp
= (struct st_geometry_program
*)prog
;
5185 stgp
->glsl_to_tgsi
= v
;
5186 stgp
->Base
.InputType
= shader_program
->Geom
.InputType
;
5187 stgp
->Base
.OutputType
= shader_program
->Geom
.OutputType
;
5188 stgp
->Base
.VerticesOut
= shader_program
->Geom
.VerticesOut
;
5191 assert(!"should not be reached");
5201 st_new_shader(struct gl_context
*ctx
, GLuint name
, GLuint type
)
5203 struct gl_shader
*shader
;
5204 assert(type
== GL_FRAGMENT_SHADER
|| type
== GL_VERTEX_SHADER
||
5205 type
== GL_GEOMETRY_SHADER_ARB
);
5206 shader
= rzalloc(NULL
, struct gl_shader
);
5208 shader
->Type
= type
;
5209 shader
->Name
= name
;
5210 _mesa_init_shader(ctx
, shader
);
5215 struct gl_shader_program
*
5216 st_new_shader_program(struct gl_context
*ctx
, GLuint name
)
5218 struct gl_shader_program
*shProg
;
5219 shProg
= rzalloc(NULL
, struct gl_shader_program
);
5221 shProg
->Name
= name
;
5222 _mesa_init_shader_program(ctx
, shProg
);
5229 * Called via ctx->Driver.LinkShader()
5230 * This actually involves converting GLSL IR into an intermediate TGSI-like IR
5231 * with code lowering and other optimizations.
5234 st_link_shader(struct gl_context
*ctx
, struct gl_shader_program
*prog
)
5236 assert(prog
->LinkStatus
);
5238 for (unsigned i
= 0; i
< MESA_SHADER_TYPES
; i
++) {
5239 if (prog
->_LinkedShaders
[i
] == NULL
)
5243 exec_list
*ir
= prog
->_LinkedShaders
[i
]->ir
;
5244 const struct gl_shader_compiler_options
*options
=
5245 &ctx
->ShaderCompilerOptions
[_mesa_shader_type_to_index(prog
->_LinkedShaders
[i
]->Type
)];
5247 /* If there are forms of indirect addressing that the driver
5248 * cannot handle, perform the lowering pass.
5250 if (options
->EmitNoIndirectInput
|| options
->EmitNoIndirectOutput
||
5251 options
->EmitNoIndirectTemp
|| options
->EmitNoIndirectUniform
) {
5252 lower_variable_index_to_cond_assign(ir
,
5253 options
->EmitNoIndirectInput
,
5254 options
->EmitNoIndirectOutput
,
5255 options
->EmitNoIndirectTemp
,
5256 options
->EmitNoIndirectUniform
);
5259 if (ctx
->Extensions
.ARB_shading_language_packing
) {
5260 unsigned lower_inst
= LOWER_PACK_SNORM_2x16
|
5261 LOWER_UNPACK_SNORM_2x16
|
5262 LOWER_PACK_UNORM_2x16
|
5263 LOWER_UNPACK_UNORM_2x16
|
5264 LOWER_PACK_SNORM_4x8
|
5265 LOWER_UNPACK_SNORM_4x8
|
5266 LOWER_UNPACK_UNORM_4x8
|
5267 LOWER_PACK_UNORM_4x8
|
5268 LOWER_PACK_HALF_2x16
|
5269 LOWER_UNPACK_HALF_2x16
;
5271 lower_packing_builtins(ir
, lower_inst
);
5274 do_mat_op_to_vec(ir
);
5275 lower_instructions(ir
,
5280 (options
->EmitNoPow
? POW_TO_EXP2
: 0) |
5281 (!ctx
->Const
.NativeIntegers
? INT_DIV_TO_MUL_RCP
: 0));
5283 lower_ubo_reference(prog
->_LinkedShaders
[i
], ir
);
5284 do_vec_index_to_cond_assign(ir
);
5285 lower_vector_insert(ir
, true);
5286 lower_quadop_vector(ir
, false);
5288 if (options
->MaxIfDepth
== 0) {
5295 progress
= do_lower_jumps(ir
, true, true, options
->EmitNoMainReturn
, options
->EmitNoCont
, options
->EmitNoLoops
) || progress
;
5297 progress
= do_common_optimization(ir
, true, true,
5298 options
->MaxUnrollIterations
, options
)
5301 progress
= lower_if_to_cond_assign(ir
, options
->MaxIfDepth
) || progress
;
5305 validate_ir_tree(ir
);
5308 for (unsigned i
= 0; i
< MESA_SHADER_TYPES
; i
++) {
5309 struct gl_program
*linked_prog
;
5311 if (prog
->_LinkedShaders
[i
] == NULL
)
5314 linked_prog
= get_mesa_program(ctx
, prog
, prog
->_LinkedShaders
[i
]);
5317 _mesa_reference_program(ctx
, &prog
->_LinkedShaders
[i
]->Program
,
5319 if (!ctx
->Driver
.ProgramStringNotify(ctx
,
5320 _mesa_program_index_to_target(i
),
5322 _mesa_reference_program(ctx
, &prog
->_LinkedShaders
[i
]->Program
,
5324 _mesa_reference_program(ctx
, &linked_prog
, NULL
);
5329 _mesa_reference_program(ctx
, &linked_prog
, NULL
);
5336 st_translate_stream_output_info(glsl_to_tgsi_visitor
*glsl_to_tgsi
,
5337 const GLuint outputMapping
[],
5338 struct pipe_stream_output_info
*so
)
5341 struct gl_transform_feedback_info
*info
=
5342 &glsl_to_tgsi
->shader_program
->LinkedTransformFeedback
;
5344 for (i
= 0; i
< info
->NumOutputs
; i
++) {
5345 so
->output
[i
].register_index
=
5346 outputMapping
[info
->Outputs
[i
].OutputRegister
];
5347 so
->output
[i
].start_component
= info
->Outputs
[i
].ComponentOffset
;
5348 so
->output
[i
].num_components
= info
->Outputs
[i
].NumComponents
;
5349 so
->output
[i
].output_buffer
= info
->Outputs
[i
].OutputBuffer
;
5350 so
->output
[i
].dst_offset
= info
->Outputs
[i
].DstOffset
;
5353 for (i
= 0; i
< PIPE_MAX_SO_BUFFERS
; i
++) {
5354 so
->stride
[i
] = info
->BufferStride
[i
];
5356 so
->num_outputs
= info
->NumOutputs
;