2 * Copyright (C) 2005-2007 Brian Paul All Rights Reserved.
3 * Copyright (C) 2008 VMware, Inc. All Rights Reserved.
4 * Copyright © 2010 Intel Corporation
5 * Copyright © 2011 Bryan Cain
7 * Permission is hereby granted, free of charge, to any person obtaining a
8 * copy of this software and associated documentation files (the "Software"),
9 * to deal in the Software without restriction, including without limitation
10 * the rights to use, copy, modify, merge, publish, distribute, sublicense,
11 * and/or sell copies of the Software, and to permit persons to whom the
12 * Software is furnished to do so, subject to the following conditions:
14 * The above copyright notice and this permission notice (including the next
15 * paragraph) shall be included in all copies or substantial portions of the
18 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
19 * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
20 * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL
21 * THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
22 * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING
23 * FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER
24 * DEALINGS IN THE SOFTWARE.
28 * \file glsl_to_tgsi.cpp
30 * Translate GLSL IR to TGSI.
34 #include "main/compiler.h"
36 #include "ir_visitor.h"
37 #include "ir_print_visitor.h"
38 #include "ir_expression_flattening.h"
39 #include "glsl_types.h"
40 #include "glsl_parser_extras.h"
41 #include "../glsl/program.h"
42 #include "ir_optimization.h"
45 #include "main/mtypes.h"
46 #include "main/shaderobj.h"
47 #include "program/hash_table.h"
50 #include "main/shaderapi.h"
51 #include "main/uniforms.h"
52 #include "program/prog_instruction.h"
53 #include "program/prog_optimize.h"
54 #include "program/prog_print.h"
55 #include "program/program.h"
56 #include "program/prog_uniform.h"
57 #include "program/prog_parameter.h"
58 #include "program/sampler.h"
60 #include "pipe/p_compiler.h"
61 #include "pipe/p_context.h"
62 #include "pipe/p_screen.h"
63 #include "pipe/p_shader_tokens.h"
64 #include "pipe/p_state.h"
65 #include "util/u_math.h"
66 #include "tgsi/tgsi_ureg.h"
67 #include "tgsi/tgsi_info.h"
68 #include "st_context.h"
69 #include "st_program.h"
70 #include "st_glsl_to_tgsi.h"
71 #include "st_mesa_to_tgsi.h"
74 #define PROGRAM_IMMEDIATE PROGRAM_FILE_MAX
75 #define PROGRAM_ANY_CONST ((1 << PROGRAM_LOCAL_PARAM) | \
76 (1 << PROGRAM_ENV_PARAM) | \
77 (1 << PROGRAM_STATE_VAR) | \
78 (1 << PROGRAM_NAMED_PARAM) | \
79 (1 << PROGRAM_CONSTANT) | \
80 (1 << PROGRAM_UNIFORM))
83 * Maximum number of temporary registers.
85 * It is too big for stack allocated arrays -- it will cause stack overflow on
86 * Windows and likely Mac OS X.
88 #define MAX_TEMPS 4096
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
;
112 this->type
= type
? type
->base_type
: GLSL_TYPE_ERROR
;
113 this->reladdr
= NULL
;
116 st_src_reg(gl_register_file file
, int index
, int type
)
121 this->swizzle
= SWIZZLE_XYZW
;
123 this->reladdr
= NULL
;
128 this->type
= GLSL_TYPE_ERROR
;
129 this->file
= PROGRAM_UNDEFINED
;
133 this->reladdr
= NULL
;
136 explicit st_src_reg(st_dst_reg reg
);
138 gl_register_file file
; /**< PROGRAM_* from Mesa */
139 int index
; /**< temporary index, VERT_ATTRIB_*, FRAG_ATTRIB_*, etc. */
140 GLuint swizzle
; /**< SWIZZLE_XYZWONEZERO swizzles from Mesa. */
141 int negate
; /**< NEGATE_XYZW mask from mesa */
142 int type
; /** GLSL_TYPE_* from GLSL IR (enum glsl_base_type) */
143 /** Register index should be offset by the integer in this reg. */
149 st_dst_reg(gl_register_file file
, int writemask
, int type
)
153 this->writemask
= writemask
;
154 this->cond_mask
= COND_TR
;
155 this->reladdr
= NULL
;
161 this->type
= GLSL_TYPE_ERROR
;
162 this->file
= PROGRAM_UNDEFINED
;
165 this->cond_mask
= COND_TR
;
166 this->reladdr
= NULL
;
169 explicit st_dst_reg(st_src_reg reg
);
171 gl_register_file file
; /**< PROGRAM_* from Mesa */
172 int index
; /**< temporary index, VERT_ATTRIB_*, FRAG_ATTRIB_*, etc. */
173 int writemask
; /**< Bitfield of WRITEMASK_[XYZW] */
175 int type
; /** GLSL_TYPE_* from GLSL IR (enum glsl_base_type) */
176 /** Register index should be offset by the integer in this reg. */
180 st_src_reg::st_src_reg(st_dst_reg reg
)
182 this->type
= reg
.type
;
183 this->file
= reg
.file
;
184 this->index
= reg
.index
;
185 this->swizzle
= SWIZZLE_XYZW
;
187 this->reladdr
= reg
.reladdr
;
190 st_dst_reg::st_dst_reg(st_src_reg reg
)
192 this->type
= reg
.type
;
193 this->file
= reg
.file
;
194 this->index
= reg
.index
;
195 this->writemask
= WRITEMASK_XYZW
;
196 this->cond_mask
= COND_TR
;
197 this->reladdr
= reg
.reladdr
;
200 class glsl_to_tgsi_instruction
: public exec_node
{
202 /* Callers of this ralloc-based new need not call delete. It's
203 * easier to just ralloc_free 'ctx' (or any of its ancestors). */
204 static void* operator new(size_t size
, void *ctx
)
208 node
= rzalloc_size(ctx
, size
);
209 assert(node
!= NULL
);
217 /** Pointer to the ir source this tree came from for debugging */
219 GLboolean cond_update
;
221 int sampler
; /**< sampler index */
222 int tex_target
; /**< One of TEXTURE_*_INDEX */
223 GLboolean tex_shadow
;
224 struct tgsi_texture_offset tex_offsets
[MAX_GLSL_TEXTURE_OFFSET
];
225 unsigned tex_offset_num_offset
;
226 int dead_mask
; /**< Used in dead code elimination */
228 class function_entry
*function
; /* Set on TGSI_OPCODE_CAL or TGSI_OPCODE_BGNSUB */
231 class variable_storage
: public exec_node
{
233 variable_storage(ir_variable
*var
, gl_register_file file
, int index
)
234 : file(file
), index(index
), var(var
)
239 gl_register_file file
;
241 ir_variable
*var
; /* variable that maps to this, if any */
244 class immediate_storage
: public exec_node
{
246 immediate_storage(gl_constant_value
*values
, int size
, int type
)
248 memcpy(this->values
, values
, size
* sizeof(gl_constant_value
));
253 gl_constant_value values
[4];
254 int size
; /**< Number of components (1-4) */
255 int type
; /**< GL_FLOAT, GL_INT, GL_BOOL, or GL_UNSIGNED_INT */
258 class function_entry
: public exec_node
{
260 ir_function_signature
*sig
;
263 * identifier of this function signature used by the program.
265 * At the point that TGSI instructions for function calls are
266 * generated, we don't know the address of the first instruction of
267 * the function body. So we make the BranchTarget that is called a
268 * small integer and rewrite them during set_branchtargets().
273 * Pointer to first instruction of the function body.
275 * Set during function body emits after main() is processed.
277 glsl_to_tgsi_instruction
*bgn_inst
;
280 * Index of the first instruction of the function body in actual TGSI.
282 * Set after conversion from glsl_to_tgsi_instruction to TGSI.
286 /** Storage for the return value. */
287 st_src_reg return_reg
;
290 class glsl_to_tgsi_visitor
: public ir_visitor
{
292 glsl_to_tgsi_visitor();
293 ~glsl_to_tgsi_visitor();
295 function_entry
*current_function
;
297 struct gl_context
*ctx
;
298 struct gl_program
*prog
;
299 struct gl_shader_program
*shader_program
;
300 struct gl_shader_compiler_options
*options
;
304 int num_address_regs
;
306 bool indirect_addr_temps
;
307 bool indirect_addr_consts
;
310 bool native_integers
;
312 variable_storage
*find_variable_storage(ir_variable
*var
);
314 int add_constant(gl_register_file file
, gl_constant_value values
[4],
315 int size
, int datatype
, GLuint
*swizzle_out
);
317 function_entry
*get_function_signature(ir_function_signature
*sig
);
319 st_src_reg
get_temp(const glsl_type
*type
);
320 void reladdr_to_temp(ir_instruction
*ir
, st_src_reg
*reg
, int *num_reladdr
);
322 st_src_reg
st_src_reg_for_float(float val
);
323 st_src_reg
st_src_reg_for_int(int val
);
324 st_src_reg
st_src_reg_for_type(int type
, int val
);
327 * \name Visit methods
329 * As typical for the visitor pattern, there must be one \c visit method for
330 * each concrete subclass of \c ir_instruction. Virtual base classes within
331 * the hierarchy should not have \c visit methods.
334 virtual void visit(ir_variable
*);
335 virtual void visit(ir_loop
*);
336 virtual void visit(ir_loop_jump
*);
337 virtual void visit(ir_function_signature
*);
338 virtual void visit(ir_function
*);
339 virtual void visit(ir_expression
*);
340 virtual void visit(ir_swizzle
*);
341 virtual void visit(ir_dereference_variable
*);
342 virtual void visit(ir_dereference_array
*);
343 virtual void visit(ir_dereference_record
*);
344 virtual void visit(ir_assignment
*);
345 virtual void visit(ir_constant
*);
346 virtual void visit(ir_call
*);
347 virtual void visit(ir_return
*);
348 virtual void visit(ir_discard
*);
349 virtual void visit(ir_texture
*);
350 virtual void visit(ir_if
*);
355 /** List of variable_storage */
358 /** List of immediate_storage */
359 exec_list immediates
;
362 /** List of function_entry */
363 exec_list function_signatures
;
364 int next_signature_id
;
366 /** List of glsl_to_tgsi_instruction */
367 exec_list instructions
;
369 glsl_to_tgsi_instruction
*emit(ir_instruction
*ir
, unsigned op
);
371 glsl_to_tgsi_instruction
*emit(ir_instruction
*ir
, unsigned op
,
372 st_dst_reg dst
, st_src_reg src0
);
374 glsl_to_tgsi_instruction
*emit(ir_instruction
*ir
, unsigned op
,
375 st_dst_reg dst
, st_src_reg src0
, st_src_reg src1
);
377 glsl_to_tgsi_instruction
*emit(ir_instruction
*ir
, unsigned op
,
379 st_src_reg src0
, st_src_reg src1
, st_src_reg src2
);
381 unsigned get_opcode(ir_instruction
*ir
, unsigned op
,
383 st_src_reg src0
, st_src_reg src1
);
386 * Emit the correct dot-product instruction for the type of arguments
388 glsl_to_tgsi_instruction
*emit_dp(ir_instruction
*ir
,
394 void emit_scalar(ir_instruction
*ir
, unsigned op
,
395 st_dst_reg dst
, st_src_reg src0
);
397 void emit_scalar(ir_instruction
*ir
, unsigned op
,
398 st_dst_reg dst
, st_src_reg src0
, st_src_reg src1
);
400 void try_emit_float_set(ir_instruction
*ir
, unsigned op
, st_dst_reg dst
);
402 void emit_arl(ir_instruction
*ir
, st_dst_reg dst
, st_src_reg src0
);
404 void emit_scs(ir_instruction
*ir
, unsigned op
,
405 st_dst_reg dst
, const st_src_reg
&src
);
407 bool try_emit_mad(ir_expression
*ir
,
409 bool try_emit_mad_for_and_not(ir_expression
*ir
,
411 bool try_emit_sat(ir_expression
*ir
);
413 void emit_swz(ir_expression
*ir
);
415 bool process_move_condition(ir_rvalue
*ir
);
417 void remove_output_reads(gl_register_file type
);
418 void simplify_cmp(void);
420 void rename_temp_register(int index
, int new_index
);
421 int get_first_temp_read(int index
);
422 int get_first_temp_write(int index
);
423 int get_last_temp_read(int index
);
424 int get_last_temp_write(int index
);
426 void copy_propagate(void);
427 void eliminate_dead_code(void);
428 int eliminate_dead_code_advanced(void);
429 void merge_registers(void);
430 void renumber_registers(void);
435 static st_src_reg undef_src
= st_src_reg(PROGRAM_UNDEFINED
, 0, GLSL_TYPE_ERROR
);
437 static st_dst_reg undef_dst
= st_dst_reg(PROGRAM_UNDEFINED
, SWIZZLE_NOOP
, GLSL_TYPE_ERROR
);
439 static st_dst_reg address_reg
= st_dst_reg(PROGRAM_ADDRESS
, WRITEMASK_X
, GLSL_TYPE_FLOAT
);
442 fail_link(struct gl_shader_program
*prog
, const char *fmt
, ...) PRINTFLIKE(2, 3);
445 fail_link(struct gl_shader_program
*prog
, const char *fmt
, ...)
449 ralloc_vasprintf_append(&prog
->InfoLog
, fmt
, args
);
452 prog
->LinkStatus
= GL_FALSE
;
456 swizzle_for_size(int size
)
458 int size_swizzles
[4] = {
459 MAKE_SWIZZLE4(SWIZZLE_X
, SWIZZLE_X
, SWIZZLE_X
, SWIZZLE_X
),
460 MAKE_SWIZZLE4(SWIZZLE_X
, SWIZZLE_Y
, SWIZZLE_Y
, SWIZZLE_Y
),
461 MAKE_SWIZZLE4(SWIZZLE_X
, SWIZZLE_Y
, SWIZZLE_Z
, SWIZZLE_Z
),
462 MAKE_SWIZZLE4(SWIZZLE_X
, SWIZZLE_Y
, SWIZZLE_Z
, SWIZZLE_W
),
465 assert((size
>= 1) && (size
<= 4));
466 return size_swizzles
[size
- 1];
470 is_tex_instruction(unsigned opcode
)
472 const tgsi_opcode_info
* info
= tgsi_get_opcode_info(opcode
);
477 num_inst_dst_regs(unsigned opcode
)
479 const tgsi_opcode_info
* info
= tgsi_get_opcode_info(opcode
);
480 return info
->num_dst
;
484 num_inst_src_regs(unsigned opcode
)
486 const tgsi_opcode_info
* info
= tgsi_get_opcode_info(opcode
);
487 return info
->is_tex
? info
->num_src
- 1 : info
->num_src
;
490 glsl_to_tgsi_instruction
*
491 glsl_to_tgsi_visitor::emit(ir_instruction
*ir
, unsigned op
,
493 st_src_reg src0
, st_src_reg src1
, st_src_reg src2
)
495 glsl_to_tgsi_instruction
*inst
= new(mem_ctx
) glsl_to_tgsi_instruction();
496 int num_reladdr
= 0, i
;
498 op
= get_opcode(ir
, op
, dst
, src0
, src1
);
500 /* If we have to do relative addressing, we want to load the ARL
501 * reg directly for one of the regs, and preload the other reladdr
502 * sources into temps.
504 num_reladdr
+= dst
.reladdr
!= NULL
;
505 num_reladdr
+= src0
.reladdr
!= NULL
;
506 num_reladdr
+= src1
.reladdr
!= NULL
;
507 num_reladdr
+= src2
.reladdr
!= NULL
;
509 reladdr_to_temp(ir
, &src2
, &num_reladdr
);
510 reladdr_to_temp(ir
, &src1
, &num_reladdr
);
511 reladdr_to_temp(ir
, &src0
, &num_reladdr
);
514 emit_arl(ir
, address_reg
, *dst
.reladdr
);
517 assert(num_reladdr
== 0);
527 inst
->function
= NULL
;
529 if (op
== TGSI_OPCODE_ARL
|| op
== TGSI_OPCODE_UARL
)
530 this->num_address_regs
= 1;
532 /* Update indirect addressing status used by TGSI */
535 case PROGRAM_TEMPORARY
:
536 this->indirect_addr_temps
= true;
538 case PROGRAM_LOCAL_PARAM
:
539 case PROGRAM_ENV_PARAM
:
540 case PROGRAM_STATE_VAR
:
541 case PROGRAM_NAMED_PARAM
:
542 case PROGRAM_CONSTANT
:
543 case PROGRAM_UNIFORM
:
544 this->indirect_addr_consts
= true;
546 case PROGRAM_IMMEDIATE
:
547 assert(!"immediates should not have indirect addressing");
554 for (i
=0; i
<3; i
++) {
555 if(inst
->src
[i
].reladdr
) {
556 switch(inst
->src
[i
].file
) {
557 case PROGRAM_TEMPORARY
:
558 this->indirect_addr_temps
= true;
560 case PROGRAM_LOCAL_PARAM
:
561 case PROGRAM_ENV_PARAM
:
562 case PROGRAM_STATE_VAR
:
563 case PROGRAM_NAMED_PARAM
:
564 case PROGRAM_CONSTANT
:
565 case PROGRAM_UNIFORM
:
566 this->indirect_addr_consts
= true;
568 case PROGRAM_IMMEDIATE
:
569 assert(!"immediates should not have indirect addressing");
578 this->instructions
.push_tail(inst
);
581 try_emit_float_set(ir
, op
, dst
);
587 glsl_to_tgsi_instruction
*
588 glsl_to_tgsi_visitor::emit(ir_instruction
*ir
, unsigned op
,
589 st_dst_reg dst
, st_src_reg src0
, st_src_reg src1
)
591 return emit(ir
, op
, dst
, src0
, src1
, undef_src
);
594 glsl_to_tgsi_instruction
*
595 glsl_to_tgsi_visitor::emit(ir_instruction
*ir
, unsigned op
,
596 st_dst_reg dst
, st_src_reg src0
)
598 assert(dst
.writemask
!= 0);
599 return emit(ir
, op
, dst
, src0
, undef_src
, undef_src
);
602 glsl_to_tgsi_instruction
*
603 glsl_to_tgsi_visitor::emit(ir_instruction
*ir
, unsigned op
)
605 return emit(ir
, op
, undef_dst
, undef_src
, undef_src
, undef_src
);
609 * Emits the code to convert the result of float SET instructions to integers.
612 glsl_to_tgsi_visitor::try_emit_float_set(ir_instruction
*ir
, unsigned op
,
615 if ((op
== TGSI_OPCODE_SEQ
||
616 op
== TGSI_OPCODE_SNE
||
617 op
== TGSI_OPCODE_SGE
||
618 op
== TGSI_OPCODE_SLT
))
620 st_src_reg src
= st_src_reg(dst
);
621 src
.negate
= ~src
.negate
;
622 dst
.type
= GLSL_TYPE_FLOAT
;
623 emit(ir
, TGSI_OPCODE_F2I
, dst
, src
);
628 * Determines whether to use an integer, unsigned integer, or float opcode
629 * based on the operands and input opcode, then emits the result.
632 glsl_to_tgsi_visitor::get_opcode(ir_instruction
*ir
, unsigned op
,
634 st_src_reg src0
, st_src_reg src1
)
636 int type
= GLSL_TYPE_FLOAT
;
638 if (src0
.type
== GLSL_TYPE_FLOAT
|| src1
.type
== GLSL_TYPE_FLOAT
)
639 type
= GLSL_TYPE_FLOAT
;
640 else if (native_integers
)
641 type
= src0
.type
== GLSL_TYPE_BOOL
? GLSL_TYPE_INT
: src0
.type
;
643 #define case4(c, f, i, u) \
644 case TGSI_OPCODE_##c: \
645 if (type == GLSL_TYPE_INT) op = TGSI_OPCODE_##i; \
646 else if (type == GLSL_TYPE_UINT) op = TGSI_OPCODE_##u; \
647 else op = TGSI_OPCODE_##f; \
649 #define case3(f, i, u) case4(f, f, i, u)
650 #define case2fi(f, i) case4(f, f, i, i)
651 #define case2iu(i, u) case4(i, LAST, i, u)
657 case3(DIV
, IDIV
, UDIV
);
658 case3(MAX
, IMAX
, UMAX
);
659 case3(MIN
, IMIN
, UMIN
);
664 case3(SGE
, ISGE
, USGE
);
665 case3(SLT
, ISLT
, USLT
);
672 assert(op
!= TGSI_OPCODE_LAST
);
676 glsl_to_tgsi_instruction
*
677 glsl_to_tgsi_visitor::emit_dp(ir_instruction
*ir
,
678 st_dst_reg dst
, st_src_reg src0
, st_src_reg src1
,
681 static const unsigned dot_opcodes
[] = {
682 TGSI_OPCODE_DP2
, TGSI_OPCODE_DP3
, TGSI_OPCODE_DP4
685 return emit(ir
, dot_opcodes
[elements
- 2], dst
, src0
, src1
);
689 * Emits TGSI scalar opcodes to produce unique answers across channels.
691 * Some TGSI opcodes are scalar-only, like ARB_fp/vp. The src X
692 * channel determines the result across all channels. So to do a vec4
693 * of this operation, we want to emit a scalar per source channel used
694 * to produce dest channels.
697 glsl_to_tgsi_visitor::emit_scalar(ir_instruction
*ir
, unsigned op
,
699 st_src_reg orig_src0
, st_src_reg orig_src1
)
702 int done_mask
= ~dst
.writemask
;
704 /* TGSI RCP is a scalar operation splatting results to all channels,
705 * like ARB_fp/vp. So emit as many RCPs as necessary to cover our
708 for (i
= 0; i
< 4; i
++) {
709 GLuint this_mask
= (1 << i
);
710 glsl_to_tgsi_instruction
*inst
;
711 st_src_reg src0
= orig_src0
;
712 st_src_reg src1
= orig_src1
;
714 if (done_mask
& this_mask
)
717 GLuint src0_swiz
= GET_SWZ(src0
.swizzle
, i
);
718 GLuint src1_swiz
= GET_SWZ(src1
.swizzle
, i
);
719 for (j
= i
+ 1; j
< 4; j
++) {
720 /* If there is another enabled component in the destination that is
721 * derived from the same inputs, generate its value on this pass as
724 if (!(done_mask
& (1 << j
)) &&
725 GET_SWZ(src0
.swizzle
, j
) == src0_swiz
&&
726 GET_SWZ(src1
.swizzle
, j
) == src1_swiz
) {
727 this_mask
|= (1 << j
);
730 src0
.swizzle
= MAKE_SWIZZLE4(src0_swiz
, src0_swiz
,
731 src0_swiz
, src0_swiz
);
732 src1
.swizzle
= MAKE_SWIZZLE4(src1_swiz
, src1_swiz
,
733 src1_swiz
, src1_swiz
);
735 inst
= emit(ir
, op
, dst
, src0
, src1
);
736 inst
->dst
.writemask
= this_mask
;
737 done_mask
|= this_mask
;
742 glsl_to_tgsi_visitor::emit_scalar(ir_instruction
*ir
, unsigned op
,
743 st_dst_reg dst
, st_src_reg src0
)
745 st_src_reg undef
= undef_src
;
747 undef
.swizzle
= SWIZZLE_XXXX
;
749 emit_scalar(ir
, op
, dst
, src0
, undef
);
753 glsl_to_tgsi_visitor::emit_arl(ir_instruction
*ir
,
754 st_dst_reg dst
, st_src_reg src0
)
756 int op
= TGSI_OPCODE_ARL
;
758 if (src0
.type
== GLSL_TYPE_INT
|| src0
.type
== GLSL_TYPE_UINT
)
759 op
= TGSI_OPCODE_UARL
;
761 emit(NULL
, op
, dst
, src0
);
765 * Emit an TGSI_OPCODE_SCS instruction
767 * The \c SCS opcode functions a bit differently than the other TGSI opcodes.
768 * Instead of splatting its result across all four components of the
769 * destination, it writes one value to the \c x component and another value to
770 * the \c y component.
772 * \param ir IR instruction being processed
773 * \param op Either \c TGSI_OPCODE_SIN or \c TGSI_OPCODE_COS depending
774 * on which value is desired.
775 * \param dst Destination register
776 * \param src Source register
779 glsl_to_tgsi_visitor::emit_scs(ir_instruction
*ir
, unsigned op
,
781 const st_src_reg
&src
)
783 /* Vertex programs cannot use the SCS opcode.
785 if (this->prog
->Target
== GL_VERTEX_PROGRAM_ARB
) {
786 emit_scalar(ir
, op
, dst
, src
);
790 const unsigned component
= (op
== TGSI_OPCODE_SIN
) ? 0 : 1;
791 const unsigned scs_mask
= (1U << component
);
792 int done_mask
= ~dst
.writemask
;
795 assert(op
== TGSI_OPCODE_SIN
|| op
== TGSI_OPCODE_COS
);
797 /* If there are compnents in the destination that differ from the component
798 * that will be written by the SCS instrution, we'll need a temporary.
800 if (scs_mask
!= unsigned(dst
.writemask
)) {
801 tmp
= get_temp(glsl_type::vec4_type
);
804 for (unsigned i
= 0; i
< 4; i
++) {
805 unsigned this_mask
= (1U << i
);
806 st_src_reg src0
= src
;
808 if ((done_mask
& this_mask
) != 0)
811 /* The source swizzle specified which component of the source generates
812 * sine / cosine for the current component in the destination. The SCS
813 * instruction requires that this value be swizzle to the X component.
814 * Replace the current swizzle with a swizzle that puts the source in
817 unsigned src0_swiz
= GET_SWZ(src
.swizzle
, i
);
819 src0
.swizzle
= MAKE_SWIZZLE4(src0_swiz
, src0_swiz
,
820 src0_swiz
, src0_swiz
);
821 for (unsigned j
= i
+ 1; j
< 4; j
++) {
822 /* If there is another enabled component in the destination that is
823 * derived from the same inputs, generate its value on this pass as
826 if (!(done_mask
& (1 << j
)) &&
827 GET_SWZ(src0
.swizzle
, j
) == src0_swiz
) {
828 this_mask
|= (1 << j
);
832 if (this_mask
!= scs_mask
) {
833 glsl_to_tgsi_instruction
*inst
;
834 st_dst_reg tmp_dst
= st_dst_reg(tmp
);
836 /* Emit the SCS instruction.
838 inst
= emit(ir
, TGSI_OPCODE_SCS
, tmp_dst
, src0
);
839 inst
->dst
.writemask
= scs_mask
;
841 /* Move the result of the SCS instruction to the desired location in
844 tmp
.swizzle
= MAKE_SWIZZLE4(component
, component
,
845 component
, component
);
846 inst
= emit(ir
, TGSI_OPCODE_SCS
, dst
, tmp
);
847 inst
->dst
.writemask
= this_mask
;
849 /* Emit the SCS instruction to write directly to the destination.
851 glsl_to_tgsi_instruction
*inst
= emit(ir
, TGSI_OPCODE_SCS
, dst
, src0
);
852 inst
->dst
.writemask
= scs_mask
;
855 done_mask
|= this_mask
;
860 glsl_to_tgsi_visitor::add_constant(gl_register_file file
,
861 gl_constant_value values
[4], int size
, int datatype
,
864 if (file
== PROGRAM_CONSTANT
) {
865 return _mesa_add_typed_unnamed_constant(this->prog
->Parameters
, values
,
866 size
, datatype
, swizzle_out
);
869 immediate_storage
*entry
;
870 assert(file
== PROGRAM_IMMEDIATE
);
872 /* Search immediate storage to see if we already have an identical
873 * immediate that we can use instead of adding a duplicate entry.
875 foreach_iter(exec_list_iterator
, iter
, this->immediates
) {
876 entry
= (immediate_storage
*)iter
.get();
878 if (entry
->size
== size
&&
879 entry
->type
== datatype
&&
880 !memcmp(entry
->values
, values
, size
* sizeof(gl_constant_value
))) {
886 /* Add this immediate to the list. */
887 entry
= new(mem_ctx
) immediate_storage(values
, size
, datatype
);
888 this->immediates
.push_tail(entry
);
889 this->num_immediates
++;
895 glsl_to_tgsi_visitor::st_src_reg_for_float(float val
)
897 st_src_reg
src(PROGRAM_IMMEDIATE
, -1, GLSL_TYPE_FLOAT
);
898 union gl_constant_value uval
;
901 src
.index
= add_constant(src
.file
, &uval
, 1, GL_FLOAT
, &src
.swizzle
);
907 glsl_to_tgsi_visitor::st_src_reg_for_int(int val
)
909 st_src_reg
src(PROGRAM_IMMEDIATE
, -1, GLSL_TYPE_INT
);
910 union gl_constant_value uval
;
912 assert(native_integers
);
915 src
.index
= add_constant(src
.file
, &uval
, 1, GL_INT
, &src
.swizzle
);
921 glsl_to_tgsi_visitor::st_src_reg_for_type(int type
, int val
)
924 return type
== GLSL_TYPE_FLOAT
? st_src_reg_for_float(val
) :
925 st_src_reg_for_int(val
);
927 return st_src_reg_for_float(val
);
931 type_size(const struct glsl_type
*type
)
936 switch (type
->base_type
) {
939 case GLSL_TYPE_FLOAT
:
941 if (type
->is_matrix()) {
942 return type
->matrix_columns
;
944 /* Regardless of size of vector, it gets a vec4. This is bad
945 * packing for things like floats, but otherwise arrays become a
946 * mess. Hopefully a later pass over the code can pack scalars
947 * down if appropriate.
951 case GLSL_TYPE_ARRAY
:
952 assert(type
->length
> 0);
953 return type_size(type
->fields
.array
) * type
->length
;
954 case GLSL_TYPE_STRUCT
:
956 for (i
= 0; i
< type
->length
; i
++) {
957 size
+= type_size(type
->fields
.structure
[i
].type
);
960 case GLSL_TYPE_SAMPLER
:
961 /* Samplers take up one slot in UNIFORMS[], but they're baked in
972 * In the initial pass of codegen, we assign temporary numbers to
973 * intermediate results. (not SSA -- variable assignments will reuse
977 glsl_to_tgsi_visitor::get_temp(const glsl_type
*type
)
981 src
.type
= native_integers
? type
->base_type
: GLSL_TYPE_FLOAT
;
982 src
.file
= PROGRAM_TEMPORARY
;
983 src
.index
= next_temp
;
985 next_temp
+= type_size(type
);
987 if (type
->is_array() || type
->is_record()) {
988 src
.swizzle
= SWIZZLE_NOOP
;
990 src
.swizzle
= swizzle_for_size(type
->vector_elements
);
998 glsl_to_tgsi_visitor::find_variable_storage(ir_variable
*var
)
1001 variable_storage
*entry
;
1003 foreach_iter(exec_list_iterator
, iter
, this->variables
) {
1004 entry
= (variable_storage
*)iter
.get();
1006 if (entry
->var
== var
)
1014 glsl_to_tgsi_visitor::visit(ir_variable
*ir
)
1016 if (strcmp(ir
->name
, "gl_FragCoord") == 0) {
1017 struct gl_fragment_program
*fp
= (struct gl_fragment_program
*)this->prog
;
1019 fp
->OriginUpperLeft
= ir
->origin_upper_left
;
1020 fp
->PixelCenterInteger
= ir
->pixel_center_integer
;
1022 } else if (strcmp(ir
->name
, "gl_FragDepth") == 0) {
1023 struct gl_fragment_program
*fp
= (struct gl_fragment_program
*)this->prog
;
1024 switch (ir
->depth_layout
) {
1025 case ir_depth_layout_none
:
1026 fp
->FragDepthLayout
= FRAG_DEPTH_LAYOUT_NONE
;
1028 case ir_depth_layout_any
:
1029 fp
->FragDepthLayout
= FRAG_DEPTH_LAYOUT_ANY
;
1031 case ir_depth_layout_greater
:
1032 fp
->FragDepthLayout
= FRAG_DEPTH_LAYOUT_GREATER
;
1034 case ir_depth_layout_less
:
1035 fp
->FragDepthLayout
= FRAG_DEPTH_LAYOUT_LESS
;
1037 case ir_depth_layout_unchanged
:
1038 fp
->FragDepthLayout
= FRAG_DEPTH_LAYOUT_UNCHANGED
;
1046 if (ir
->mode
== ir_var_uniform
&& strncmp(ir
->name
, "gl_", 3) == 0) {
1048 const ir_state_slot
*const slots
= ir
->state_slots
;
1049 assert(ir
->state_slots
!= NULL
);
1051 /* Check if this statevar's setup in the STATE file exactly
1052 * matches how we'll want to reference it as a
1053 * struct/array/whatever. If not, then we need to move it into
1054 * temporary storage and hope that it'll get copy-propagated
1057 for (i
= 0; i
< ir
->num_state_slots
; i
++) {
1058 if (slots
[i
].swizzle
!= SWIZZLE_XYZW
) {
1063 variable_storage
*storage
;
1065 if (i
== ir
->num_state_slots
) {
1066 /* We'll set the index later. */
1067 storage
= new(mem_ctx
) variable_storage(ir
, PROGRAM_STATE_VAR
, -1);
1068 this->variables
.push_tail(storage
);
1072 /* The variable_storage constructor allocates slots based on the size
1073 * of the type. However, this had better match the number of state
1074 * elements that we're going to copy into the new temporary.
1076 assert((int) ir
->num_state_slots
== type_size(ir
->type
));
1078 storage
= new(mem_ctx
) variable_storage(ir
, PROGRAM_TEMPORARY
,
1080 this->variables
.push_tail(storage
);
1081 this->next_temp
+= type_size(ir
->type
);
1083 dst
= st_dst_reg(st_src_reg(PROGRAM_TEMPORARY
, storage
->index
,
1084 native_integers
? ir
->type
->base_type
: GLSL_TYPE_FLOAT
));
1088 for (unsigned int i
= 0; i
< ir
->num_state_slots
; i
++) {
1089 int index
= _mesa_add_state_reference(this->prog
->Parameters
,
1090 (gl_state_index
*)slots
[i
].tokens
);
1092 if (storage
->file
== PROGRAM_STATE_VAR
) {
1093 if (storage
->index
== -1) {
1094 storage
->index
= index
;
1096 assert(index
== storage
->index
+ (int)i
);
1099 st_src_reg
src(PROGRAM_STATE_VAR
, index
,
1100 native_integers
? ir
->type
->base_type
: GLSL_TYPE_FLOAT
);
1101 src
.swizzle
= slots
[i
].swizzle
;
1102 emit(ir
, TGSI_OPCODE_MOV
, dst
, src
);
1103 /* even a float takes up a whole vec4 reg in a struct/array. */
1108 if (storage
->file
== PROGRAM_TEMPORARY
&&
1109 dst
.index
!= storage
->index
+ (int) ir
->num_state_slots
) {
1110 fail_link(this->shader_program
,
1111 "failed to load builtin uniform `%s' (%d/%d regs loaded)\n",
1112 ir
->name
, dst
.index
- storage
->index
,
1113 type_size(ir
->type
));
1119 glsl_to_tgsi_visitor::visit(ir_loop
*ir
)
1121 ir_dereference_variable
*counter
= NULL
;
1123 if (ir
->counter
!= NULL
)
1124 counter
= new(ir
) ir_dereference_variable(ir
->counter
);
1126 if (ir
->from
!= NULL
) {
1127 assert(ir
->counter
!= NULL
);
1129 ir_assignment
*a
= new(ir
) ir_assignment(counter
, ir
->from
, NULL
);
1135 emit(NULL
, TGSI_OPCODE_BGNLOOP
);
1139 new(ir
) ir_expression(ir
->cmp
, glsl_type::bool_type
,
1141 ir_if
*if_stmt
= new(ir
) ir_if(e
);
1143 ir_loop_jump
*brk
= new(ir
) ir_loop_jump(ir_loop_jump::jump_break
);
1145 if_stmt
->then_instructions
.push_tail(brk
);
1147 if_stmt
->accept(this);
1154 visit_exec_list(&ir
->body_instructions
, this);
1156 if (ir
->increment
) {
1158 new(ir
) ir_expression(ir_binop_add
, counter
->type
,
1159 counter
, ir
->increment
);
1161 ir_assignment
*a
= new(ir
) ir_assignment(counter
, e
, NULL
);
1168 emit(NULL
, TGSI_OPCODE_ENDLOOP
);
1172 glsl_to_tgsi_visitor::visit(ir_loop_jump
*ir
)
1175 case ir_loop_jump::jump_break
:
1176 emit(NULL
, TGSI_OPCODE_BRK
);
1178 case ir_loop_jump::jump_continue
:
1179 emit(NULL
, TGSI_OPCODE_CONT
);
1186 glsl_to_tgsi_visitor::visit(ir_function_signature
*ir
)
1193 glsl_to_tgsi_visitor::visit(ir_function
*ir
)
1195 /* Ignore function bodies other than main() -- we shouldn't see calls to
1196 * them since they should all be inlined before we get to glsl_to_tgsi.
1198 if (strcmp(ir
->name
, "main") == 0) {
1199 const ir_function_signature
*sig
;
1202 sig
= ir
->matching_signature(&empty
);
1206 foreach_iter(exec_list_iterator
, iter
, sig
->body
) {
1207 ir_instruction
*ir
= (ir_instruction
*)iter
.get();
1215 glsl_to_tgsi_visitor::try_emit_mad(ir_expression
*ir
, int mul_operand
)
1217 int nonmul_operand
= 1 - mul_operand
;
1219 st_dst_reg result_dst
;
1221 ir_expression
*expr
= ir
->operands
[mul_operand
]->as_expression();
1222 if (!expr
|| expr
->operation
!= ir_binop_mul
)
1225 expr
->operands
[0]->accept(this);
1227 expr
->operands
[1]->accept(this);
1229 ir
->operands
[nonmul_operand
]->accept(this);
1232 this->result
= get_temp(ir
->type
);
1233 result_dst
= st_dst_reg(this->result
);
1234 result_dst
.writemask
= (1 << ir
->type
->vector_elements
) - 1;
1235 emit(ir
, TGSI_OPCODE_MAD
, result_dst
, a
, b
, c
);
1241 * Emit MAD(a, -b, a) instead of AND(a, NOT(b))
1243 * The logic values are 1.0 for true and 0.0 for false. Logical-and is
1244 * implemented using multiplication, and logical-or is implemented using
1245 * addition. Logical-not can be implemented as (true - x), or (1.0 - x).
1246 * As result, the logical expression (a & !b) can be rewritten as:
1250 * - (a * 1) - (a * b)
1254 * This final expression can be implemented as a single MAD(a, -b, a)
1258 glsl_to_tgsi_visitor::try_emit_mad_for_and_not(ir_expression
*ir
, int try_operand
)
1260 const int other_operand
= 1 - try_operand
;
1263 ir_expression
*expr
= ir
->operands
[try_operand
]->as_expression();
1264 if (!expr
|| expr
->operation
!= ir_unop_logic_not
)
1267 ir
->operands
[other_operand
]->accept(this);
1269 expr
->operands
[0]->accept(this);
1272 b
.negate
= ~b
.negate
;
1274 this->result
= get_temp(ir
->type
);
1275 emit(ir
, TGSI_OPCODE_MAD
, st_dst_reg(this->result
), a
, b
, a
);
1281 glsl_to_tgsi_visitor::try_emit_sat(ir_expression
*ir
)
1283 /* Saturates were only introduced to vertex programs in
1284 * NV_vertex_program3, so don't give them to drivers in the VP.
1286 if (this->prog
->Target
== GL_VERTEX_PROGRAM_ARB
)
1289 ir_rvalue
*sat_src
= ir
->as_rvalue_to_saturate();
1293 sat_src
->accept(this);
1294 st_src_reg src
= this->result
;
1296 /* If we generated an expression instruction into a temporary in
1297 * processing the saturate's operand, apply the saturate to that
1298 * instruction. Otherwise, generate a MOV to do the saturate.
1300 * Note that we have to be careful to only do this optimization if
1301 * the instruction in question was what generated src->result. For
1302 * example, ir_dereference_array might generate a MUL instruction
1303 * to create the reladdr, and return us a src reg using that
1304 * reladdr. That MUL result is not the value we're trying to
1307 ir_expression
*sat_src_expr
= sat_src
->as_expression();
1308 if (sat_src_expr
&& (sat_src_expr
->operation
== ir_binop_mul
||
1309 sat_src_expr
->operation
== ir_binop_add
||
1310 sat_src_expr
->operation
== ir_binop_dot
)) {
1311 glsl_to_tgsi_instruction
*new_inst
;
1312 new_inst
= (glsl_to_tgsi_instruction
*)this->instructions
.get_tail();
1313 new_inst
->saturate
= true;
1315 this->result
= get_temp(ir
->type
);
1316 st_dst_reg result_dst
= st_dst_reg(this->result
);
1317 result_dst
.writemask
= (1 << ir
->type
->vector_elements
) - 1;
1318 glsl_to_tgsi_instruction
*inst
;
1319 inst
= emit(ir
, TGSI_OPCODE_MOV
, result_dst
, src
);
1320 inst
->saturate
= true;
1327 glsl_to_tgsi_visitor::reladdr_to_temp(ir_instruction
*ir
,
1328 st_src_reg
*reg
, int *num_reladdr
)
1333 emit_arl(ir
, address_reg
, *reg
->reladdr
);
1335 if (*num_reladdr
!= 1) {
1336 st_src_reg temp
= get_temp(glsl_type::vec4_type
);
1338 emit(ir
, TGSI_OPCODE_MOV
, st_dst_reg(temp
), *reg
);
1346 glsl_to_tgsi_visitor::visit(ir_expression
*ir
)
1348 unsigned int operand
;
1349 st_src_reg op
[Elements(ir
->operands
)];
1350 st_src_reg result_src
;
1351 st_dst_reg result_dst
;
1353 /* Quick peephole: Emit MAD(a, b, c) instead of ADD(MUL(a, b), c)
1355 if (ir
->operation
== ir_binop_add
) {
1356 if (try_emit_mad(ir
, 1))
1358 if (try_emit_mad(ir
, 0))
1362 /* Quick peephole: Emit OPCODE_MAD(-a, -b, a) instead of AND(a, NOT(b))
1364 if (ir
->operation
== ir_binop_logic_and
) {
1365 if (try_emit_mad_for_and_not(ir
, 1))
1367 if (try_emit_mad_for_and_not(ir
, 0))
1371 if (try_emit_sat(ir
))
1374 if (ir
->operation
== ir_quadop_vector
)
1375 assert(!"ir_quadop_vector should have been lowered");
1377 for (operand
= 0; operand
< ir
->get_num_operands(); operand
++) {
1378 this->result
.file
= PROGRAM_UNDEFINED
;
1379 ir
->operands
[operand
]->accept(this);
1380 if (this->result
.file
== PROGRAM_UNDEFINED
) {
1382 printf("Failed to get tree for expression operand:\n");
1383 ir
->operands
[operand
]->accept(&v
);
1386 op
[operand
] = this->result
;
1388 /* Matrix expression operands should have been broken down to vector
1389 * operations already.
1391 assert(!ir
->operands
[operand
]->type
->is_matrix());
1394 int vector_elements
= ir
->operands
[0]->type
->vector_elements
;
1395 if (ir
->operands
[1]) {
1396 vector_elements
= MAX2(vector_elements
,
1397 ir
->operands
[1]->type
->vector_elements
);
1400 this->result
.file
= PROGRAM_UNDEFINED
;
1402 /* Storage for our result. Ideally for an assignment we'd be using
1403 * the actual storage for the result here, instead.
1405 result_src
= get_temp(ir
->type
);
1406 /* convenience for the emit functions below. */
1407 result_dst
= st_dst_reg(result_src
);
1408 /* Limit writes to the channels that will be used by result_src later.
1409 * This does limit this temp's use as a temporary for multi-instruction
1412 result_dst
.writemask
= (1 << ir
->type
->vector_elements
) - 1;
1414 switch (ir
->operation
) {
1415 case ir_unop_logic_not
:
1416 if (result_dst
.type
!= GLSL_TYPE_FLOAT
)
1417 emit(ir
, TGSI_OPCODE_NOT
, result_dst
, op
[0]);
1419 /* Previously 'SEQ dst, src, 0.0' was used for this. However, many
1420 * older GPUs implement SEQ using multiple instructions (i915 uses two
1421 * SGE instructions and a MUL instruction). Since our logic values are
1422 * 0.0 and 1.0, 1-x also implements !x.
1424 op
[0].negate
= ~op
[0].negate
;
1425 emit(ir
, TGSI_OPCODE_ADD
, result_dst
, op
[0], st_src_reg_for_float(1.0));
1429 assert(result_dst
.type
== GLSL_TYPE_FLOAT
|| result_dst
.type
== GLSL_TYPE_INT
);
1430 if (result_dst
.type
== GLSL_TYPE_INT
)
1431 emit(ir
, TGSI_OPCODE_INEG
, result_dst
, op
[0]);
1433 op
[0].negate
= ~op
[0].negate
;
1438 assert(result_dst
.type
== GLSL_TYPE_FLOAT
);
1439 emit(ir
, TGSI_OPCODE_ABS
, result_dst
, op
[0]);
1442 emit(ir
, TGSI_OPCODE_SSG
, result_dst
, op
[0]);
1445 emit_scalar(ir
, TGSI_OPCODE_RCP
, result_dst
, op
[0]);
1449 emit_scalar(ir
, TGSI_OPCODE_EX2
, result_dst
, op
[0]);
1453 assert(!"not reached: should be handled by ir_explog_to_explog2");
1456 emit_scalar(ir
, TGSI_OPCODE_LG2
, result_dst
, op
[0]);
1459 emit_scalar(ir
, TGSI_OPCODE_SIN
, result_dst
, op
[0]);
1462 emit_scalar(ir
, TGSI_OPCODE_COS
, result_dst
, op
[0]);
1464 case ir_unop_sin_reduced
:
1465 emit_scs(ir
, TGSI_OPCODE_SIN
, result_dst
, op
[0]);
1467 case ir_unop_cos_reduced
:
1468 emit_scs(ir
, TGSI_OPCODE_COS
, result_dst
, op
[0]);
1472 emit(ir
, TGSI_OPCODE_DDX
, result_dst
, op
[0]);
1475 op
[0].negate
= ~op
[0].negate
;
1476 emit(ir
, TGSI_OPCODE_DDY
, result_dst
, op
[0]);
1479 case ir_unop_noise
: {
1480 /* At some point, a motivated person could add a better
1481 * implementation of noise. Currently not even the nvidia
1482 * binary drivers do anything more than this. In any case, the
1483 * place to do this is in the GL state tracker, not the poor
1486 emit(ir
, TGSI_OPCODE_MOV
, result_dst
, st_src_reg_for_float(0.5));
1491 emit(ir
, TGSI_OPCODE_ADD
, result_dst
, op
[0], op
[1]);
1494 emit(ir
, TGSI_OPCODE_SUB
, result_dst
, op
[0], op
[1]);
1498 emit(ir
, TGSI_OPCODE_MUL
, result_dst
, op
[0], op
[1]);
1501 if (result_dst
.type
== GLSL_TYPE_FLOAT
)
1502 assert(!"not reached: should be handled by ir_div_to_mul_rcp");
1504 emit(ir
, TGSI_OPCODE_DIV
, result_dst
, op
[0], op
[1]);
1507 if (result_dst
.type
== GLSL_TYPE_FLOAT
)
1508 assert(!"ir_binop_mod should have been converted to b * fract(a/b)");
1510 emit(ir
, TGSI_OPCODE_MOD
, result_dst
, op
[0], op
[1]);
1514 emit(ir
, TGSI_OPCODE_SLT
, result_dst
, op
[0], op
[1]);
1516 case ir_binop_greater
:
1517 emit(ir
, TGSI_OPCODE_SLT
, result_dst
, op
[1], op
[0]);
1519 case ir_binop_lequal
:
1520 emit(ir
, TGSI_OPCODE_SGE
, result_dst
, op
[1], op
[0]);
1522 case ir_binop_gequal
:
1523 emit(ir
, TGSI_OPCODE_SGE
, result_dst
, op
[0], op
[1]);
1525 case ir_binop_equal
:
1526 emit(ir
, TGSI_OPCODE_SEQ
, result_dst
, op
[0], op
[1]);
1528 case ir_binop_nequal
:
1529 emit(ir
, TGSI_OPCODE_SNE
, result_dst
, op
[0], op
[1]);
1531 case ir_binop_all_equal
:
1532 /* "==" operator producing a scalar boolean. */
1533 if (ir
->operands
[0]->type
->is_vector() ||
1534 ir
->operands
[1]->type
->is_vector()) {
1535 st_src_reg temp
= get_temp(native_integers
?
1536 glsl_type::get_instance(ir
->operands
[0]->type
->base_type
, 4, 1) :
1537 glsl_type::vec4_type
);
1539 if (native_integers
) {
1540 st_dst_reg temp_dst
= st_dst_reg(temp
);
1541 st_src_reg temp1
= st_src_reg(temp
), temp2
= st_src_reg(temp
);
1543 emit(ir
, TGSI_OPCODE_SEQ
, st_dst_reg(temp
), op
[0], op
[1]);
1545 /* Emit 1-3 AND operations to combine the SEQ results. */
1546 switch (ir
->operands
[0]->type
->vector_elements
) {
1550 temp_dst
.writemask
= WRITEMASK_Y
;
1551 temp1
.swizzle
= SWIZZLE_YYYY
;
1552 temp2
.swizzle
= SWIZZLE_ZZZZ
;
1553 emit(ir
, TGSI_OPCODE_AND
, temp_dst
, temp1
, temp2
);
1556 temp_dst
.writemask
= WRITEMASK_X
;
1557 temp1
.swizzle
= SWIZZLE_XXXX
;
1558 temp2
.swizzle
= SWIZZLE_YYYY
;
1559 emit(ir
, TGSI_OPCODE_AND
, temp_dst
, temp1
, temp2
);
1560 temp_dst
.writemask
= WRITEMASK_Y
;
1561 temp1
.swizzle
= SWIZZLE_ZZZZ
;
1562 temp2
.swizzle
= SWIZZLE_WWWW
;
1563 emit(ir
, TGSI_OPCODE_AND
, temp_dst
, temp1
, temp2
);
1566 temp1
.swizzle
= SWIZZLE_XXXX
;
1567 temp2
.swizzle
= SWIZZLE_YYYY
;
1568 emit(ir
, TGSI_OPCODE_AND
, result_dst
, temp1
, temp2
);
1570 emit(ir
, TGSI_OPCODE_SNE
, st_dst_reg(temp
), op
[0], op
[1]);
1572 /* After the dot-product, the value will be an integer on the
1573 * range [0,4]. Zero becomes 1.0, and positive values become zero.
1575 emit_dp(ir
, result_dst
, temp
, temp
, vector_elements
);
1577 /* Negating the result of the dot-product gives values on the range
1578 * [-4, 0]. Zero becomes 1.0, and negative values become zero.
1579 * This is achieved using SGE.
1581 st_src_reg sge_src
= result_src
;
1582 sge_src
.negate
= ~sge_src
.negate
;
1583 emit(ir
, TGSI_OPCODE_SGE
, result_dst
, sge_src
, st_src_reg_for_float(0.0));
1586 emit(ir
, TGSI_OPCODE_SEQ
, result_dst
, op
[0], op
[1]);
1589 case ir_binop_any_nequal
:
1590 /* "!=" operator producing a scalar boolean. */
1591 if (ir
->operands
[0]->type
->is_vector() ||
1592 ir
->operands
[1]->type
->is_vector()) {
1593 st_src_reg temp
= get_temp(native_integers
?
1594 glsl_type::get_instance(ir
->operands
[0]->type
->base_type
, 4, 1) :
1595 glsl_type::vec4_type
);
1596 emit(ir
, TGSI_OPCODE_SNE
, st_dst_reg(temp
), op
[0], op
[1]);
1598 if (native_integers
) {
1599 st_dst_reg temp_dst
= st_dst_reg(temp
);
1600 st_src_reg temp1
= st_src_reg(temp
), temp2
= st_src_reg(temp
);
1602 /* Emit 1-3 OR operations to combine the SNE results. */
1603 switch (ir
->operands
[0]->type
->vector_elements
) {
1607 temp_dst
.writemask
= WRITEMASK_Y
;
1608 temp1
.swizzle
= SWIZZLE_YYYY
;
1609 temp2
.swizzle
= SWIZZLE_ZZZZ
;
1610 emit(ir
, TGSI_OPCODE_OR
, temp_dst
, temp1
, temp2
);
1613 temp_dst
.writemask
= WRITEMASK_X
;
1614 temp1
.swizzle
= SWIZZLE_XXXX
;
1615 temp2
.swizzle
= SWIZZLE_YYYY
;
1616 emit(ir
, TGSI_OPCODE_OR
, temp_dst
, temp1
, temp2
);
1617 temp_dst
.writemask
= WRITEMASK_Y
;
1618 temp1
.swizzle
= SWIZZLE_ZZZZ
;
1619 temp2
.swizzle
= SWIZZLE_WWWW
;
1620 emit(ir
, TGSI_OPCODE_OR
, temp_dst
, temp1
, temp2
);
1623 temp1
.swizzle
= SWIZZLE_XXXX
;
1624 temp2
.swizzle
= SWIZZLE_YYYY
;
1625 emit(ir
, TGSI_OPCODE_OR
, result_dst
, temp1
, temp2
);
1627 /* After the dot-product, the value will be an integer on the
1628 * range [0,4]. Zero stays zero, and positive values become 1.0.
1630 glsl_to_tgsi_instruction
*const dp
=
1631 emit_dp(ir
, result_dst
, temp
, temp
, vector_elements
);
1632 if (this->prog
->Target
== GL_FRAGMENT_PROGRAM_ARB
) {
1633 /* The clamping to [0,1] can be done for free in the fragment
1634 * shader with a saturate.
1636 dp
->saturate
= true;
1638 /* Negating the result of the dot-product gives values on the range
1639 * [-4, 0]. Zero stays zero, and negative values become 1.0. This
1640 * achieved using SLT.
1642 st_src_reg slt_src
= result_src
;
1643 slt_src
.negate
= ~slt_src
.negate
;
1644 emit(ir
, TGSI_OPCODE_SLT
, result_dst
, slt_src
, st_src_reg_for_float(0.0));
1648 emit(ir
, TGSI_OPCODE_SNE
, result_dst
, op
[0], op
[1]);
1653 assert(ir
->operands
[0]->type
->is_vector());
1655 /* After the dot-product, the value will be an integer on the
1656 * range [0,4]. Zero stays zero, and positive values become 1.0.
1658 glsl_to_tgsi_instruction
*const dp
=
1659 emit_dp(ir
, result_dst
, op
[0], op
[0],
1660 ir
->operands
[0]->type
->vector_elements
);
1661 if (this->prog
->Target
== GL_FRAGMENT_PROGRAM_ARB
&&
1662 result_dst
.type
== GLSL_TYPE_FLOAT
) {
1663 /* The clamping to [0,1] can be done for free in the fragment
1664 * shader with a saturate.
1666 dp
->saturate
= true;
1667 } else if (result_dst
.type
== GLSL_TYPE_FLOAT
) {
1668 /* Negating the result of the dot-product gives values on the range
1669 * [-4, 0]. Zero stays zero, and negative values become 1.0. This
1670 * is achieved using SLT.
1672 st_src_reg slt_src
= result_src
;
1673 slt_src
.negate
= ~slt_src
.negate
;
1674 emit(ir
, TGSI_OPCODE_SLT
, result_dst
, slt_src
, st_src_reg_for_float(0.0));
1677 /* Use SNE 0 if integers are being used as boolean values. */
1678 emit(ir
, TGSI_OPCODE_SNE
, result_dst
, result_src
, st_src_reg_for_int(0));
1683 case ir_binop_logic_xor
:
1684 if (native_integers
)
1685 emit(ir
, TGSI_OPCODE_XOR
, result_dst
, op
[0], op
[1]);
1687 emit(ir
, TGSI_OPCODE_SNE
, result_dst
, op
[0], op
[1]);
1690 case ir_binop_logic_or
: {
1691 if (native_integers
) {
1692 /* If integers are used as booleans, we can use an actual "or"
1695 assert(native_integers
);
1696 emit(ir
, TGSI_OPCODE_OR
, result_dst
, op
[0], op
[1]);
1698 /* After the addition, the value will be an integer on the
1699 * range [0,2]. Zero stays zero, and positive values become 1.0.
1701 glsl_to_tgsi_instruction
*add
=
1702 emit(ir
, TGSI_OPCODE_ADD
, result_dst
, op
[0], op
[1]);
1703 if (this->prog
->Target
== GL_FRAGMENT_PROGRAM_ARB
) {
1704 /* The clamping to [0,1] can be done for free in the fragment
1705 * shader with a saturate if floats are being used as boolean values.
1707 add
->saturate
= true;
1709 /* Negating the result of the addition gives values on the range
1710 * [-2, 0]. Zero stays zero, and negative values become 1.0. This
1711 * is achieved using SLT.
1713 st_src_reg slt_src
= result_src
;
1714 slt_src
.negate
= ~slt_src
.negate
;
1715 emit(ir
, TGSI_OPCODE_SLT
, result_dst
, slt_src
, st_src_reg_for_float(0.0));
1721 case ir_binop_logic_and
:
1722 /* If native integers are disabled, the bool args are stored as float 0.0
1723 * or 1.0, so "mul" gives us "and". If they're enabled, just use the
1724 * actual AND opcode.
1726 if (native_integers
)
1727 emit(ir
, TGSI_OPCODE_AND
, result_dst
, op
[0], op
[1]);
1729 emit(ir
, TGSI_OPCODE_MUL
, result_dst
, op
[0], op
[1]);
1733 assert(ir
->operands
[0]->type
->is_vector());
1734 assert(ir
->operands
[0]->type
== ir
->operands
[1]->type
);
1735 emit_dp(ir
, result_dst
, op
[0], op
[1],
1736 ir
->operands
[0]->type
->vector_elements
);
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));
1749 emit_scalar(ir
, TGSI_OPCODE_RSQ
, result_dst
, op
[0]);
1752 if (native_integers
) {
1753 emit(ir
, TGSI_OPCODE_I2F
, result_dst
, op
[0]);
1756 /* fallthrough to next case otherwise */
1758 if (native_integers
) {
1759 emit(ir
, TGSI_OPCODE_AND
, result_dst
, op
[0], st_src_reg_for_float(1.0));
1762 /* fallthrough to next case otherwise */
1765 /* Converting between signed and unsigned integers is a no-op. */
1769 if (native_integers
) {
1770 /* Booleans are stored as integers using ~0 for true and 0 for false.
1771 * GLSL requires that int(bool) return 1 for true and 0 for false.
1772 * This conversion is done with AND, but it could be done with NEG.
1774 emit(ir
, TGSI_OPCODE_AND
, result_dst
, op
[0], st_src_reg_for_int(1));
1776 /* Booleans and integers are both stored as floats when native
1777 * integers are disabled.
1783 if (native_integers
)
1784 emit(ir
, TGSI_OPCODE_F2I
, result_dst
, op
[0]);
1786 emit(ir
, TGSI_OPCODE_TRUNC
, result_dst
, op
[0]);
1789 emit(ir
, TGSI_OPCODE_SNE
, result_dst
, op
[0], st_src_reg_for_float(0.0));
1792 if (native_integers
)
1793 emit(ir
, TGSI_OPCODE_INEG
, result_dst
, op
[0]);
1795 emit(ir
, TGSI_OPCODE_SNE
, result_dst
, op
[0], st_src_reg_for_float(0.0));
1798 emit(ir
, TGSI_OPCODE_TRUNC
, result_dst
, op
[0]);
1801 op
[0].negate
= ~op
[0].negate
;
1802 emit(ir
, TGSI_OPCODE_FLR
, result_dst
, op
[0]);
1803 result_src
.negate
= ~result_src
.negate
;
1806 emit(ir
, TGSI_OPCODE_FLR
, result_dst
, op
[0]);
1809 emit(ir
, TGSI_OPCODE_FRC
, result_dst
, op
[0]);
1813 emit(ir
, TGSI_OPCODE_MIN
, result_dst
, op
[0], op
[1]);
1816 emit(ir
, TGSI_OPCODE_MAX
, result_dst
, op
[0], op
[1]);
1819 emit_scalar(ir
, TGSI_OPCODE_POW
, result_dst
, op
[0], op
[1]);
1822 case ir_unop_bit_not
:
1823 if (native_integers
) {
1824 emit(ir
, TGSI_OPCODE_NOT
, result_dst
, op
[0]);
1828 if (native_integers
) {
1829 emit(ir
, TGSI_OPCODE_U2F
, result_dst
, op
[0]);
1832 case ir_binop_lshift
:
1833 if (native_integers
) {
1834 emit(ir
, TGSI_OPCODE_SHL
, result_dst
, op
[0]);
1837 case ir_binop_rshift
:
1838 if (native_integers
) {
1839 emit(ir
, TGSI_OPCODE_ISHR
, result_dst
, op
[0]);
1842 case ir_binop_bit_and
:
1843 if (native_integers
) {
1844 emit(ir
, TGSI_OPCODE_AND
, result_dst
, op
[0]);
1847 case ir_binop_bit_xor
:
1848 if (native_integers
) {
1849 emit(ir
, TGSI_OPCODE_XOR
, result_dst
, op
[0]);
1852 case ir_binop_bit_or
:
1853 if (native_integers
) {
1854 emit(ir
, TGSI_OPCODE_OR
, result_dst
, op
[0]);
1857 case ir_unop_round_even
:
1858 assert(!"GLSL 1.30 features unsupported");
1861 case ir_quadop_vector
:
1862 /* This operation should have already been handled.
1864 assert(!"Should not get here.");
1868 this->result
= result_src
;
1873 glsl_to_tgsi_visitor::visit(ir_swizzle
*ir
)
1879 /* Note that this is only swizzles in expressions, not those on the left
1880 * hand side of an assignment, which do write masking. See ir_assignment
1884 ir
->val
->accept(this);
1886 assert(src
.file
!= PROGRAM_UNDEFINED
);
1888 for (i
= 0; i
< 4; i
++) {
1889 if (i
< ir
->type
->vector_elements
) {
1892 swizzle
[i
] = GET_SWZ(src
.swizzle
, ir
->mask
.x
);
1895 swizzle
[i
] = GET_SWZ(src
.swizzle
, ir
->mask
.y
);
1898 swizzle
[i
] = GET_SWZ(src
.swizzle
, ir
->mask
.z
);
1901 swizzle
[i
] = GET_SWZ(src
.swizzle
, ir
->mask
.w
);
1905 /* If the type is smaller than a vec4, replicate the last
1908 swizzle
[i
] = swizzle
[ir
->type
->vector_elements
- 1];
1912 src
.swizzle
= MAKE_SWIZZLE4(swizzle
[0], swizzle
[1], swizzle
[2], swizzle
[3]);
1918 glsl_to_tgsi_visitor::visit(ir_dereference_variable
*ir
)
1920 variable_storage
*entry
= find_variable_storage(ir
->var
);
1921 ir_variable
*var
= ir
->var
;
1924 switch (var
->mode
) {
1925 case ir_var_uniform
:
1926 entry
= new(mem_ctx
) variable_storage(var
, PROGRAM_UNIFORM
,
1928 this->variables
.push_tail(entry
);
1932 /* The linker assigns locations for varyings and attributes,
1933 * including deprecated builtins (like gl_Color), user-assign
1934 * generic attributes (glBindVertexLocation), and
1935 * user-defined varyings.
1937 * FINISHME: We would hit this path for function arguments. Fix!
1939 assert(var
->location
!= -1);
1940 entry
= new(mem_ctx
) variable_storage(var
,
1945 assert(var
->location
!= -1);
1946 entry
= new(mem_ctx
) variable_storage(var
,
1950 case ir_var_system_value
:
1951 entry
= new(mem_ctx
) variable_storage(var
,
1952 PROGRAM_SYSTEM_VALUE
,
1956 case ir_var_temporary
:
1957 entry
= new(mem_ctx
) variable_storage(var
, PROGRAM_TEMPORARY
,
1959 this->variables
.push_tail(entry
);
1961 next_temp
+= type_size(var
->type
);
1966 printf("Failed to make storage for %s\n", var
->name
);
1971 this->result
= st_src_reg(entry
->file
, entry
->index
, var
->type
);
1972 if (!native_integers
)
1973 this->result
.type
= GLSL_TYPE_FLOAT
;
1977 glsl_to_tgsi_visitor::visit(ir_dereference_array
*ir
)
1981 int element_size
= type_size(ir
->type
);
1983 index
= ir
->array_index
->constant_expression_value();
1985 ir
->array
->accept(this);
1989 src
.index
+= index
->value
.i
[0] * element_size
;
1991 /* Variable index array dereference. It eats the "vec4" of the
1992 * base of the array and an index that offsets the TGSI register
1995 ir
->array_index
->accept(this);
1997 st_src_reg index_reg
;
1999 if (element_size
== 1) {
2000 index_reg
= this->result
;
2002 index_reg
= get_temp(native_integers
?
2003 glsl_type::int_type
: glsl_type::float_type
);
2005 emit(ir
, TGSI_OPCODE_MUL
, st_dst_reg(index_reg
),
2006 this->result
, st_src_reg_for_type(index_reg
.type
, element_size
));
2009 /* If there was already a relative address register involved, add the
2010 * new and the old together to get the new offset.
2012 if (src
.reladdr
!= NULL
) {
2013 st_src_reg accum_reg
= get_temp(native_integers
?
2014 glsl_type::int_type
: glsl_type::float_type
);
2016 emit(ir
, TGSI_OPCODE_ADD
, st_dst_reg(accum_reg
),
2017 index_reg
, *src
.reladdr
);
2019 index_reg
= accum_reg
;
2022 src
.reladdr
= ralloc(mem_ctx
, st_src_reg
);
2023 memcpy(src
.reladdr
, &index_reg
, sizeof(index_reg
));
2026 /* If the type is smaller than a vec4, replicate the last channel out. */
2027 if (ir
->type
->is_scalar() || ir
->type
->is_vector())
2028 src
.swizzle
= swizzle_for_size(ir
->type
->vector_elements
);
2030 src
.swizzle
= SWIZZLE_NOOP
;
2036 glsl_to_tgsi_visitor::visit(ir_dereference_record
*ir
)
2039 const glsl_type
*struct_type
= ir
->record
->type
;
2042 ir
->record
->accept(this);
2044 for (i
= 0; i
< struct_type
->length
; i
++) {
2045 if (strcmp(struct_type
->fields
.structure
[i
].name
, ir
->field
) == 0)
2047 offset
+= type_size(struct_type
->fields
.structure
[i
].type
);
2050 /* If the type is smaller than a vec4, replicate the last channel out. */
2051 if (ir
->type
->is_scalar() || ir
->type
->is_vector())
2052 this->result
.swizzle
= swizzle_for_size(ir
->type
->vector_elements
);
2054 this->result
.swizzle
= SWIZZLE_NOOP
;
2056 this->result
.index
+= offset
;
2060 * We want to be careful in assignment setup to hit the actual storage
2061 * instead of potentially using a temporary like we might with the
2062 * ir_dereference handler.
2065 get_assignment_lhs(ir_dereference
*ir
, glsl_to_tgsi_visitor
*v
)
2067 /* The LHS must be a dereference. If the LHS is a variable indexed array
2068 * access of a vector, it must be separated into a series conditional moves
2069 * before reaching this point (see ir_vec_index_to_cond_assign).
2071 assert(ir
->as_dereference());
2072 ir_dereference_array
*deref_array
= ir
->as_dereference_array();
2074 assert(!deref_array
->array
->type
->is_vector());
2077 /* Use the rvalue deref handler for the most part. We'll ignore
2078 * swizzles in it and write swizzles using writemask, though.
2081 return st_dst_reg(v
->result
);
2085 * Process the condition of a conditional assignment
2087 * Examines the condition of a conditional assignment to generate the optimal
2088 * first operand of a \c CMP instruction. If the condition is a relational
2089 * operator with 0 (e.g., \c ir_binop_less), the value being compared will be
2090 * used as the source for the \c CMP instruction. Otherwise the comparison
2091 * is processed to a boolean result, and the boolean result is used as the
2092 * operand to the CMP instruction.
2095 glsl_to_tgsi_visitor::process_move_condition(ir_rvalue
*ir
)
2097 ir_rvalue
*src_ir
= ir
;
2099 bool switch_order
= false;
2101 ir_expression
*const expr
= ir
->as_expression();
2102 if ((expr
!= NULL
) && (expr
->get_num_operands() == 2)) {
2103 bool zero_on_left
= false;
2105 if (expr
->operands
[0]->is_zero()) {
2106 src_ir
= expr
->operands
[1];
2107 zero_on_left
= true;
2108 } else if (expr
->operands
[1]->is_zero()) {
2109 src_ir
= expr
->operands
[0];
2110 zero_on_left
= false;
2114 * (a < 0) T F F ( a < 0) T F F
2115 * (0 < a) F F T (-a < 0) F F T
2116 * (a <= 0) T T F (-a < 0) F F T (swap order of other operands)
2117 * (0 <= a) F T T ( a < 0) T F F (swap order of other operands)
2118 * (a > 0) F F T (-a < 0) F F T
2119 * (0 > a) T F F ( a < 0) T F F
2120 * (a >= 0) F T T ( a < 0) T F F (swap order of other operands)
2121 * (0 >= a) T T F (-a < 0) F F T (swap order of other operands)
2123 * Note that exchanging the order of 0 and 'a' in the comparison simply
2124 * means that the value of 'a' should be negated.
2127 switch (expr
->operation
) {
2129 switch_order
= false;
2130 negate
= zero_on_left
;
2133 case ir_binop_greater
:
2134 switch_order
= false;
2135 negate
= !zero_on_left
;
2138 case ir_binop_lequal
:
2139 switch_order
= true;
2140 negate
= !zero_on_left
;
2143 case ir_binop_gequal
:
2144 switch_order
= true;
2145 negate
= zero_on_left
;
2149 /* This isn't the right kind of comparison afterall, so make sure
2150 * the whole condition is visited.
2158 src_ir
->accept(this);
2160 /* We use the TGSI_OPCODE_CMP (a < 0 ? b : c) for conditional moves, and the
2161 * condition we produced is 0.0 or 1.0. By flipping the sign, we can
2162 * choose which value TGSI_OPCODE_CMP produces without an extra instruction
2163 * computing the condition.
2166 this->result
.negate
= ~this->result
.negate
;
2168 return switch_order
;
2172 glsl_to_tgsi_visitor::visit(ir_assignment
*ir
)
2178 ir
->rhs
->accept(this);
2181 l
= get_assignment_lhs(ir
->lhs
, this);
2183 /* FINISHME: This should really set to the correct maximal writemask for each
2184 * FINISHME: component written (in the loops below). This case can only
2185 * FINISHME: occur for matrices, arrays, and structures.
2187 if (ir
->write_mask
== 0) {
2188 assert(!ir
->lhs
->type
->is_scalar() && !ir
->lhs
->type
->is_vector());
2189 l
.writemask
= WRITEMASK_XYZW
;
2190 } else if (ir
->lhs
->type
->is_scalar() &&
2191 ir
->lhs
->variable_referenced()->mode
== ir_var_out
) {
2192 /* FINISHME: This hack makes writing to gl_FragDepth, which lives in the
2193 * FINISHME: W component of fragment shader output zero, work correctly.
2195 l
.writemask
= WRITEMASK_XYZW
;
2198 int first_enabled_chan
= 0;
2201 l
.writemask
= ir
->write_mask
;
2203 for (int i
= 0; i
< 4; i
++) {
2204 if (l
.writemask
& (1 << i
)) {
2205 first_enabled_chan
= GET_SWZ(r
.swizzle
, i
);
2210 /* Swizzle a small RHS vector into the channels being written.
2212 * glsl ir treats write_mask as dictating how many channels are
2213 * present on the RHS while TGSI treats write_mask as just
2214 * showing which channels of the vec4 RHS get written.
2216 for (int i
= 0; i
< 4; i
++) {
2217 if (l
.writemask
& (1 << i
))
2218 swizzles
[i
] = GET_SWZ(r
.swizzle
, rhs_chan
++);
2220 swizzles
[i
] = first_enabled_chan
;
2222 r
.swizzle
= MAKE_SWIZZLE4(swizzles
[0], swizzles
[1],
2223 swizzles
[2], swizzles
[3]);
2226 assert(l
.file
!= PROGRAM_UNDEFINED
);
2227 assert(r
.file
!= PROGRAM_UNDEFINED
);
2229 if (ir
->condition
) {
2230 const bool switch_order
= this->process_move_condition(ir
->condition
);
2231 st_src_reg condition
= this->result
;
2233 for (i
= 0; i
< type_size(ir
->lhs
->type
); i
++) {
2234 st_src_reg l_src
= st_src_reg(l
);
2235 st_src_reg condition_temp
= condition
;
2236 l_src
.swizzle
= swizzle_for_size(ir
->lhs
->type
->vector_elements
);
2238 if (native_integers
) {
2239 /* This is necessary because TGSI's CMP instruction expects the
2240 * condition to be a float, and we store booleans as integers.
2241 * If TGSI had a UCMP instruction or similar, this extra
2242 * instruction would not be necessary.
2244 condition_temp
= get_temp(glsl_type::vec4_type
);
2245 condition
.negate
= 0;
2246 emit(ir
, TGSI_OPCODE_I2F
, st_dst_reg(condition_temp
), condition
);
2247 condition_temp
.swizzle
= condition
.swizzle
;
2251 emit(ir
, TGSI_OPCODE_CMP
, l
, condition_temp
, l_src
, r
);
2253 emit(ir
, TGSI_OPCODE_CMP
, l
, condition_temp
, r
, l_src
);
2259 } else if (ir
->rhs
->as_expression() &&
2260 this->instructions
.get_tail() &&
2261 ir
->rhs
== ((glsl_to_tgsi_instruction
*)this->instructions
.get_tail())->ir
&&
2262 type_size(ir
->lhs
->type
) == 1 &&
2263 l
.writemask
== ((glsl_to_tgsi_instruction
*)this->instructions
.get_tail())->dst
.writemask
) {
2264 /* To avoid emitting an extra MOV when assigning an expression to a
2265 * variable, emit the last instruction of the expression again, but
2266 * replace the destination register with the target of the assignment.
2267 * Dead code elimination will remove the original instruction.
2269 glsl_to_tgsi_instruction
*inst
, *new_inst
;
2270 inst
= (glsl_to_tgsi_instruction
*)this->instructions
.get_tail();
2271 new_inst
= emit(ir
, inst
->op
, l
, inst
->src
[0], inst
->src
[1], inst
->src
[2]);
2272 new_inst
->saturate
= inst
->saturate
;
2273 inst
->dead_mask
= inst
->dst
.writemask
;
2275 for (i
= 0; i
< type_size(ir
->lhs
->type
); i
++) {
2276 emit(ir
, TGSI_OPCODE_MOV
, l
, r
);
2285 glsl_to_tgsi_visitor::visit(ir_constant
*ir
)
2288 GLfloat stack_vals
[4] = { 0 };
2289 gl_constant_value
*values
= (gl_constant_value
*) stack_vals
;
2290 GLenum gl_type
= GL_NONE
;
2292 static int in_array
= 0;
2293 gl_register_file file
= in_array
? PROGRAM_CONSTANT
: PROGRAM_IMMEDIATE
;
2295 /* Unfortunately, 4 floats is all we can get into
2296 * _mesa_add_typed_unnamed_constant. So, make a temp to store an
2297 * aggregate constant and move each constant value into it. If we
2298 * get lucky, copy propagation will eliminate the extra moves.
2300 if (ir
->type
->base_type
== GLSL_TYPE_STRUCT
) {
2301 st_src_reg temp_base
= get_temp(ir
->type
);
2302 st_dst_reg temp
= st_dst_reg(temp_base
);
2304 foreach_iter(exec_list_iterator
, iter
, ir
->components
) {
2305 ir_constant
*field_value
= (ir_constant
*)iter
.get();
2306 int size
= type_size(field_value
->type
);
2310 field_value
->accept(this);
2313 for (i
= 0; i
< (unsigned int)size
; i
++) {
2314 emit(ir
, TGSI_OPCODE_MOV
, temp
, src
);
2320 this->result
= temp_base
;
2324 if (ir
->type
->is_array()) {
2325 st_src_reg temp_base
= get_temp(ir
->type
);
2326 st_dst_reg temp
= st_dst_reg(temp_base
);
2327 int size
= type_size(ir
->type
->fields
.array
);
2332 for (i
= 0; i
< ir
->type
->length
; i
++) {
2333 ir
->array_elements
[i
]->accept(this);
2335 for (int j
= 0; j
< size
; j
++) {
2336 emit(ir
, TGSI_OPCODE_MOV
, temp
, src
);
2342 this->result
= temp_base
;
2347 if (ir
->type
->is_matrix()) {
2348 st_src_reg mat
= get_temp(ir
->type
);
2349 st_dst_reg mat_column
= st_dst_reg(mat
);
2351 for (i
= 0; i
< ir
->type
->matrix_columns
; i
++) {
2352 assert(ir
->type
->base_type
== GLSL_TYPE_FLOAT
);
2353 values
= (gl_constant_value
*) &ir
->value
.f
[i
* ir
->type
->vector_elements
];
2355 src
= st_src_reg(file
, -1, ir
->type
->base_type
);
2356 src
.index
= add_constant(file
,
2358 ir
->type
->vector_elements
,
2361 emit(ir
, TGSI_OPCODE_MOV
, mat_column
, src
);
2370 switch (ir
->type
->base_type
) {
2371 case GLSL_TYPE_FLOAT
:
2373 for (i
= 0; i
< ir
->type
->vector_elements
; i
++) {
2374 values
[i
].f
= ir
->value
.f
[i
];
2377 case GLSL_TYPE_UINT
:
2378 gl_type
= native_integers
? GL_UNSIGNED_INT
: GL_FLOAT
;
2379 for (i
= 0; i
< ir
->type
->vector_elements
; i
++) {
2380 if (native_integers
)
2381 values
[i
].u
= ir
->value
.u
[i
];
2383 values
[i
].f
= ir
->value
.u
[i
];
2387 gl_type
= native_integers
? GL_INT
: GL_FLOAT
;
2388 for (i
= 0; i
< ir
->type
->vector_elements
; i
++) {
2389 if (native_integers
)
2390 values
[i
].i
= ir
->value
.i
[i
];
2392 values
[i
].f
= ir
->value
.i
[i
];
2395 case GLSL_TYPE_BOOL
:
2396 gl_type
= native_integers
? GL_BOOL
: GL_FLOAT
;
2397 for (i
= 0; i
< ir
->type
->vector_elements
; i
++) {
2398 if (native_integers
)
2399 values
[i
].b
= ir
->value
.b
[i
];
2401 values
[i
].f
= ir
->value
.b
[i
];
2405 assert(!"Non-float/uint/int/bool constant");
2408 this->result
= st_src_reg(file
, -1, ir
->type
);
2409 this->result
.index
= add_constant(file
,
2411 ir
->type
->vector_elements
,
2413 &this->result
.swizzle
);
2417 glsl_to_tgsi_visitor::get_function_signature(ir_function_signature
*sig
)
2419 function_entry
*entry
;
2421 foreach_iter(exec_list_iterator
, iter
, this->function_signatures
) {
2422 entry
= (function_entry
*)iter
.get();
2424 if (entry
->sig
== sig
)
2428 entry
= ralloc(mem_ctx
, function_entry
);
2430 entry
->sig_id
= this->next_signature_id
++;
2431 entry
->bgn_inst
= NULL
;
2433 /* Allocate storage for all the parameters. */
2434 foreach_iter(exec_list_iterator
, iter
, sig
->parameters
) {
2435 ir_variable
*param
= (ir_variable
*)iter
.get();
2436 variable_storage
*storage
;
2438 storage
= find_variable_storage(param
);
2441 storage
= new(mem_ctx
) variable_storage(param
, PROGRAM_TEMPORARY
,
2443 this->variables
.push_tail(storage
);
2445 this->next_temp
+= type_size(param
->type
);
2448 if (!sig
->return_type
->is_void()) {
2449 entry
->return_reg
= get_temp(sig
->return_type
);
2451 entry
->return_reg
= undef_src
;
2454 this->function_signatures
.push_tail(entry
);
2459 glsl_to_tgsi_visitor::visit(ir_call
*ir
)
2461 glsl_to_tgsi_instruction
*call_inst
;
2462 ir_function_signature
*sig
= ir
->get_callee();
2463 function_entry
*entry
= get_function_signature(sig
);
2466 /* Process in parameters. */
2467 exec_list_iterator sig_iter
= sig
->parameters
.iterator();
2468 foreach_iter(exec_list_iterator
, iter
, *ir
) {
2469 ir_rvalue
*param_rval
= (ir_rvalue
*)iter
.get();
2470 ir_variable
*param
= (ir_variable
*)sig_iter
.get();
2472 if (param
->mode
== ir_var_in
||
2473 param
->mode
== ir_var_inout
) {
2474 variable_storage
*storage
= find_variable_storage(param
);
2477 param_rval
->accept(this);
2478 st_src_reg r
= this->result
;
2481 l
.file
= storage
->file
;
2482 l
.index
= storage
->index
;
2484 l
.writemask
= WRITEMASK_XYZW
;
2485 l
.cond_mask
= COND_TR
;
2487 for (i
= 0; i
< type_size(param
->type
); i
++) {
2488 emit(ir
, TGSI_OPCODE_MOV
, l
, r
);
2496 assert(!sig_iter
.has_next());
2498 /* Emit call instruction */
2499 call_inst
= emit(ir
, TGSI_OPCODE_CAL
);
2500 call_inst
->function
= entry
;
2502 /* Process out parameters. */
2503 sig_iter
= sig
->parameters
.iterator();
2504 foreach_iter(exec_list_iterator
, iter
, *ir
) {
2505 ir_rvalue
*param_rval
= (ir_rvalue
*)iter
.get();
2506 ir_variable
*param
= (ir_variable
*)sig_iter
.get();
2508 if (param
->mode
== ir_var_out
||
2509 param
->mode
== ir_var_inout
) {
2510 variable_storage
*storage
= find_variable_storage(param
);
2514 r
.file
= storage
->file
;
2515 r
.index
= storage
->index
;
2517 r
.swizzle
= SWIZZLE_NOOP
;
2520 param_rval
->accept(this);
2521 st_dst_reg l
= st_dst_reg(this->result
);
2523 for (i
= 0; i
< type_size(param
->type
); i
++) {
2524 emit(ir
, TGSI_OPCODE_MOV
, l
, r
);
2532 assert(!sig_iter
.has_next());
2534 /* Process return value. */
2535 this->result
= entry
->return_reg
;
2539 glsl_to_tgsi_visitor::visit(ir_texture
*ir
)
2541 st_src_reg result_src
, coord
, lod_info
, projector
, dx
, dy
, offset
;
2542 st_dst_reg result_dst
, coord_dst
;
2543 glsl_to_tgsi_instruction
*inst
= NULL
;
2544 unsigned opcode
= TGSI_OPCODE_NOP
;
2546 if (ir
->coordinate
) {
2547 ir
->coordinate
->accept(this);
2549 /* Put our coords in a temp. We'll need to modify them for shadow,
2550 * projection, or LOD, so the only case we'd use it as is is if
2551 * we're doing plain old texturing. The optimization passes on
2552 * glsl_to_tgsi_visitor should handle cleaning up our mess in that case.
2554 coord
= get_temp(glsl_type::vec4_type
);
2555 coord_dst
= st_dst_reg(coord
);
2556 emit(ir
, TGSI_OPCODE_MOV
, coord_dst
, this->result
);
2559 if (ir
->projector
) {
2560 ir
->projector
->accept(this);
2561 projector
= this->result
;
2564 /* Storage for our result. Ideally for an assignment we'd be using
2565 * the actual storage for the result here, instead.
2567 result_src
= get_temp(glsl_type::vec4_type
);
2568 result_dst
= st_dst_reg(result_src
);
2572 opcode
= TGSI_OPCODE_TEX
;
2575 opcode
= TGSI_OPCODE_TXB
;
2576 ir
->lod_info
.bias
->accept(this);
2577 lod_info
= this->result
;
2580 opcode
= TGSI_OPCODE_TXL
;
2581 ir
->lod_info
.lod
->accept(this);
2582 lod_info
= this->result
;
2585 opcode
= TGSI_OPCODE_TXD
;
2586 ir
->lod_info
.grad
.dPdx
->accept(this);
2588 ir
->lod_info
.grad
.dPdy
->accept(this);
2592 opcode
= TGSI_OPCODE_TXQ
;
2593 ir
->lod_info
.lod
->accept(this);
2594 lod_info
= this->result
;
2597 opcode
= TGSI_OPCODE_TXF
;
2598 ir
->lod_info
.lod
->accept(this);
2599 lod_info
= this->result
;
2601 ir
->offset
->accept(this);
2602 offset
= this->result
;
2607 const glsl_type
*sampler_type
= ir
->sampler
->type
;
2609 if (ir
->projector
) {
2610 if (opcode
== TGSI_OPCODE_TEX
) {
2611 /* Slot the projector in as the last component of the coord. */
2612 coord_dst
.writemask
= WRITEMASK_W
;
2613 emit(ir
, TGSI_OPCODE_MOV
, coord_dst
, projector
);
2614 coord_dst
.writemask
= WRITEMASK_XYZW
;
2615 opcode
= TGSI_OPCODE_TXP
;
2617 st_src_reg coord_w
= coord
;
2618 coord_w
.swizzle
= SWIZZLE_WWWW
;
2620 /* For the other TEX opcodes there's no projective version
2621 * since the last slot is taken up by LOD info. Do the
2622 * projective divide now.
2624 coord_dst
.writemask
= WRITEMASK_W
;
2625 emit(ir
, TGSI_OPCODE_RCP
, coord_dst
, projector
);
2627 /* In the case where we have to project the coordinates "by hand,"
2628 * the shadow comparator value must also be projected.
2630 st_src_reg tmp_src
= coord
;
2631 if (ir
->shadow_comparitor
) {
2632 /* Slot the shadow value in as the second to last component of the
2635 ir
->shadow_comparitor
->accept(this);
2637 tmp_src
= get_temp(glsl_type::vec4_type
);
2638 st_dst_reg tmp_dst
= st_dst_reg(tmp_src
);
2640 /* Projective division not allowed for array samplers. */
2641 assert(!sampler_type
->sampler_array
);
2643 tmp_dst
.writemask
= WRITEMASK_Z
;
2644 emit(ir
, TGSI_OPCODE_MOV
, tmp_dst
, this->result
);
2646 tmp_dst
.writemask
= WRITEMASK_XY
;
2647 emit(ir
, TGSI_OPCODE_MOV
, tmp_dst
, coord
);
2650 coord_dst
.writemask
= WRITEMASK_XYZ
;
2651 emit(ir
, TGSI_OPCODE_MUL
, coord_dst
, tmp_src
, coord_w
);
2653 coord_dst
.writemask
= WRITEMASK_XYZW
;
2654 coord
.swizzle
= SWIZZLE_XYZW
;
2658 /* If projection is done and the opcode is not TGSI_OPCODE_TXP, then the shadow
2659 * comparator was put in the correct place (and projected) by the code,
2660 * above, that handles by-hand projection.
2662 if (ir
->shadow_comparitor
&& (!ir
->projector
|| opcode
== TGSI_OPCODE_TXP
)) {
2663 /* Slot the shadow value in as the second to last component of the
2666 ir
->shadow_comparitor
->accept(this);
2668 /* XXX This will need to be updated for cubemap array samplers. */
2669 if (sampler_type
->sampler_dimensionality
== GLSL_SAMPLER_DIM_2D
&&
2670 sampler_type
->sampler_array
) {
2671 coord_dst
.writemask
= WRITEMASK_W
;
2673 coord_dst
.writemask
= WRITEMASK_Z
;
2676 emit(ir
, TGSI_OPCODE_MOV
, coord_dst
, this->result
);
2677 coord_dst
.writemask
= WRITEMASK_XYZW
;
2680 if (opcode
== TGSI_OPCODE_TXL
|| opcode
== TGSI_OPCODE_TXB
||
2681 opcode
== TGSI_OPCODE_TXF
) {
2682 /* TGSI stores LOD or LOD bias in the last channel of the coords. */
2683 coord_dst
.writemask
= WRITEMASK_W
;
2684 emit(ir
, TGSI_OPCODE_MOV
, coord_dst
, lod_info
);
2685 coord_dst
.writemask
= WRITEMASK_XYZW
;
2688 if (opcode
== TGSI_OPCODE_TXD
)
2689 inst
= emit(ir
, opcode
, result_dst
, coord
, dx
, dy
);
2690 else if (opcode
== TGSI_OPCODE_TXQ
)
2691 inst
= emit(ir
, opcode
, result_dst
, lod_info
);
2692 else if (opcode
== TGSI_OPCODE_TXF
) {
2693 inst
= emit(ir
, opcode
, result_dst
, coord
);
2695 inst
= emit(ir
, opcode
, result_dst
, coord
);
2697 if (ir
->shadow_comparitor
)
2698 inst
->tex_shadow
= GL_TRUE
;
2700 inst
->sampler
= _mesa_get_sampler_uniform_value(ir
->sampler
,
2701 this->shader_program
,
2705 inst
->tex_offset_num_offset
= 1;
2706 inst
->tex_offsets
[0].Index
= offset
.index
;
2707 inst
->tex_offsets
[0].File
= offset
.file
;
2708 inst
->tex_offsets
[0].SwizzleX
= GET_SWZ(offset
.swizzle
, 0);
2709 inst
->tex_offsets
[0].SwizzleY
= GET_SWZ(offset
.swizzle
, 1);
2710 inst
->tex_offsets
[0].SwizzleZ
= GET_SWZ(offset
.swizzle
, 2);
2713 switch (sampler_type
->sampler_dimensionality
) {
2714 case GLSL_SAMPLER_DIM_1D
:
2715 inst
->tex_target
= (sampler_type
->sampler_array
)
2716 ? TEXTURE_1D_ARRAY_INDEX
: TEXTURE_1D_INDEX
;
2718 case GLSL_SAMPLER_DIM_2D
:
2719 inst
->tex_target
= (sampler_type
->sampler_array
)
2720 ? TEXTURE_2D_ARRAY_INDEX
: TEXTURE_2D_INDEX
;
2722 case GLSL_SAMPLER_DIM_3D
:
2723 inst
->tex_target
= TEXTURE_3D_INDEX
;
2725 case GLSL_SAMPLER_DIM_CUBE
:
2726 inst
->tex_target
= TEXTURE_CUBE_INDEX
;
2728 case GLSL_SAMPLER_DIM_RECT
:
2729 inst
->tex_target
= TEXTURE_RECT_INDEX
;
2731 case GLSL_SAMPLER_DIM_BUF
:
2732 assert(!"FINISHME: Implement ARB_texture_buffer_object");
2734 case GLSL_SAMPLER_DIM_EXTERNAL
:
2735 inst
->tex_target
= TEXTURE_EXTERNAL_INDEX
;
2738 assert(!"Should not get here.");
2741 this->result
= result_src
;
2745 glsl_to_tgsi_visitor::visit(ir_return
*ir
)
2747 if (ir
->get_value()) {
2751 assert(current_function
);
2753 ir
->get_value()->accept(this);
2754 st_src_reg r
= this->result
;
2756 l
= st_dst_reg(current_function
->return_reg
);
2758 for (i
= 0; i
< type_size(current_function
->sig
->return_type
); i
++) {
2759 emit(ir
, TGSI_OPCODE_MOV
, l
, r
);
2765 emit(ir
, TGSI_OPCODE_RET
);
2769 glsl_to_tgsi_visitor::visit(ir_discard
*ir
)
2771 struct gl_fragment_program
*fp
= (struct gl_fragment_program
*)this->prog
;
2773 if (ir
->condition
) {
2774 ir
->condition
->accept(this);
2775 this->result
.negate
= ~this->result
.negate
;
2776 emit(ir
, TGSI_OPCODE_KIL
, undef_dst
, this->result
);
2778 emit(ir
, TGSI_OPCODE_KILP
);
2781 fp
->UsesKill
= GL_TRUE
;
2785 glsl_to_tgsi_visitor::visit(ir_if
*ir
)
2787 glsl_to_tgsi_instruction
*cond_inst
, *if_inst
;
2788 glsl_to_tgsi_instruction
*prev_inst
;
2790 prev_inst
= (glsl_to_tgsi_instruction
*)this->instructions
.get_tail();
2792 ir
->condition
->accept(this);
2793 assert(this->result
.file
!= PROGRAM_UNDEFINED
);
2795 if (this->options
->EmitCondCodes
) {
2796 cond_inst
= (glsl_to_tgsi_instruction
*)this->instructions
.get_tail();
2798 /* See if we actually generated any instruction for generating
2799 * the condition. If not, then cook up a move to a temp so we
2800 * have something to set cond_update on.
2802 if (cond_inst
== prev_inst
) {
2803 st_src_reg temp
= get_temp(glsl_type::bool_type
);
2804 cond_inst
= emit(ir
->condition
, TGSI_OPCODE_MOV
, st_dst_reg(temp
), result
);
2806 cond_inst
->cond_update
= GL_TRUE
;
2808 if_inst
= emit(ir
->condition
, TGSI_OPCODE_IF
);
2809 if_inst
->dst
.cond_mask
= COND_NE
;
2811 if_inst
= emit(ir
->condition
, TGSI_OPCODE_IF
, undef_dst
, this->result
);
2814 this->instructions
.push_tail(if_inst
);
2816 visit_exec_list(&ir
->then_instructions
, this);
2818 if (!ir
->else_instructions
.is_empty()) {
2819 emit(ir
->condition
, TGSI_OPCODE_ELSE
);
2820 visit_exec_list(&ir
->else_instructions
, this);
2823 if_inst
= emit(ir
->condition
, TGSI_OPCODE_ENDIF
);
2826 glsl_to_tgsi_visitor::glsl_to_tgsi_visitor()
2828 result
.file
= PROGRAM_UNDEFINED
;
2830 next_signature_id
= 1;
2832 current_function
= NULL
;
2833 num_address_regs
= 0;
2834 indirect_addr_temps
= false;
2835 indirect_addr_consts
= false;
2836 mem_ctx
= ralloc_context(NULL
);
2839 glsl_to_tgsi_visitor::~glsl_to_tgsi_visitor()
2841 ralloc_free(mem_ctx
);
2844 extern "C" void free_glsl_to_tgsi_visitor(glsl_to_tgsi_visitor
*v
)
2851 * Count resources used by the given gpu program (number of texture
2855 count_resources(glsl_to_tgsi_visitor
*v
, gl_program
*prog
)
2857 v
->samplers_used
= 0;
2859 foreach_iter(exec_list_iterator
, iter
, v
->instructions
) {
2860 glsl_to_tgsi_instruction
*inst
= (glsl_to_tgsi_instruction
*)iter
.get();
2862 if (is_tex_instruction(inst
->op
)) {
2863 v
->samplers_used
|= 1 << inst
->sampler
;
2865 prog
->SamplerTargets
[inst
->sampler
] =
2866 (gl_texture_index
)inst
->tex_target
;
2867 if (inst
->tex_shadow
) {
2868 prog
->ShadowSamplers
|= 1 << inst
->sampler
;
2873 prog
->SamplersUsed
= v
->samplers_used
;
2874 _mesa_update_shader_textures_used(prog
);
2879 * Check if the given vertex/fragment/shader program is within the
2880 * resource limits of the context (number of texture units, etc).
2881 * If any of those checks fail, record a linker error.
2883 * XXX more checks are needed...
2886 check_resources(const struct gl_context
*ctx
,
2887 struct gl_shader_program
*shader_program
,
2888 glsl_to_tgsi_visitor
*prog
,
2889 struct gl_program
*proginfo
)
2891 switch (proginfo
->Target
) {
2892 case GL_VERTEX_PROGRAM_ARB
:
2893 if (_mesa_bitcount(prog
->samplers_used
) >
2894 ctx
->Const
.MaxVertexTextureImageUnits
) {
2895 fail_link(shader_program
, "Too many vertex shader texture samplers");
2897 if (proginfo
->Parameters
->NumParameters
> MAX_UNIFORMS
) {
2898 fail_link(shader_program
, "Too many vertex shader constants");
2901 case MESA_GEOMETRY_PROGRAM
:
2902 if (_mesa_bitcount(prog
->samplers_used
) >
2903 ctx
->Const
.MaxGeometryTextureImageUnits
) {
2904 fail_link(shader_program
, "Too many geometry shader texture samplers");
2906 if (proginfo
->Parameters
->NumParameters
>
2907 MAX_GEOMETRY_UNIFORM_COMPONENTS
/ 4) {
2908 fail_link(shader_program
, "Too many geometry shader constants");
2911 case GL_FRAGMENT_PROGRAM_ARB
:
2912 if (_mesa_bitcount(prog
->samplers_used
) >
2913 ctx
->Const
.MaxTextureImageUnits
) {
2914 fail_link(shader_program
, "Too many fragment shader texture samplers");
2916 if (proginfo
->Parameters
->NumParameters
> MAX_UNIFORMS
) {
2917 fail_link(shader_program
, "Too many fragment shader constants");
2921 _mesa_problem(ctx
, "unexpected program type in check_resources()");
2927 set_uniform_initializer(struct gl_context
*ctx
, void *mem_ctx
,
2928 struct gl_shader_program
*shader_program
,
2929 const char *name
, const glsl_type
*type
,
2932 if (type
->is_record()) {
2933 ir_constant
*field_constant
;
2935 field_constant
= (ir_constant
*)val
->components
.get_head();
2937 for (unsigned int i
= 0; i
< type
->length
; i
++) {
2938 const glsl_type
*field_type
= type
->fields
.structure
[i
].type
;
2939 const char *field_name
= ralloc_asprintf(mem_ctx
, "%s.%s", name
,
2940 type
->fields
.structure
[i
].name
);
2941 set_uniform_initializer(ctx
, mem_ctx
, shader_program
, field_name
,
2942 field_type
, field_constant
);
2943 field_constant
= (ir_constant
*)field_constant
->next
;
2948 int loc
= _mesa_get_uniform_location(ctx
, shader_program
, name
);
2951 fail_link(shader_program
,
2952 "Couldn't find uniform for initializer %s\n", name
);
2956 for (unsigned int i
= 0; i
< (type
->is_array() ? type
->length
: 1); i
++) {
2957 ir_constant
*element
;
2958 const glsl_type
*element_type
;
2959 if (type
->is_array()) {
2960 element
= val
->array_elements
[i
];
2961 element_type
= type
->fields
.array
;
2964 element_type
= type
;
2969 if (element_type
->base_type
== GLSL_TYPE_BOOL
) {
2970 int *conv
= ralloc_array(mem_ctx
, int, element_type
->components());
2971 for (unsigned int j
= 0; j
< element_type
->components(); j
++) {
2972 conv
[j
] = element
->value
.b
[j
];
2974 values
= (void *)conv
;
2975 element_type
= glsl_type::get_instance(GLSL_TYPE_INT
,
2976 element_type
->vector_elements
,
2979 values
= &element
->value
;
2982 if (element_type
->is_matrix()) {
2983 _mesa_uniform_matrix(ctx
, shader_program
,
2984 element_type
->matrix_columns
,
2985 element_type
->vector_elements
,
2986 loc
, 1, GL_FALSE
, (GLfloat
*)values
);
2988 _mesa_uniform(ctx
, shader_program
, loc
, element_type
->matrix_columns
,
2989 values
, element_type
->gl_type
);
2997 * Scan/rewrite program to remove reads of custom (output) registers.
2998 * The passed type has to be either PROGRAM_OUTPUT or PROGRAM_VARYING
2999 * (for vertex shaders).
3000 * In GLSL shaders, varying vars can be read and written.
3001 * On some hardware, trying to read an output register causes trouble.
3002 * So, rewrite the program to use a temporary register in this case.
3004 * Based on _mesa_remove_output_reads from programopt.c.
3007 glsl_to_tgsi_visitor::remove_output_reads(gl_register_file type
)
3010 GLint outputMap
[VERT_RESULT_MAX
];
3011 GLint outputTypes
[VERT_RESULT_MAX
];
3012 GLuint numVaryingReads
= 0;
3013 GLboolean
*usedTemps
;
3014 GLuint firstTemp
= 0;
3016 usedTemps
= new GLboolean
[MAX_TEMPS
];
3020 _mesa_find_used_registers(prog
, PROGRAM_TEMPORARY
,
3021 usedTemps
, MAX_TEMPS
);
3023 assert(type
== PROGRAM_VARYING
|| type
== PROGRAM_OUTPUT
);
3024 assert(prog
->Target
== GL_VERTEX_PROGRAM_ARB
|| type
!= PROGRAM_VARYING
);
3026 for (i
= 0; i
< VERT_RESULT_MAX
; i
++)
3029 /* look for instructions which read from varying vars */
3030 foreach_iter(exec_list_iterator
, iter
, this->instructions
) {
3031 glsl_to_tgsi_instruction
*inst
= (glsl_to_tgsi_instruction
*)iter
.get();
3032 const GLuint numSrc
= num_inst_src_regs(inst
->op
);
3034 for (j
= 0; j
< numSrc
; j
++) {
3035 if (inst
->src
[j
].file
== type
) {
3036 /* replace the read with a temp reg */
3037 const GLuint var
= inst
->src
[j
].index
;
3038 if (outputMap
[var
] == -1) {
3040 outputMap
[var
] = _mesa_find_free_register(usedTemps
,
3043 outputTypes
[var
] = inst
->src
[j
].type
;
3044 firstTemp
= outputMap
[var
] + 1;
3046 inst
->src
[j
].file
= PROGRAM_TEMPORARY
;
3047 inst
->src
[j
].index
= outputMap
[var
];
3052 delete [] usedTemps
;
3054 if (numVaryingReads
== 0)
3055 return; /* nothing to be done */
3057 /* look for instructions which write to the varying vars identified above */
3058 foreach_iter(exec_list_iterator
, iter
, this->instructions
) {
3059 glsl_to_tgsi_instruction
*inst
= (glsl_to_tgsi_instruction
*)iter
.get();
3060 if (inst
->dst
.file
== type
&& outputMap
[inst
->dst
.index
] >= 0) {
3061 /* change inst to write to the temp reg, instead of the varying */
3062 inst
->dst
.file
= PROGRAM_TEMPORARY
;
3063 inst
->dst
.index
= outputMap
[inst
->dst
.index
];
3067 /* insert new MOV instructions at the end */
3068 for (i
= 0; i
< VERT_RESULT_MAX
; i
++) {
3069 if (outputMap
[i
] >= 0) {
3070 /* MOV VAR[i], TEMP[tmp]; */
3071 st_src_reg src
= st_src_reg(PROGRAM_TEMPORARY
, outputMap
[i
], outputTypes
[i
]);
3072 st_dst_reg dst
= st_dst_reg(type
, WRITEMASK_XYZW
, outputTypes
[i
]);
3074 this->emit(NULL
, TGSI_OPCODE_MOV
, dst
, src
);
3080 * Returns the mask of channels (bitmask of WRITEMASK_X,Y,Z,W) which
3081 * are read from the given src in this instruction
3084 get_src_arg_mask(st_dst_reg dst
, st_src_reg src
)
3086 int read_mask
= 0, comp
;
3088 /* Now, given the src swizzle and the written channels, find which
3089 * components are actually read
3091 for (comp
= 0; comp
< 4; ++comp
) {
3092 const unsigned coord
= GET_SWZ(src
.swizzle
, comp
);
3094 if (dst
.writemask
& (1 << comp
) && coord
<= SWIZZLE_W
)
3095 read_mask
|= 1 << coord
;
3102 * This pass replaces CMP T0, T1 T2 T0 with MOV T0, T2 when the CMP
3103 * instruction is the first instruction to write to register T0. There are
3104 * several lowering passes done in GLSL IR (e.g. branches and
3105 * relative addressing) that create a large number of conditional assignments
3106 * that ir_to_mesa converts to CMP instructions like the one mentioned above.
3108 * Here is why this conversion is safe:
3109 * CMP T0, T1 T2 T0 can be expanded to:
3115 * If (T1 < 0.0) evaluates to true then our replacement MOV T0, T2 is the same
3116 * as the original program. If (T1 < 0.0) evaluates to false, executing
3117 * MOV T0, T0 will store a garbage value in T0 since T0 is uninitialized.
3118 * Therefore, it doesn't matter that we are replacing MOV T0, T0 with MOV T0, T2
3119 * because any instruction that was going to read from T0 after this was going
3120 * to read a garbage value anyway.
3123 glsl_to_tgsi_visitor::simplify_cmp(void)
3125 unsigned *tempWrites
;
3126 unsigned outputWrites
[MAX_PROGRAM_OUTPUTS
];
3128 tempWrites
= new unsigned[MAX_TEMPS
];
3132 memset(tempWrites
, 0, sizeof(tempWrites
));
3133 memset(outputWrites
, 0, sizeof(outputWrites
));
3135 foreach_iter(exec_list_iterator
, iter
, this->instructions
) {
3136 glsl_to_tgsi_instruction
*inst
= (glsl_to_tgsi_instruction
*)iter
.get();
3137 unsigned prevWriteMask
= 0;
3139 /* Give up if we encounter relative addressing or flow control. */
3140 if (inst
->dst
.reladdr
||
3141 tgsi_get_opcode_info(inst
->op
)->is_branch
||
3142 inst
->op
== TGSI_OPCODE_BGNSUB
||
3143 inst
->op
== TGSI_OPCODE_CONT
||
3144 inst
->op
== TGSI_OPCODE_END
||
3145 inst
->op
== TGSI_OPCODE_ENDSUB
||
3146 inst
->op
== TGSI_OPCODE_RET
) {
3150 if (inst
->dst
.file
== PROGRAM_OUTPUT
) {
3151 assert(inst
->dst
.index
< MAX_PROGRAM_OUTPUTS
);
3152 prevWriteMask
= outputWrites
[inst
->dst
.index
];
3153 outputWrites
[inst
->dst
.index
] |= inst
->dst
.writemask
;
3154 } else if (inst
->dst
.file
== PROGRAM_TEMPORARY
) {
3155 assert(inst
->dst
.index
< MAX_TEMPS
);
3156 prevWriteMask
= tempWrites
[inst
->dst
.index
];
3157 tempWrites
[inst
->dst
.index
] |= inst
->dst
.writemask
;
3160 /* For a CMP to be considered a conditional write, the destination
3161 * register and source register two must be the same. */
3162 if (inst
->op
== TGSI_OPCODE_CMP
3163 && !(inst
->dst
.writemask
& prevWriteMask
)
3164 && inst
->src
[2].file
== inst
->dst
.file
3165 && inst
->src
[2].index
== inst
->dst
.index
3166 && inst
->dst
.writemask
== get_src_arg_mask(inst
->dst
, inst
->src
[2])) {
3168 inst
->op
= TGSI_OPCODE_MOV
;
3169 inst
->src
[0] = inst
->src
[1];
3173 delete [] tempWrites
;
3176 /* Replaces all references to a temporary register index with another index. */
3178 glsl_to_tgsi_visitor::rename_temp_register(int index
, int new_index
)
3180 foreach_iter(exec_list_iterator
, iter
, this->instructions
) {
3181 glsl_to_tgsi_instruction
*inst
= (glsl_to_tgsi_instruction
*)iter
.get();
3184 for (j
=0; j
< num_inst_src_regs(inst
->op
); j
++) {
3185 if (inst
->src
[j
].file
== PROGRAM_TEMPORARY
&&
3186 inst
->src
[j
].index
== index
) {
3187 inst
->src
[j
].index
= new_index
;
3191 if (inst
->dst
.file
== PROGRAM_TEMPORARY
&& inst
->dst
.index
== index
) {
3192 inst
->dst
.index
= new_index
;
3198 glsl_to_tgsi_visitor::get_first_temp_read(int index
)
3200 int depth
= 0; /* loop depth */
3201 int loop_start
= -1; /* index of the first active BGNLOOP (if any) */
3204 foreach_iter(exec_list_iterator
, iter
, this->instructions
) {
3205 glsl_to_tgsi_instruction
*inst
= (glsl_to_tgsi_instruction
*)iter
.get();
3207 for (j
=0; j
< num_inst_src_regs(inst
->op
); j
++) {
3208 if (inst
->src
[j
].file
== PROGRAM_TEMPORARY
&&
3209 inst
->src
[j
].index
== index
) {
3210 return (depth
== 0) ? i
: loop_start
;
3214 if (inst
->op
== TGSI_OPCODE_BGNLOOP
) {
3217 } else if (inst
->op
== TGSI_OPCODE_ENDLOOP
) {
3230 glsl_to_tgsi_visitor::get_first_temp_write(int index
)
3232 int depth
= 0; /* loop depth */
3233 int loop_start
= -1; /* index of the first active BGNLOOP (if any) */
3236 foreach_iter(exec_list_iterator
, iter
, this->instructions
) {
3237 glsl_to_tgsi_instruction
*inst
= (glsl_to_tgsi_instruction
*)iter
.get();
3239 if (inst
->dst
.file
== PROGRAM_TEMPORARY
&& inst
->dst
.index
== index
) {
3240 return (depth
== 0) ? i
: loop_start
;
3243 if (inst
->op
== TGSI_OPCODE_BGNLOOP
) {
3246 } else if (inst
->op
== TGSI_OPCODE_ENDLOOP
) {
3259 glsl_to_tgsi_visitor::get_last_temp_read(int index
)
3261 int depth
= 0; /* loop depth */
3262 int last
= -1; /* index of last instruction that reads the temporary */
3265 foreach_iter(exec_list_iterator
, iter
, this->instructions
) {
3266 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 last
= (depth
== 0) ? i
: -2;
3275 if (inst
->op
== TGSI_OPCODE_BGNLOOP
)
3277 else if (inst
->op
== TGSI_OPCODE_ENDLOOP
)
3278 if (--depth
== 0 && last
== -2)
3290 glsl_to_tgsi_visitor::get_last_temp_write(int index
)
3292 int depth
= 0; /* loop depth */
3293 int last
= -1; /* index of last instruction that writes to the temporary */
3296 foreach_iter(exec_list_iterator
, iter
, this->instructions
) {
3297 glsl_to_tgsi_instruction
*inst
= (glsl_to_tgsi_instruction
*)iter
.get();
3299 if (inst
->dst
.file
== PROGRAM_TEMPORARY
&& inst
->dst
.index
== index
)
3300 last
= (depth
== 0) ? i
: -2;
3302 if (inst
->op
== TGSI_OPCODE_BGNLOOP
)
3304 else if (inst
->op
== TGSI_OPCODE_ENDLOOP
)
3305 if (--depth
== 0 && last
== -2)
3317 * On a basic block basis, tracks available PROGRAM_TEMPORARY register
3318 * channels for copy propagation and updates following instructions to
3319 * use the original versions.
3321 * The glsl_to_tgsi_visitor lazily produces code assuming that this pass
3322 * will occur. As an example, a TXP production before this pass:
3324 * 0: MOV TEMP[1], INPUT[4].xyyy;
3325 * 1: MOV TEMP[1].w, INPUT[4].wwww;
3326 * 2: TXP TEMP[2], TEMP[1], texture[0], 2D;
3330 * 0: MOV TEMP[1], INPUT[4].xyyy;
3331 * 1: MOV TEMP[1].w, INPUT[4].wwww;
3332 * 2: TXP TEMP[2], INPUT[4].xyyw, texture[0], 2D;
3334 * which allows for dead code elimination on TEMP[1]'s writes.
3337 glsl_to_tgsi_visitor::copy_propagate(void)
3339 glsl_to_tgsi_instruction
**acp
= rzalloc_array(mem_ctx
,
3340 glsl_to_tgsi_instruction
*,
3341 this->next_temp
* 4);
3342 int *acp_level
= rzalloc_array(mem_ctx
, int, this->next_temp
* 4);
3345 foreach_iter(exec_list_iterator
, iter
, this->instructions
) {
3346 glsl_to_tgsi_instruction
*inst
= (glsl_to_tgsi_instruction
*)iter
.get();
3348 assert(inst
->dst
.file
!= PROGRAM_TEMPORARY
3349 || inst
->dst
.index
< this->next_temp
);
3351 /* First, do any copy propagation possible into the src regs. */
3352 for (int r
= 0; r
< 3; r
++) {
3353 glsl_to_tgsi_instruction
*first
= NULL
;
3355 int acp_base
= inst
->src
[r
].index
* 4;
3357 if (inst
->src
[r
].file
!= PROGRAM_TEMPORARY
||
3358 inst
->src
[r
].reladdr
)
3361 /* See if we can find entries in the ACP consisting of MOVs
3362 * from the same src register for all the swizzled channels
3363 * of this src register reference.
3365 for (int i
= 0; i
< 4; i
++) {
3366 int src_chan
= GET_SWZ(inst
->src
[r
].swizzle
, i
);
3367 glsl_to_tgsi_instruction
*copy_chan
= acp
[acp_base
+ src_chan
];
3374 assert(acp_level
[acp_base
+ src_chan
] <= level
);
3379 if (first
->src
[0].file
!= copy_chan
->src
[0].file
||
3380 first
->src
[0].index
!= copy_chan
->src
[0].index
) {
3388 /* We've now validated that we can copy-propagate to
3389 * replace this src register reference. Do it.
3391 inst
->src
[r
].file
= first
->src
[0].file
;
3392 inst
->src
[r
].index
= first
->src
[0].index
;
3395 for (int i
= 0; i
< 4; i
++) {
3396 int src_chan
= GET_SWZ(inst
->src
[r
].swizzle
, i
);
3397 glsl_to_tgsi_instruction
*copy_inst
= acp
[acp_base
+ src_chan
];
3398 swizzle
|= (GET_SWZ(copy_inst
->src
[0].swizzle
, src_chan
) <<
3401 inst
->src
[r
].swizzle
= swizzle
;
3406 case TGSI_OPCODE_BGNLOOP
:
3407 case TGSI_OPCODE_ENDLOOP
:
3408 /* End of a basic block, clear the ACP entirely. */
3409 memset(acp
, 0, sizeof(*acp
) * this->next_temp
* 4);
3412 case TGSI_OPCODE_IF
:
3416 case TGSI_OPCODE_ENDIF
:
3417 case TGSI_OPCODE_ELSE
:
3418 /* Clear all channels written inside the block from the ACP, but
3419 * leaving those that were not touched.
3421 for (int r
= 0; r
< this->next_temp
; r
++) {
3422 for (int c
= 0; c
< 4; c
++) {
3423 if (!acp
[4 * r
+ c
])
3426 if (acp_level
[4 * r
+ c
] >= level
)
3427 acp
[4 * r
+ c
] = NULL
;
3430 if (inst
->op
== TGSI_OPCODE_ENDIF
)
3435 /* Continuing the block, clear any written channels from
3438 if (inst
->dst
.file
== PROGRAM_TEMPORARY
&& inst
->dst
.reladdr
) {
3439 /* Any temporary might be written, so no copy propagation
3440 * across this instruction.
3442 memset(acp
, 0, sizeof(*acp
) * this->next_temp
* 4);
3443 } else if (inst
->dst
.file
== PROGRAM_OUTPUT
&&
3444 inst
->dst
.reladdr
) {
3445 /* Any output might be written, so no copy propagation
3446 * from outputs across this instruction.
3448 for (int r
= 0; r
< this->next_temp
; r
++) {
3449 for (int c
= 0; c
< 4; c
++) {
3450 if (!acp
[4 * r
+ c
])
3453 if (acp
[4 * r
+ c
]->src
[0].file
== PROGRAM_OUTPUT
)
3454 acp
[4 * r
+ c
] = NULL
;
3457 } else if (inst
->dst
.file
== PROGRAM_TEMPORARY
||
3458 inst
->dst
.file
== PROGRAM_OUTPUT
) {
3459 /* Clear where it's used as dst. */
3460 if (inst
->dst
.file
== PROGRAM_TEMPORARY
) {
3461 for (int c
= 0; c
< 4; c
++) {
3462 if (inst
->dst
.writemask
& (1 << c
)) {
3463 acp
[4 * inst
->dst
.index
+ c
] = NULL
;
3468 /* Clear where it's used as src. */
3469 for (int r
= 0; r
< this->next_temp
; r
++) {
3470 for (int c
= 0; c
< 4; c
++) {
3471 if (!acp
[4 * r
+ c
])
3474 int src_chan
= GET_SWZ(acp
[4 * r
+ c
]->src
[0].swizzle
, c
);
3476 if (acp
[4 * r
+ c
]->src
[0].file
== inst
->dst
.file
&&
3477 acp
[4 * r
+ c
]->src
[0].index
== inst
->dst
.index
&&
3478 inst
->dst
.writemask
& (1 << src_chan
))
3480 acp
[4 * r
+ c
] = NULL
;
3488 /* If this is a copy, add it to the ACP. */
3489 if (inst
->op
== TGSI_OPCODE_MOV
&&
3490 inst
->dst
.file
== PROGRAM_TEMPORARY
&&
3491 !inst
->dst
.reladdr
&&
3493 !inst
->src
[0].reladdr
&&
3494 !inst
->src
[0].negate
) {
3495 for (int i
= 0; i
< 4; i
++) {
3496 if (inst
->dst
.writemask
& (1 << i
)) {
3497 acp
[4 * inst
->dst
.index
+ i
] = inst
;
3498 acp_level
[4 * inst
->dst
.index
+ i
] = level
;
3504 ralloc_free(acp_level
);
3509 * Tracks available PROGRAM_TEMPORARY registers for dead code elimination.
3511 * The glsl_to_tgsi_visitor lazily produces code assuming that this pass
3512 * will occur. As an example, a TXP production after copy propagation but
3515 * 0: MOV TEMP[1], INPUT[4].xyyy;
3516 * 1: MOV TEMP[1].w, INPUT[4].wwww;
3517 * 2: TXP TEMP[2], INPUT[4].xyyw, texture[0], 2D;
3519 * and after this pass:
3521 * 0: TXP TEMP[2], INPUT[4].xyyw, texture[0], 2D;
3523 * FIXME: assumes that all functions are inlined (no support for BGNSUB/ENDSUB)
3524 * FIXME: doesn't eliminate all dead code inside of loops; it steps around them
3527 glsl_to_tgsi_visitor::eliminate_dead_code(void)
3531 for (i
=0; i
< this->next_temp
; i
++) {
3532 int last_read
= get_last_temp_read(i
);
3535 foreach_iter(exec_list_iterator
, iter
, this->instructions
) {
3536 glsl_to_tgsi_instruction
*inst
= (glsl_to_tgsi_instruction
*)iter
.get();
3538 if (inst
->dst
.file
== PROGRAM_TEMPORARY
&& inst
->dst
.index
== i
&&
3551 * On a basic block basis, tracks available PROGRAM_TEMPORARY registers for dead
3552 * code elimination. This is less primitive than eliminate_dead_code(), as it
3553 * is per-channel and can detect consecutive writes without a read between them
3554 * as dead code. However, there is some dead code that can be eliminated by
3555 * eliminate_dead_code() but not this function - for example, this function
3556 * cannot eliminate an instruction writing to a register that is never read and
3557 * is the only instruction writing to that register.
3559 * The glsl_to_tgsi_visitor lazily produces code assuming that this pass
3563 glsl_to_tgsi_visitor::eliminate_dead_code_advanced(void)
3565 glsl_to_tgsi_instruction
**writes
= rzalloc_array(mem_ctx
,
3566 glsl_to_tgsi_instruction
*,
3567 this->next_temp
* 4);
3568 int *write_level
= rzalloc_array(mem_ctx
, int, this->next_temp
* 4);
3572 foreach_iter(exec_list_iterator
, iter
, this->instructions
) {
3573 glsl_to_tgsi_instruction
*inst
= (glsl_to_tgsi_instruction
*)iter
.get();
3575 assert(inst
->dst
.file
!= PROGRAM_TEMPORARY
3576 || inst
->dst
.index
< this->next_temp
);
3579 case TGSI_OPCODE_BGNLOOP
:
3580 case TGSI_OPCODE_ENDLOOP
:
3581 /* End of a basic block, clear the write array entirely.
3582 * FIXME: This keeps us from killing dead code when the writes are
3583 * on either side of a loop, even when the register isn't touched
3586 memset(writes
, 0, sizeof(*writes
) * this->next_temp
* 4);
3589 case TGSI_OPCODE_ENDIF
:
3593 case TGSI_OPCODE_ELSE
:
3594 /* Clear all channels written inside the preceding if block from the
3595 * write array, but leave those that were not touched.
3597 * FIXME: This destroys opportunities to remove dead code inside of
3598 * IF blocks that are followed by an ELSE block.
3600 for (int r
= 0; r
< this->next_temp
; r
++) {
3601 for (int c
= 0; c
< 4; c
++) {
3602 if (!writes
[4 * r
+ c
])
3605 if (write_level
[4 * r
+ c
] >= level
)
3606 writes
[4 * r
+ c
] = NULL
;
3611 case TGSI_OPCODE_IF
:
3613 /* fallthrough to default case to mark the condition as read */
3616 /* Continuing the block, clear any channels from the write array that
3617 * are read by this instruction.
3619 for (unsigned i
= 0; i
< Elements(inst
->src
); i
++) {
3620 if (inst
->src
[i
].file
== PROGRAM_TEMPORARY
&& inst
->src
[i
].reladdr
){
3621 /* Any temporary might be read, so no dead code elimination
3622 * across this instruction.
3624 memset(writes
, 0, sizeof(*writes
) * this->next_temp
* 4);
3625 } else if (inst
->src
[i
].file
== PROGRAM_TEMPORARY
) {
3626 /* Clear where it's used as src. */
3627 int src_chans
= 1 << GET_SWZ(inst
->src
[i
].swizzle
, 0);
3628 src_chans
|= 1 << GET_SWZ(inst
->src
[i
].swizzle
, 1);
3629 src_chans
|= 1 << GET_SWZ(inst
->src
[i
].swizzle
, 2);
3630 src_chans
|= 1 << GET_SWZ(inst
->src
[i
].swizzle
, 3);
3632 for (int c
= 0; c
< 4; c
++) {
3633 if (src_chans
& (1 << c
)) {
3634 writes
[4 * inst
->src
[i
].index
+ c
] = NULL
;
3642 /* If this instruction writes to a temporary, add it to the write array.
3643 * If there is already an instruction in the write array for one or more
3644 * of the channels, flag that channel write as dead.
3646 if (inst
->dst
.file
== PROGRAM_TEMPORARY
&&
3647 !inst
->dst
.reladdr
&&
3649 for (int c
= 0; c
< 4; c
++) {
3650 if (inst
->dst
.writemask
& (1 << c
)) {
3651 if (writes
[4 * inst
->dst
.index
+ c
]) {
3652 if (write_level
[4 * inst
->dst
.index
+ c
] < level
)
3655 writes
[4 * inst
->dst
.index
+ c
]->dead_mask
|= (1 << c
);
3657 writes
[4 * inst
->dst
.index
+ c
] = inst
;
3658 write_level
[4 * inst
->dst
.index
+ c
] = level
;
3664 /* Anything still in the write array at this point is dead code. */
3665 for (int r
= 0; r
< this->next_temp
; r
++) {
3666 for (int c
= 0; c
< 4; c
++) {
3667 glsl_to_tgsi_instruction
*inst
= writes
[4 * r
+ c
];
3669 inst
->dead_mask
|= (1 << c
);
3673 /* Now actually remove the instructions that are completely dead and update
3674 * the writemask of other instructions with dead channels.
3676 foreach_iter(exec_list_iterator
, iter
, this->instructions
) {
3677 glsl_to_tgsi_instruction
*inst
= (glsl_to_tgsi_instruction
*)iter
.get();
3679 if (!inst
->dead_mask
|| !inst
->dst
.writemask
)
3681 else if ((inst
->dst
.writemask
& ~inst
->dead_mask
) == 0) {
3686 inst
->dst
.writemask
&= ~(inst
->dead_mask
);
3689 ralloc_free(write_level
);
3690 ralloc_free(writes
);
3695 /* Merges temporary registers together where possible to reduce the number of
3696 * registers needed to run a program.
3698 * Produces optimal code only after copy propagation and dead code elimination
3701 glsl_to_tgsi_visitor::merge_registers(void)
3703 int *last_reads
= rzalloc_array(mem_ctx
, int, this->next_temp
);
3704 int *first_writes
= rzalloc_array(mem_ctx
, int, this->next_temp
);
3707 /* Read the indices of the last read and first write to each temp register
3708 * into an array so that we don't have to traverse the instruction list as
3710 for (i
=0; i
< this->next_temp
; i
++) {
3711 last_reads
[i
] = get_last_temp_read(i
);
3712 first_writes
[i
] = get_first_temp_write(i
);
3715 /* Start looking for registers with non-overlapping usages that can be
3716 * merged together. */
3717 for (i
=0; i
< this->next_temp
; i
++) {
3718 /* Don't touch unused registers. */
3719 if (last_reads
[i
] < 0 || first_writes
[i
] < 0) continue;
3721 for (j
=0; j
< this->next_temp
; j
++) {
3722 /* Don't touch unused registers. */
3723 if (last_reads
[j
] < 0 || first_writes
[j
] < 0) continue;
3725 /* We can merge the two registers if the first write to j is after or
3726 * in the same instruction as the last read from i. Note that the
3727 * register at index i will always be used earlier or at the same time
3728 * as the register at index j. */
3729 if (first_writes
[i
] <= first_writes
[j
] &&
3730 last_reads
[i
] <= first_writes
[j
])
3732 rename_temp_register(j
, i
); /* Replace all references to j with i.*/
3734 /* Update the first_writes and last_reads arrays with the new
3735 * values for the merged register index, and mark the newly unused
3736 * register index as such. */
3737 last_reads
[i
] = last_reads
[j
];
3738 first_writes
[j
] = -1;
3744 ralloc_free(last_reads
);
3745 ralloc_free(first_writes
);
3748 /* Reassign indices to temporary registers by reusing unused indices created
3749 * by optimization passes. */
3751 glsl_to_tgsi_visitor::renumber_registers(void)
3756 for (i
=0; i
< this->next_temp
; i
++) {
3757 if (get_first_temp_read(i
) < 0) continue;
3759 rename_temp_register(i
, new_index
);
3763 this->next_temp
= new_index
;
3767 * Returns a fragment program which implements the current pixel transfer ops.
3768 * Based on get_pixel_transfer_program in st_atom_pixeltransfer.c.
3771 get_pixel_transfer_visitor(struct st_fragment_program
*fp
,
3772 glsl_to_tgsi_visitor
*original
,
3773 int scale_and_bias
, int pixel_maps
)
3775 glsl_to_tgsi_visitor
*v
= new glsl_to_tgsi_visitor();
3776 struct st_context
*st
= st_context(original
->ctx
);
3777 struct gl_program
*prog
= &fp
->Base
.Base
;
3778 struct gl_program_parameter_list
*params
= _mesa_new_parameter_list();
3779 st_src_reg coord
, src0
;
3781 glsl_to_tgsi_instruction
*inst
;
3783 /* Copy attributes of the glsl_to_tgsi_visitor in the original shader. */
3784 v
->ctx
= original
->ctx
;
3786 v
->glsl_version
= original
->glsl_version
;
3787 v
->native_integers
= original
->native_integers
;
3788 v
->options
= original
->options
;
3789 v
->next_temp
= original
->next_temp
;
3790 v
->num_address_regs
= original
->num_address_regs
;
3791 v
->samplers_used
= prog
->SamplersUsed
= original
->samplers_used
;
3792 v
->indirect_addr_temps
= original
->indirect_addr_temps
;
3793 v
->indirect_addr_consts
= original
->indirect_addr_consts
;
3794 memcpy(&v
->immediates
, &original
->immediates
, sizeof(v
->immediates
));
3797 * Get initial pixel color from the texture.
3798 * TEX colorTemp, fragment.texcoord[0], texture[0], 2D;
3800 coord
= st_src_reg(PROGRAM_INPUT
, FRAG_ATTRIB_TEX0
, glsl_type::vec2_type
);
3801 src0
= v
->get_temp(glsl_type::vec4_type
);
3802 dst0
= st_dst_reg(src0
);
3803 inst
= v
->emit(NULL
, TGSI_OPCODE_TEX
, dst0
, coord
);
3805 inst
->tex_target
= TEXTURE_2D_INDEX
;
3807 prog
->InputsRead
|= (1 << FRAG_ATTRIB_TEX0
);
3808 prog
->SamplersUsed
|= (1 << 0); /* mark sampler 0 as used */
3809 v
->samplers_used
|= (1 << 0);
3811 if (scale_and_bias
) {
3812 static const gl_state_index scale_state
[STATE_LENGTH
] =
3813 { STATE_INTERNAL
, STATE_PT_SCALE
,
3814 (gl_state_index
) 0, (gl_state_index
) 0, (gl_state_index
) 0 };
3815 static const gl_state_index bias_state
[STATE_LENGTH
] =
3816 { STATE_INTERNAL
, STATE_PT_BIAS
,
3817 (gl_state_index
) 0, (gl_state_index
) 0, (gl_state_index
) 0 };
3818 GLint scale_p
, bias_p
;
3819 st_src_reg scale
, bias
;
3821 scale_p
= _mesa_add_state_reference(params
, scale_state
);
3822 bias_p
= _mesa_add_state_reference(params
, bias_state
);
3824 /* MAD colorTemp, colorTemp, scale, bias; */
3825 scale
= st_src_reg(PROGRAM_STATE_VAR
, scale_p
, GLSL_TYPE_FLOAT
);
3826 bias
= st_src_reg(PROGRAM_STATE_VAR
, bias_p
, GLSL_TYPE_FLOAT
);
3827 inst
= v
->emit(NULL
, TGSI_OPCODE_MAD
, dst0
, src0
, scale
, bias
);
3831 st_src_reg temp
= v
->get_temp(glsl_type::vec4_type
);
3832 st_dst_reg temp_dst
= st_dst_reg(temp
);
3834 assert(st
->pixel_xfer
.pixelmap_texture
);
3836 /* With a little effort, we can do four pixel map look-ups with
3837 * two TEX instructions:
3840 /* TEX temp.rg, colorTemp.rgba, texture[1], 2D; */
3841 temp_dst
.writemask
= WRITEMASK_XY
; /* write R,G */
3842 inst
= v
->emit(NULL
, TGSI_OPCODE_TEX
, temp_dst
, src0
);
3844 inst
->tex_target
= TEXTURE_2D_INDEX
;
3846 /* TEX temp.ba, colorTemp.baba, texture[1], 2D; */
3847 src0
.swizzle
= MAKE_SWIZZLE4(SWIZZLE_Z
, SWIZZLE_W
, SWIZZLE_Z
, SWIZZLE_W
);
3848 temp_dst
.writemask
= WRITEMASK_ZW
; /* write B,A */
3849 inst
= v
->emit(NULL
, TGSI_OPCODE_TEX
, temp_dst
, src0
);
3851 inst
->tex_target
= TEXTURE_2D_INDEX
;
3853 prog
->SamplersUsed
|= (1 << 1); /* mark sampler 1 as used */
3854 v
->samplers_used
|= (1 << 1);
3856 /* MOV colorTemp, temp; */
3857 inst
= v
->emit(NULL
, TGSI_OPCODE_MOV
, dst0
, temp
);
3860 /* Now copy the instructions from the original glsl_to_tgsi_visitor into the
3862 foreach_iter(exec_list_iterator
, iter
, original
->instructions
) {
3863 glsl_to_tgsi_instruction
*inst
= (glsl_to_tgsi_instruction
*)iter
.get();
3864 st_src_reg src_regs
[3];
3866 if (inst
->dst
.file
== PROGRAM_OUTPUT
)
3867 prog
->OutputsWritten
|= BITFIELD64_BIT(inst
->dst
.index
);
3869 for (int i
=0; i
<3; i
++) {
3870 src_regs
[i
] = inst
->src
[i
];
3871 if (src_regs
[i
].file
== PROGRAM_INPUT
&&
3872 src_regs
[i
].index
== FRAG_ATTRIB_COL0
)
3874 src_regs
[i
].file
= PROGRAM_TEMPORARY
;
3875 src_regs
[i
].index
= src0
.index
;
3877 else if (src_regs
[i
].file
== PROGRAM_INPUT
)
3878 prog
->InputsRead
|= (1 << src_regs
[i
].index
);
3881 v
->emit(NULL
, inst
->op
, inst
->dst
, src_regs
[0], src_regs
[1], src_regs
[2]);
3884 /* Make modifications to fragment program info. */
3885 prog
->Parameters
= _mesa_combine_parameter_lists(params
,
3886 original
->prog
->Parameters
);
3887 _mesa_free_parameter_list(params
);
3888 count_resources(v
, prog
);
3889 fp
->glsl_to_tgsi
= v
;
3893 * Make fragment program for glBitmap:
3894 * Sample the texture and kill the fragment if the bit is 0.
3895 * This program will be combined with the user's fragment program.
3897 * Based on make_bitmap_fragment_program in st_cb_bitmap.c.
3900 get_bitmap_visitor(struct st_fragment_program
*fp
,
3901 glsl_to_tgsi_visitor
*original
, int samplerIndex
)
3903 glsl_to_tgsi_visitor
*v
= new glsl_to_tgsi_visitor();
3904 struct st_context
*st
= st_context(original
->ctx
);
3905 struct gl_program
*prog
= &fp
->Base
.Base
;
3906 st_src_reg coord
, src0
;
3908 glsl_to_tgsi_instruction
*inst
;
3910 /* Copy attributes of the glsl_to_tgsi_visitor in the original shader. */
3911 v
->ctx
= original
->ctx
;
3913 v
->glsl_version
= original
->glsl_version
;
3914 v
->native_integers
= original
->native_integers
;
3915 v
->options
= original
->options
;
3916 v
->next_temp
= original
->next_temp
;
3917 v
->num_address_regs
= original
->num_address_regs
;
3918 v
->samplers_used
= prog
->SamplersUsed
= original
->samplers_used
;
3919 v
->indirect_addr_temps
= original
->indirect_addr_temps
;
3920 v
->indirect_addr_consts
= original
->indirect_addr_consts
;
3921 memcpy(&v
->immediates
, &original
->immediates
, sizeof(v
->immediates
));
3923 /* TEX tmp0, fragment.texcoord[0], texture[0], 2D; */
3924 coord
= st_src_reg(PROGRAM_INPUT
, FRAG_ATTRIB_TEX0
, glsl_type::vec2_type
);
3925 src0
= v
->get_temp(glsl_type::vec4_type
);
3926 dst0
= st_dst_reg(src0
);
3927 inst
= v
->emit(NULL
, TGSI_OPCODE_TEX
, dst0
, coord
);
3928 inst
->sampler
= samplerIndex
;
3929 inst
->tex_target
= TEXTURE_2D_INDEX
;
3931 prog
->InputsRead
|= (1 << FRAG_ATTRIB_TEX0
);
3932 prog
->SamplersUsed
|= (1 << samplerIndex
); /* mark sampler as used */
3933 v
->samplers_used
|= (1 << samplerIndex
);
3935 /* KIL if -tmp0 < 0 # texel=0 -> keep / texel=0 -> discard */
3936 src0
.negate
= NEGATE_XYZW
;
3937 if (st
->bitmap
.tex_format
== PIPE_FORMAT_L8_UNORM
)
3938 src0
.swizzle
= SWIZZLE_XXXX
;
3939 inst
= v
->emit(NULL
, TGSI_OPCODE_KIL
, undef_dst
, src0
);
3941 /* Now copy the instructions from the original glsl_to_tgsi_visitor into the
3943 foreach_iter(exec_list_iterator
, iter
, original
->instructions
) {
3944 glsl_to_tgsi_instruction
*inst
= (glsl_to_tgsi_instruction
*)iter
.get();
3945 st_src_reg src_regs
[3];
3947 if (inst
->dst
.file
== PROGRAM_OUTPUT
)
3948 prog
->OutputsWritten
|= BITFIELD64_BIT(inst
->dst
.index
);
3950 for (int i
=0; i
<3; i
++) {
3951 src_regs
[i
] = inst
->src
[i
];
3952 if (src_regs
[i
].file
== PROGRAM_INPUT
)
3953 prog
->InputsRead
|= (1 << src_regs
[i
].index
);
3956 v
->emit(NULL
, inst
->op
, inst
->dst
, src_regs
[0], src_regs
[1], src_regs
[2]);
3959 /* Make modifications to fragment program info. */
3960 prog
->Parameters
= _mesa_clone_parameter_list(original
->prog
->Parameters
);
3961 count_resources(v
, prog
);
3962 fp
->glsl_to_tgsi
= v
;
3965 /* ------------------------- TGSI conversion stuff -------------------------- */
3967 unsigned branch_target
;
3972 * Intermediate state used during shader translation.
3974 struct st_translate
{
3975 struct ureg_program
*ureg
;
3977 struct ureg_dst temps
[MAX_TEMPS
];
3978 struct ureg_src
*constants
;
3979 struct ureg_src
*immediates
;
3980 struct ureg_dst outputs
[PIPE_MAX_SHADER_OUTPUTS
];
3981 struct ureg_src inputs
[PIPE_MAX_SHADER_INPUTS
];
3982 struct ureg_dst address
[1];
3983 struct ureg_src samplers
[PIPE_MAX_SAMPLERS
];
3984 struct ureg_src systemValues
[SYSTEM_VALUE_MAX
];
3986 /* Extra info for handling point size clamping in vertex shader */
3987 struct ureg_dst pointSizeResult
; /**< Actual point size output register */
3988 struct ureg_src pointSizeConst
; /**< Point size range constant register */
3989 GLint pointSizeOutIndex
; /**< Temp point size output register */
3990 GLboolean prevInstWrotePointSize
;
3992 const GLuint
*inputMapping
;
3993 const GLuint
*outputMapping
;
3995 /* For every instruction that contains a label (eg CALL), keep
3996 * details so that we can go back afterwards and emit the correct
3997 * tgsi instruction number for each label.
3999 struct label
*labels
;
4000 unsigned labels_size
;
4001 unsigned labels_count
;
4003 /* Keep a record of the tgsi instruction number that each mesa
4004 * instruction starts at, will be used to fix up labels after
4009 unsigned insn_count
;
4011 unsigned procType
; /**< TGSI_PROCESSOR_VERTEX/FRAGMENT */
4016 /** Map Mesa's SYSTEM_VALUE_x to TGSI_SEMANTIC_x */
4017 static unsigned mesa_sysval_to_semantic
[SYSTEM_VALUE_MAX
] = {
4019 TGSI_SEMANTIC_INSTANCEID
4023 * Make note of a branch to a label in the TGSI code.
4024 * After we've emitted all instructions, we'll go over the list
4025 * of labels built here and patch the TGSI code with the actual
4026 * location of each label.
4028 static unsigned *get_label(struct st_translate
*t
, unsigned branch_target
)
4032 if (t
->labels_count
+ 1 >= t
->labels_size
) {
4033 t
->labels_size
= 1 << (util_logbase2(t
->labels_size
) + 1);
4034 t
->labels
= (struct label
*)realloc(t
->labels
,
4035 t
->labels_size
* sizeof(struct label
));
4036 if (t
->labels
== NULL
) {
4037 static unsigned dummy
;
4043 i
= t
->labels_count
++;
4044 t
->labels
[i
].branch_target
= branch_target
;
4045 return &t
->labels
[i
].token
;
4049 * Called prior to emitting the TGSI code for each instruction.
4050 * Allocate additional space for instructions if needed.
4051 * Update the insn[] array so the next glsl_to_tgsi_instruction points to
4052 * the next TGSI instruction.
4054 static void set_insn_start(struct st_translate
*t
, unsigned start
)
4056 if (t
->insn_count
+ 1 >= t
->insn_size
) {
4057 t
->insn_size
= 1 << (util_logbase2(t
->insn_size
) + 1);
4058 t
->insn
= (unsigned *)realloc(t
->insn
, t
->insn_size
* sizeof(t
->insn
[0]));
4059 if (t
->insn
== NULL
) {
4065 t
->insn
[t
->insn_count
++] = start
;
4069 * Map a glsl_to_tgsi constant/immediate to a TGSI immediate.
4071 static struct ureg_src
4072 emit_immediate(struct st_translate
*t
,
4073 gl_constant_value values
[4],
4076 struct ureg_program
*ureg
= t
->ureg
;
4081 return ureg_DECL_immediate(ureg
, &values
[0].f
, size
);
4083 return ureg_DECL_immediate_int(ureg
, &values
[0].i
, size
);
4084 case GL_UNSIGNED_INT
:
4086 return ureg_DECL_immediate_uint(ureg
, &values
[0].u
, size
);
4088 assert(!"should not get here - type must be float, int, uint, or bool");
4089 return ureg_src_undef();
4094 * Map a glsl_to_tgsi dst register to a TGSI ureg_dst register.
4096 static struct ureg_dst
4097 dst_register(struct st_translate
*t
,
4098 gl_register_file file
,
4102 case PROGRAM_UNDEFINED
:
4103 return ureg_dst_undef();
4105 case PROGRAM_TEMPORARY
:
4106 if (ureg_dst_is_undef(t
->temps
[index
]))
4107 t
->temps
[index
] = ureg_DECL_temporary(t
->ureg
);
4109 return t
->temps
[index
];
4111 case PROGRAM_OUTPUT
:
4112 if (t
->procType
== TGSI_PROCESSOR_VERTEX
&& index
== VERT_RESULT_PSIZ
)
4113 t
->prevInstWrotePointSize
= GL_TRUE
;
4115 if (t
->procType
== TGSI_PROCESSOR_VERTEX
)
4116 assert(index
< VERT_RESULT_MAX
);
4117 else if (t
->procType
== TGSI_PROCESSOR_FRAGMENT
)
4118 assert(index
< FRAG_RESULT_MAX
);
4120 assert(index
< GEOM_RESULT_MAX
);
4122 assert(t
->outputMapping
[index
] < Elements(t
->outputs
));
4124 return t
->outputs
[t
->outputMapping
[index
]];
4126 case PROGRAM_ADDRESS
:
4127 return t
->address
[index
];
4130 assert(!"unknown dst register file");
4131 return ureg_dst_undef();
4136 * Map a glsl_to_tgsi src register to a TGSI ureg_src register.
4138 static struct ureg_src
4139 src_register(struct st_translate
*t
,
4140 gl_register_file file
,
4144 case PROGRAM_UNDEFINED
:
4145 return ureg_src_undef();
4147 case PROGRAM_TEMPORARY
:
4149 assert(index
< Elements(t
->temps
));
4150 if (ureg_dst_is_undef(t
->temps
[index
]))
4151 t
->temps
[index
] = ureg_DECL_temporary(t
->ureg
);
4152 return ureg_src(t
->temps
[index
]);
4154 case PROGRAM_NAMED_PARAM
:
4155 case PROGRAM_ENV_PARAM
:
4156 case PROGRAM_LOCAL_PARAM
:
4157 case PROGRAM_UNIFORM
:
4159 return t
->constants
[index
];
4160 case PROGRAM_STATE_VAR
:
4161 case PROGRAM_CONSTANT
: /* ie, immediate */
4163 return ureg_DECL_constant(t
->ureg
, 0);
4165 return t
->constants
[index
];
4167 case PROGRAM_IMMEDIATE
:
4168 return t
->immediates
[index
];
4171 assert(t
->inputMapping
[index
] < Elements(t
->inputs
));
4172 return t
->inputs
[t
->inputMapping
[index
]];
4174 case PROGRAM_OUTPUT
:
4175 assert(t
->outputMapping
[index
] < Elements(t
->outputs
));
4176 return ureg_src(t
->outputs
[t
->outputMapping
[index
]]); /* not needed? */
4178 case PROGRAM_ADDRESS
:
4179 return ureg_src(t
->address
[index
]);
4181 case PROGRAM_SYSTEM_VALUE
:
4182 assert(index
< Elements(t
->systemValues
));
4183 return t
->systemValues
[index
];
4186 assert(!"unknown src register file");
4187 return ureg_src_undef();
4192 * Create a TGSI ureg_dst register from an st_dst_reg.
4194 static struct ureg_dst
4195 translate_dst(struct st_translate
*t
,
4196 const st_dst_reg
*dst_reg
,
4199 struct ureg_dst dst
= dst_register(t
,
4203 dst
= ureg_writemask(dst
, dst_reg
->writemask
);
4206 dst
= ureg_saturate(dst
);
4208 if (dst_reg
->reladdr
!= NULL
)
4209 dst
= ureg_dst_indirect(dst
, ureg_src(t
->address
[0]));
4215 * Create a TGSI ureg_src register from an st_src_reg.
4217 static struct ureg_src
4218 translate_src(struct st_translate
*t
, const st_src_reg
*src_reg
)
4220 struct ureg_src src
= src_register(t
, src_reg
->file
, src_reg
->index
);
4222 src
= ureg_swizzle(src
,
4223 GET_SWZ(src_reg
->swizzle
, 0) & 0x3,
4224 GET_SWZ(src_reg
->swizzle
, 1) & 0x3,
4225 GET_SWZ(src_reg
->swizzle
, 2) & 0x3,
4226 GET_SWZ(src_reg
->swizzle
, 3) & 0x3);
4228 if ((src_reg
->negate
& 0xf) == NEGATE_XYZW
)
4229 src
= ureg_negate(src
);
4231 if (src_reg
->reladdr
!= NULL
) {
4232 /* Normally ureg_src_indirect() would be used here, but a stupid compiler
4233 * bug in g++ makes ureg_src_indirect (an inline C function) erroneously
4234 * set the bit for src.Negate. So we have to do the operation manually
4235 * here to work around the compiler's problems. */
4236 /*src = ureg_src_indirect(src, ureg_src(t->address[0]));*/
4237 struct ureg_src addr
= ureg_src(t
->address
[0]);
4239 src
.IndirectFile
= addr
.File
;
4240 src
.IndirectIndex
= addr
.Index
;
4241 src
.IndirectSwizzle
= addr
.SwizzleX
;
4243 if (src_reg
->file
!= PROGRAM_INPUT
&&
4244 src_reg
->file
!= PROGRAM_OUTPUT
) {
4245 /* If src_reg->index was negative, it was set to zero in
4246 * src_register(). Reassign it now. But don't do this
4247 * for input/output regs since they get remapped while
4248 * const buffers don't.
4250 src
.Index
= src_reg
->index
;
4257 static struct tgsi_texture_offset
4258 translate_tex_offset(struct st_translate
*t
,
4259 const struct tgsi_texture_offset
*in_offset
)
4261 struct tgsi_texture_offset offset
;
4263 assert(in_offset
->File
== PROGRAM_IMMEDIATE
);
4265 offset
.File
= TGSI_FILE_IMMEDIATE
;
4266 offset
.Index
= in_offset
->Index
;
4267 offset
.SwizzleX
= in_offset
->SwizzleX
;
4268 offset
.SwizzleY
= in_offset
->SwizzleY
;
4269 offset
.SwizzleZ
= in_offset
->SwizzleZ
;
4275 compile_tgsi_instruction(struct st_translate
*t
,
4276 const glsl_to_tgsi_instruction
*inst
)
4278 struct ureg_program
*ureg
= t
->ureg
;
4280 struct ureg_dst dst
[1];
4281 struct ureg_src src
[4];
4282 struct tgsi_texture_offset texoffsets
[MAX_GLSL_TEXTURE_OFFSET
];
4287 num_dst
= num_inst_dst_regs(inst
->op
);
4288 num_src
= num_inst_src_regs(inst
->op
);
4291 dst
[0] = translate_dst(t
,
4295 for (i
= 0; i
< num_src
; i
++)
4296 src
[i
] = translate_src(t
, &inst
->src
[i
]);
4299 case TGSI_OPCODE_BGNLOOP
:
4300 case TGSI_OPCODE_CAL
:
4301 case TGSI_OPCODE_ELSE
:
4302 case TGSI_OPCODE_ENDLOOP
:
4303 case TGSI_OPCODE_IF
:
4304 assert(num_dst
== 0);
4305 ureg_label_insn(ureg
,
4309 inst
->op
== TGSI_OPCODE_CAL
? inst
->function
->sig_id
: 0));
4312 case TGSI_OPCODE_TEX
:
4313 case TGSI_OPCODE_TXB
:
4314 case TGSI_OPCODE_TXD
:
4315 case TGSI_OPCODE_TXL
:
4316 case TGSI_OPCODE_TXP
:
4317 case TGSI_OPCODE_TXQ
:
4318 case TGSI_OPCODE_TXF
:
4319 src
[num_src
++] = t
->samplers
[inst
->sampler
];
4320 for (i
= 0; i
< inst
->tex_offset_num_offset
; i
++) {
4321 texoffsets
[i
] = translate_tex_offset(t
, &inst
->tex_offsets
[i
]);
4326 translate_texture_target(inst
->tex_target
, inst
->tex_shadow
),
4327 texoffsets
, inst
->tex_offset_num_offset
,
4331 case TGSI_OPCODE_SCS
:
4332 dst
[0] = ureg_writemask(dst
[0], TGSI_WRITEMASK_XY
);
4333 ureg_insn(ureg
, inst
->op
, dst
, num_dst
, src
, num_src
);
4346 * Emit the TGSI instructions for inverting and adjusting WPOS.
4347 * This code is unavoidable because it also depends on whether
4348 * a FBO is bound (STATE_FB_WPOS_Y_TRANSFORM).
4351 emit_wpos_adjustment( struct st_translate
*t
,
4352 const struct gl_program
*program
,
4354 GLfloat adjX
, GLfloat adjY
[2])
4356 struct ureg_program
*ureg
= t
->ureg
;
4358 /* Fragment program uses fragment position input.
4359 * Need to replace instances of INPUT[WPOS] with temp T
4360 * where T = INPUT[WPOS] by y is inverted.
4362 static const gl_state_index wposTransformState
[STATE_LENGTH
]
4363 = { STATE_INTERNAL
, STATE_FB_WPOS_Y_TRANSFORM
,
4364 (gl_state_index
)0, (gl_state_index
)0, (gl_state_index
)0 };
4366 /* XXX: note we are modifying the incoming shader here! Need to
4367 * do this before emitting the constant decls below, or this
4370 unsigned wposTransConst
= _mesa_add_state_reference(program
->Parameters
,
4371 wposTransformState
);
4373 struct ureg_src wpostrans
= ureg_DECL_constant( ureg
, wposTransConst
);
4374 struct ureg_dst wpos_temp
= ureg_DECL_temporary( ureg
);
4375 struct ureg_src wpos_input
= t
->inputs
[t
->inputMapping
[FRAG_ATTRIB_WPOS
]];
4377 /* First, apply the coordinate shift: */
4378 if (adjX
|| adjY
[0] || adjY
[1]) {
4379 if (adjY
[0] != adjY
[1]) {
4380 /* Adjust the y coordinate by adjY[1] or adjY[0] respectively
4381 * depending on whether inversion is actually going to be applied
4382 * or not, which is determined by testing against the inversion
4383 * state variable used below, which will be either +1 or -1.
4385 struct ureg_dst adj_temp
= ureg_DECL_temporary(ureg
);
4387 ureg_CMP(ureg
, adj_temp
,
4388 ureg_scalar(wpostrans
, invert
? 2 : 0),
4389 ureg_imm4f(ureg
, adjX
, adjY
[0], 0.0f
, 0.0f
),
4390 ureg_imm4f(ureg
, adjX
, adjY
[1], 0.0f
, 0.0f
));
4391 ureg_ADD(ureg
, wpos_temp
, wpos_input
, ureg_src(adj_temp
));
4393 ureg_ADD(ureg
, wpos_temp
, wpos_input
,
4394 ureg_imm4f(ureg
, adjX
, adjY
[0], 0.0f
, 0.0f
));
4396 wpos_input
= ureg_src(wpos_temp
);
4398 /* MOV wpos_temp, input[wpos]
4400 ureg_MOV( ureg
, wpos_temp
, wpos_input
);
4403 /* Now the conditional y flip: STATE_FB_WPOS_Y_TRANSFORM.xy/zw will be
4404 * inversion/identity, or the other way around if we're drawing to an FBO.
4407 /* MAD wpos_temp.y, wpos_input, wpostrans.xxxx, wpostrans.yyyy
4410 ureg_writemask(wpos_temp
, TGSI_WRITEMASK_Y
),
4412 ureg_scalar(wpostrans
, 0),
4413 ureg_scalar(wpostrans
, 1));
4415 /* MAD wpos_temp.y, wpos_input, wpostrans.zzzz, wpostrans.wwww
4418 ureg_writemask(wpos_temp
, TGSI_WRITEMASK_Y
),
4420 ureg_scalar(wpostrans
, 2),
4421 ureg_scalar(wpostrans
, 3));
4424 /* Use wpos_temp as position input from here on:
4426 t
->inputs
[t
->inputMapping
[FRAG_ATTRIB_WPOS
]] = ureg_src(wpos_temp
);
4431 * Emit fragment position/ooordinate code.
4434 emit_wpos(struct st_context
*st
,
4435 struct st_translate
*t
,
4436 const struct gl_program
*program
,
4437 struct ureg_program
*ureg
)
4439 const struct gl_fragment_program
*fp
=
4440 (const struct gl_fragment_program
*) program
;
4441 struct pipe_screen
*pscreen
= st
->pipe
->screen
;
4442 GLfloat adjX
= 0.0f
;
4443 GLfloat adjY
[2] = { 0.0f
, 0.0f
};
4444 boolean invert
= FALSE
;
4446 /* Query the pixel center conventions supported by the pipe driver and set
4447 * adjX, adjY to help out if it cannot handle the requested one internally.
4449 * The bias of the y-coordinate depends on whether y-inversion takes place
4450 * (adjY[1]) or not (adjY[0]), which is in turn dependent on whether we are
4451 * drawing to an FBO (causes additional inversion), and whether the the pipe
4452 * driver origin and the requested origin differ (the latter condition is
4453 * stored in the 'invert' variable).
4455 * For height = 100 (i = integer, h = half-integer, l = lower, u = upper):
4457 * center shift only:
4462 * l,i -> u,i: ( 0.0 + 1.0) * -1 + 100 = 99
4463 * l,h -> u,h: ( 0.5 + 0.0) * -1 + 100 = 99.5
4464 * u,i -> l,i: (99.0 + 1.0) * -1 + 100 = 0
4465 * u,h -> l,h: (99.5 + 0.0) * -1 + 100 = 0.5
4467 * inversion and center shift:
4468 * l,i -> u,h: ( 0.0 + 0.5) * -1 + 100 = 99.5
4469 * l,h -> u,i: ( 0.5 + 0.5) * -1 + 100 = 99
4470 * u,i -> l,h: (99.0 + 0.5) * -1 + 100 = 0.5
4471 * u,h -> l,i: (99.5 + 0.5) * -1 + 100 = 0
4473 if (fp
->OriginUpperLeft
) {
4474 /* Fragment shader wants origin in upper-left */
4475 if (pscreen
->get_param(pscreen
, PIPE_CAP_TGSI_FS_COORD_ORIGIN_UPPER_LEFT
)) {
4476 /* the driver supports upper-left origin */
4478 else if (pscreen
->get_param(pscreen
, PIPE_CAP_TGSI_FS_COORD_ORIGIN_LOWER_LEFT
)) {
4479 /* the driver supports lower-left origin, need to invert Y */
4480 ureg_property_fs_coord_origin(ureg
, TGSI_FS_COORD_ORIGIN_LOWER_LEFT
);
4487 /* Fragment shader wants origin in lower-left */
4488 if (pscreen
->get_param(pscreen
, PIPE_CAP_TGSI_FS_COORD_ORIGIN_LOWER_LEFT
))
4489 /* the driver supports lower-left origin */
4490 ureg_property_fs_coord_origin(ureg
, TGSI_FS_COORD_ORIGIN_LOWER_LEFT
);
4491 else if (pscreen
->get_param(pscreen
, PIPE_CAP_TGSI_FS_COORD_ORIGIN_UPPER_LEFT
))
4492 /* the driver supports upper-left origin, need to invert Y */
4498 if (fp
->PixelCenterInteger
) {
4499 /* Fragment shader wants pixel center integer */
4500 if (pscreen
->get_param(pscreen
, PIPE_CAP_TGSI_FS_COORD_PIXEL_CENTER_INTEGER
)) {
4501 /* the driver supports pixel center integer */
4503 ureg_property_fs_coord_pixel_center(ureg
, TGSI_FS_COORD_PIXEL_CENTER_INTEGER
);
4505 else if (pscreen
->get_param(pscreen
, PIPE_CAP_TGSI_FS_COORD_PIXEL_CENTER_HALF_INTEGER
)) {
4506 /* the driver supports pixel center half integer, need to bias X,Y */
4515 /* Fragment shader wants pixel center half integer */
4516 if (pscreen
->get_param(pscreen
, PIPE_CAP_TGSI_FS_COORD_PIXEL_CENTER_HALF_INTEGER
)) {
4517 /* the driver supports pixel center half integer */
4519 else if (pscreen
->get_param(pscreen
, PIPE_CAP_TGSI_FS_COORD_PIXEL_CENTER_INTEGER
)) {
4520 /* the driver supports pixel center integer, need to bias X,Y */
4521 adjX
= adjY
[0] = adjY
[1] = 0.5f
;
4522 ureg_property_fs_coord_pixel_center(ureg
, TGSI_FS_COORD_PIXEL_CENTER_INTEGER
);
4528 /* we invert after adjustment so that we avoid the MOV to temporary,
4529 * and reuse the adjustment ADD instead */
4530 emit_wpos_adjustment(t
, program
, invert
, adjX
, adjY
);
4534 * OpenGL's fragment gl_FrontFace input is 1 for front-facing, 0 for back.
4535 * TGSI uses +1 for front, -1 for back.
4536 * This function converts the TGSI value to the GL value. Simply clamping/
4537 * saturating the value to [0,1] does the job.
4540 emit_face_var(struct st_translate
*t
)
4542 struct ureg_program
*ureg
= t
->ureg
;
4543 struct ureg_dst face_temp
= ureg_DECL_temporary(ureg
);
4544 struct ureg_src face_input
= t
->inputs
[t
->inputMapping
[FRAG_ATTRIB_FACE
]];
4546 /* MOV_SAT face_temp, input[face] */
4547 face_temp
= ureg_saturate(face_temp
);
4548 ureg_MOV(ureg
, face_temp
, face_input
);
4550 /* Use face_temp as face input from here on: */
4551 t
->inputs
[t
->inputMapping
[FRAG_ATTRIB_FACE
]] = ureg_src(face_temp
);
4555 emit_edgeflags(struct st_translate
*t
)
4557 struct ureg_program
*ureg
= t
->ureg
;
4558 struct ureg_dst edge_dst
= t
->outputs
[t
->outputMapping
[VERT_RESULT_EDGE
]];
4559 struct ureg_src edge_src
= t
->inputs
[t
->inputMapping
[VERT_ATTRIB_EDGEFLAG
]];
4561 ureg_MOV(ureg
, edge_dst
, edge_src
);
4565 * Translate intermediate IR (glsl_to_tgsi_instruction) to TGSI format.
4566 * \param program the program to translate
4567 * \param numInputs number of input registers used
4568 * \param inputMapping maps Mesa fragment program inputs to TGSI generic
4570 * \param inputSemanticName the TGSI_SEMANTIC flag for each input
4571 * \param inputSemanticIndex the semantic index (ex: which texcoord) for
4573 * \param interpMode the TGSI_INTERPOLATE_LINEAR/PERSP mode for each input
4574 * \param numOutputs number of output registers used
4575 * \param outputMapping maps Mesa fragment program outputs to TGSI
4577 * \param outputSemanticName the TGSI_SEMANTIC flag for each output
4578 * \param outputSemanticIndex the semantic index (ex: which texcoord) for
4581 * \return PIPE_OK or PIPE_ERROR_OUT_OF_MEMORY
4583 extern "C" enum pipe_error
4584 st_translate_program(
4585 struct gl_context
*ctx
,
4587 struct ureg_program
*ureg
,
4588 glsl_to_tgsi_visitor
*program
,
4589 const struct gl_program
*proginfo
,
4591 const GLuint inputMapping
[],
4592 const ubyte inputSemanticName
[],
4593 const ubyte inputSemanticIndex
[],
4594 const GLuint interpMode
[],
4596 const GLuint outputMapping
[],
4597 const ubyte outputSemanticName
[],
4598 const ubyte outputSemanticIndex
[],
4599 boolean passthrough_edgeflags
)
4601 struct st_translate
*t
;
4603 enum pipe_error ret
= PIPE_OK
;
4605 assert(numInputs
<= Elements(t
->inputs
));
4606 assert(numOutputs
<= Elements(t
->outputs
));
4608 t
= CALLOC_STRUCT(st_translate
);
4610 ret
= PIPE_ERROR_OUT_OF_MEMORY
;
4614 memset(t
, 0, sizeof *t
);
4616 t
->procType
= procType
;
4617 t
->inputMapping
= inputMapping
;
4618 t
->outputMapping
= outputMapping
;
4620 t
->pointSizeOutIndex
= -1;
4621 t
->prevInstWrotePointSize
= GL_FALSE
;
4624 * Declare input attributes.
4626 if (procType
== TGSI_PROCESSOR_FRAGMENT
) {
4627 for (i
= 0; i
< numInputs
; i
++) {
4628 t
->inputs
[i
] = ureg_DECL_fs_input(ureg
,
4629 inputSemanticName
[i
],
4630 inputSemanticIndex
[i
],
4634 if (proginfo
->InputsRead
& FRAG_BIT_WPOS
) {
4635 /* Must do this after setting up t->inputs, and before
4636 * emitting constant references, below:
4638 emit_wpos(st_context(ctx
), t
, proginfo
, ureg
);
4641 if (proginfo
->InputsRead
& FRAG_BIT_FACE
)
4645 * Declare output attributes.
4647 for (i
= 0; i
< numOutputs
; i
++) {
4648 switch (outputSemanticName
[i
]) {
4649 case TGSI_SEMANTIC_POSITION
:
4650 t
->outputs
[i
] = ureg_DECL_output(ureg
,
4651 TGSI_SEMANTIC_POSITION
, /* Z/Depth */
4652 outputSemanticIndex
[i
]);
4653 t
->outputs
[i
] = ureg_writemask(t
->outputs
[i
], TGSI_WRITEMASK_Z
);
4655 case TGSI_SEMANTIC_STENCIL
:
4656 t
->outputs
[i
] = ureg_DECL_output(ureg
,
4657 TGSI_SEMANTIC_STENCIL
, /* Stencil */
4658 outputSemanticIndex
[i
]);
4659 t
->outputs
[i
] = ureg_writemask(t
->outputs
[i
], TGSI_WRITEMASK_Y
);
4661 case TGSI_SEMANTIC_COLOR
:
4662 t
->outputs
[i
] = ureg_DECL_output(ureg
,
4663 TGSI_SEMANTIC_COLOR
,
4664 outputSemanticIndex
[i
]);
4667 assert(!"fragment shader outputs must be POSITION/STENCIL/COLOR");
4668 ret
= PIPE_ERROR_BAD_INPUT
;
4673 else if (procType
== TGSI_PROCESSOR_GEOMETRY
) {
4674 for (i
= 0; i
< numInputs
; i
++) {
4675 t
->inputs
[i
] = ureg_DECL_gs_input(ureg
,
4677 inputSemanticName
[i
],
4678 inputSemanticIndex
[i
]);
4681 for (i
= 0; i
< numOutputs
; i
++) {
4682 t
->outputs
[i
] = ureg_DECL_output(ureg
,
4683 outputSemanticName
[i
],
4684 outputSemanticIndex
[i
]);
4688 assert(procType
== TGSI_PROCESSOR_VERTEX
);
4690 for (i
= 0; i
< numInputs
; i
++) {
4691 t
->inputs
[i
] = ureg_DECL_vs_input(ureg
, i
);
4694 for (i
= 0; i
< numOutputs
; i
++) {
4695 t
->outputs
[i
] = ureg_DECL_output(ureg
,
4696 outputSemanticName
[i
],
4697 outputSemanticIndex
[i
]);
4698 if ((outputSemanticName
[i
] == TGSI_SEMANTIC_PSIZE
) && proginfo
->Id
) {
4699 /* Writing to the point size result register requires special
4700 * handling to implement clamping.
4702 static const gl_state_index pointSizeClampState
[STATE_LENGTH
]
4703 = { STATE_INTERNAL
, STATE_POINT_SIZE_IMPL_CLAMP
, (gl_state_index
)0, (gl_state_index
)0, (gl_state_index
)0 };
4704 /* XXX: note we are modifying the incoming shader here! Need to
4705 * do this before emitting the constant decls below, or this
4708 unsigned pointSizeClampConst
=
4709 _mesa_add_state_reference(proginfo
->Parameters
,
4710 pointSizeClampState
);
4711 struct ureg_dst psizregtemp
= ureg_DECL_temporary(ureg
);
4712 t
->pointSizeConst
= ureg_DECL_constant(ureg
, pointSizeClampConst
);
4713 t
->pointSizeResult
= t
->outputs
[i
];
4714 t
->pointSizeOutIndex
= i
;
4715 t
->outputs
[i
] = psizregtemp
;
4718 if (passthrough_edgeflags
)
4722 /* Declare address register.
4724 if (program
->num_address_regs
> 0) {
4725 assert(program
->num_address_regs
== 1);
4726 t
->address
[0] = ureg_DECL_address(ureg
);
4729 /* Declare misc input registers
4732 GLbitfield sysInputs
= proginfo
->SystemValuesRead
;
4733 unsigned numSys
= 0;
4734 for (i
= 0; sysInputs
; i
++) {
4735 if (sysInputs
& (1 << i
)) {
4736 unsigned semName
= mesa_sysval_to_semantic
[i
];
4737 t
->systemValues
[i
] = ureg_DECL_system_value(ureg
, numSys
, semName
, 0);
4739 sysInputs
&= ~(1 << i
);
4744 if (program
->indirect_addr_temps
) {
4745 /* If temps are accessed with indirect addressing, declare temporaries
4746 * in sequential order. Else, we declare them on demand elsewhere.
4747 * (Note: the number of temporaries is equal to program->next_temp)
4749 for (i
= 0; i
< (unsigned)program
->next_temp
; i
++) {
4750 /* XXX use TGSI_FILE_TEMPORARY_ARRAY when it's supported by ureg */
4751 t
->temps
[i
] = ureg_DECL_temporary(t
->ureg
);
4755 /* Emit constants and uniforms. TGSI uses a single index space for these,
4756 * so we put all the translated regs in t->constants.
4758 if (proginfo
->Parameters
) {
4759 t
->constants
= (struct ureg_src
*)CALLOC(proginfo
->Parameters
->NumParameters
* sizeof(t
->constants
[0]));
4760 if (t
->constants
== NULL
) {
4761 ret
= PIPE_ERROR_OUT_OF_MEMORY
;
4765 for (i
= 0; i
< proginfo
->Parameters
->NumParameters
; i
++) {
4766 switch (proginfo
->Parameters
->Parameters
[i
].Type
) {
4767 case PROGRAM_ENV_PARAM
:
4768 case PROGRAM_LOCAL_PARAM
:
4769 case PROGRAM_STATE_VAR
:
4770 case PROGRAM_NAMED_PARAM
:
4771 case PROGRAM_UNIFORM
:
4772 t
->constants
[i
] = ureg_DECL_constant(ureg
, i
);
4775 /* Emit immediates for PROGRAM_CONSTANT only when there's no indirect
4776 * addressing of the const buffer.
4777 * FIXME: Be smarter and recognize param arrays:
4778 * indirect addressing is only valid within the referenced
4781 case PROGRAM_CONSTANT
:
4782 if (program
->indirect_addr_consts
)
4783 t
->constants
[i
] = ureg_DECL_constant(ureg
, i
);
4785 t
->constants
[i
] = emit_immediate(t
,
4786 proginfo
->Parameters
->ParameterValues
[i
],
4787 proginfo
->Parameters
->Parameters
[i
].DataType
,
4796 /* Emit immediate values.
4798 t
->immediates
= (struct ureg_src
*)CALLOC(program
->num_immediates
* sizeof(struct ureg_src
));
4799 if (t
->immediates
== NULL
) {
4800 ret
= PIPE_ERROR_OUT_OF_MEMORY
;
4804 foreach_iter(exec_list_iterator
, iter
, program
->immediates
) {
4805 immediate_storage
*imm
= (immediate_storage
*)iter
.get();
4806 t
->immediates
[i
++] = emit_immediate(t
, imm
->values
, imm
->type
, imm
->size
);
4809 /* texture samplers */
4810 for (i
= 0; i
< ctx
->Const
.MaxTextureImageUnits
; i
++) {
4811 if (program
->samplers_used
& (1 << i
)) {
4812 t
->samplers
[i
] = ureg_DECL_sampler(ureg
, i
);
4816 /* Emit each instruction in turn:
4818 foreach_iter(exec_list_iterator
, iter
, program
->instructions
) {
4819 set_insn_start(t
, ureg_get_instruction_number(ureg
));
4820 compile_tgsi_instruction(t
, (glsl_to_tgsi_instruction
*)iter
.get());
4822 if (t
->prevInstWrotePointSize
&& proginfo
->Id
) {
4823 /* The previous instruction wrote to the (fake) vertex point size
4824 * result register. Now we need to clamp that value to the min/max
4825 * point size range, putting the result into the real point size
4827 * Note that we can't do this easily at the end of program due to
4828 * possible early return.
4830 set_insn_start(t
, ureg_get_instruction_number(ureg
));
4832 ureg_writemask(t
->outputs
[t
->pointSizeOutIndex
], WRITEMASK_X
),
4833 ureg_src(t
->outputs
[t
->pointSizeOutIndex
]),
4834 ureg_swizzle(t
->pointSizeConst
, 1,1,1,1));
4835 ureg_MIN(t
->ureg
, ureg_writemask(t
->pointSizeResult
, WRITEMASK_X
),
4836 ureg_src(t
->outputs
[t
->pointSizeOutIndex
]),
4837 ureg_swizzle(t
->pointSizeConst
, 2,2,2,2));
4839 t
->prevInstWrotePointSize
= GL_FALSE
;
4842 /* Fix up all emitted labels:
4844 for (i
= 0; i
< t
->labels_count
; i
++) {
4845 ureg_fixup_label(ureg
, t
->labels
[i
].token
,
4846 t
->insn
[t
->labels
[i
].branch_target
]);
4854 FREE(t
->immediates
);
4857 debug_printf("%s: translate error flag set\n", __FUNCTION__
);
4865 /* ----------------------------- End TGSI code ------------------------------ */
4868 * Convert a shader's GLSL IR into a Mesa gl_program, although without
4869 * generating Mesa IR.
4871 static struct gl_program
*
4872 get_mesa_program(struct gl_context
*ctx
,
4873 struct gl_shader_program
*shader_program
,
4874 struct gl_shader
*shader
)
4876 glsl_to_tgsi_visitor
* v
= new glsl_to_tgsi_visitor();
4877 struct gl_program
*prog
;
4879 const char *target_string
;
4881 struct gl_shader_compiler_options
*options
=
4882 &ctx
->ShaderCompilerOptions
[_mesa_shader_type_to_index(shader
->Type
)];
4884 switch (shader
->Type
) {
4885 case GL_VERTEX_SHADER
:
4886 target
= GL_VERTEX_PROGRAM_ARB
;
4887 target_string
= "vertex";
4889 case GL_FRAGMENT_SHADER
:
4890 target
= GL_FRAGMENT_PROGRAM_ARB
;
4891 target_string
= "fragment";
4893 case GL_GEOMETRY_SHADER
:
4894 target
= GL_GEOMETRY_PROGRAM_NV
;
4895 target_string
= "geometry";
4898 assert(!"should not be reached");
4902 validate_ir_tree(shader
->ir
);
4904 prog
= ctx
->Driver
.NewProgram(ctx
, target
, shader_program
->Name
);
4907 prog
->Parameters
= _mesa_new_parameter_list();
4910 v
->shader_program
= shader_program
;
4911 v
->options
= options
;
4912 v
->glsl_version
= ctx
->Const
.GLSLVersion
;
4913 v
->native_integers
= ctx
->Const
.NativeIntegers
;
4915 _mesa_generate_parameters_list_for_uniforms(shader_program
, shader
,
4918 /* Emit intermediate IR for main(). */
4919 visit_exec_list(shader
->ir
, v
);
4921 /* Now emit bodies for any functions that were used. */
4923 progress
= GL_FALSE
;
4925 foreach_iter(exec_list_iterator
, iter
, v
->function_signatures
) {
4926 function_entry
*entry
= (function_entry
*)iter
.get();
4928 if (!entry
->bgn_inst
) {
4929 v
->current_function
= entry
;
4931 entry
->bgn_inst
= v
->emit(NULL
, TGSI_OPCODE_BGNSUB
);
4932 entry
->bgn_inst
->function
= entry
;
4934 visit_exec_list(&entry
->sig
->body
, v
);
4936 glsl_to_tgsi_instruction
*last
;
4937 last
= (glsl_to_tgsi_instruction
*)v
->instructions
.get_tail();
4938 if (last
->op
!= TGSI_OPCODE_RET
)
4939 v
->emit(NULL
, TGSI_OPCODE_RET
);
4941 glsl_to_tgsi_instruction
*end
;
4942 end
= v
->emit(NULL
, TGSI_OPCODE_ENDSUB
);
4943 end
->function
= entry
;
4951 /* Print out some information (for debugging purposes) used by the
4952 * optimization passes. */
4953 for (i
=0; i
< v
->next_temp
; i
++) {
4954 int fr
= v
->get_first_temp_read(i
);
4955 int fw
= v
->get_first_temp_write(i
);
4956 int lr
= v
->get_last_temp_read(i
);
4957 int lw
= v
->get_last_temp_write(i
);
4959 printf("Temp %d: FR=%3d FW=%3d LR=%3d LW=%3d\n", i
, fr
, fw
, lr
, lw
);
4964 /* Remove reads to output registers, and to varyings in vertex shaders. */
4965 v
->remove_output_reads(PROGRAM_OUTPUT
);
4966 if (target
== GL_VERTEX_PROGRAM_ARB
)
4967 v
->remove_output_reads(PROGRAM_VARYING
);
4969 /* Perform optimizations on the instructions in the glsl_to_tgsi_visitor. */
4971 v
->copy_propagate();
4972 while (v
->eliminate_dead_code_advanced());
4974 /* FIXME: These passes to optimize temporary registers don't work when there
4975 * is indirect addressing of the temporary register space. We need proper
4976 * array support so that we don't have to give up these passes in every
4977 * shader that uses arrays.
4979 if (!v
->indirect_addr_temps
) {
4980 v
->eliminate_dead_code();
4981 v
->merge_registers();
4982 v
->renumber_registers();
4985 /* Write the END instruction. */
4986 v
->emit(NULL
, TGSI_OPCODE_END
);
4988 if (ctx
->Shader
.Flags
& GLSL_DUMP
) {
4990 printf("GLSL IR for linked %s program %d:\n", target_string
,
4991 shader_program
->Name
);
4992 _mesa_print_ir(shader
->ir
, NULL
);
4997 prog
->Instructions
= NULL
;
4998 prog
->NumInstructions
= 0;
5000 do_set_program_inouts(shader
->ir
, prog
, shader
->Type
== GL_FRAGMENT_SHADER
);
5001 count_resources(v
, prog
);
5003 check_resources(ctx
, shader_program
, v
, prog
);
5005 _mesa_reference_program(ctx
, &shader
->Program
, prog
);
5007 /* This has to be done last. Any operation the can cause
5008 * prog->ParameterValues to get reallocated (e.g., anything that adds a
5009 * program constant) has to happen before creating this linkage.
5011 _mesa_associate_uniform_storage(ctx
, shader_program
, prog
->Parameters
);
5012 if (!shader_program
->LinkStatus
) {
5016 struct st_vertex_program
*stvp
;
5017 struct st_fragment_program
*stfp
;
5018 struct st_geometry_program
*stgp
;
5020 switch (shader
->Type
) {
5021 case GL_VERTEX_SHADER
:
5022 stvp
= (struct st_vertex_program
*)prog
;
5023 stvp
->glsl_to_tgsi
= v
;
5025 case GL_FRAGMENT_SHADER
:
5026 stfp
= (struct st_fragment_program
*)prog
;
5027 stfp
->glsl_to_tgsi
= v
;
5029 case GL_GEOMETRY_SHADER
:
5030 stgp
= (struct st_geometry_program
*)prog
;
5031 stgp
->glsl_to_tgsi
= v
;
5034 assert(!"should not be reached");
5044 st_new_shader(struct gl_context
*ctx
, GLuint name
, GLuint type
)
5046 struct gl_shader
*shader
;
5047 assert(type
== GL_FRAGMENT_SHADER
|| type
== GL_VERTEX_SHADER
||
5048 type
== GL_GEOMETRY_SHADER_ARB
);
5049 shader
= rzalloc(NULL
, struct gl_shader
);
5051 shader
->Type
= type
;
5052 shader
->Name
= name
;
5053 _mesa_init_shader(ctx
, shader
);
5058 struct gl_shader_program
*
5059 st_new_shader_program(struct gl_context
*ctx
, GLuint name
)
5061 struct gl_shader_program
*shProg
;
5062 shProg
= rzalloc(NULL
, struct gl_shader_program
);
5064 shProg
->Name
= name
;
5065 _mesa_init_shader_program(ctx
, shProg
);
5072 * Called via ctx->Driver.LinkShader()
5073 * This actually involves converting GLSL IR into an intermediate TGSI-like IR
5074 * with code lowering and other optimizations.
5077 st_link_shader(struct gl_context
*ctx
, struct gl_shader_program
*prog
)
5079 assert(prog
->LinkStatus
);
5081 for (unsigned i
= 0; i
< MESA_SHADER_TYPES
; i
++) {
5082 if (prog
->_LinkedShaders
[i
] == NULL
)
5086 exec_list
*ir
= prog
->_LinkedShaders
[i
]->ir
;
5087 const struct gl_shader_compiler_options
*options
=
5088 &ctx
->ShaderCompilerOptions
[_mesa_shader_type_to_index(prog
->_LinkedShaders
[i
]->Type
)];
5094 do_mat_op_to_vec(ir
);
5095 lower_instructions(ir
, (MOD_TO_FRACT
| DIV_TO_MUL_RCP
| EXP_TO_EXP2
5096 | LOG_TO_LOG2
| INT_DIV_TO_MUL_RCP
5097 | ((options
->EmitNoPow
) ? POW_TO_EXP2
: 0)));
5099 progress
= do_lower_jumps(ir
, true, true, options
->EmitNoMainReturn
, options
->EmitNoCont
, options
->EmitNoLoops
) || progress
;
5101 progress
= do_common_optimization(ir
, true, true,
5102 options
->MaxUnrollIterations
)
5105 progress
= lower_quadop_vector(ir
, false) || progress
;
5107 if (options
->MaxIfDepth
== 0)
5108 progress
= lower_discard(ir
) || progress
;
5110 progress
= lower_if_to_cond_assign(ir
, options
->MaxIfDepth
) || progress
;
5112 if (options
->EmitNoNoise
)
5113 progress
= lower_noise(ir
) || progress
;
5115 /* If there are forms of indirect addressing that the driver
5116 * cannot handle, perform the lowering pass.
5118 if (options
->EmitNoIndirectInput
|| options
->EmitNoIndirectOutput
5119 || options
->EmitNoIndirectTemp
|| options
->EmitNoIndirectUniform
)
5121 lower_variable_index_to_cond_assign(ir
,
5122 options
->EmitNoIndirectInput
,
5123 options
->EmitNoIndirectOutput
,
5124 options
->EmitNoIndirectTemp
,
5125 options
->EmitNoIndirectUniform
)
5128 progress
= do_vec_index_to_cond_assign(ir
) || progress
;
5131 validate_ir_tree(ir
);
5134 for (unsigned i
= 0; i
< MESA_SHADER_TYPES
; i
++) {
5135 struct gl_program
*linked_prog
;
5137 if (prog
->_LinkedShaders
[i
] == NULL
)
5140 linked_prog
= get_mesa_program(ctx
, prog
, prog
->_LinkedShaders
[i
]);
5143 static const GLenum targets
[] = {
5144 GL_VERTEX_PROGRAM_ARB
,
5145 GL_FRAGMENT_PROGRAM_ARB
,
5146 GL_GEOMETRY_PROGRAM_NV
5149 _mesa_reference_program(ctx
, &prog
->_LinkedShaders
[i
]->Program
,
5151 if (!ctx
->Driver
.ProgramStringNotify(ctx
, targets
[i
], linked_prog
)) {
5152 _mesa_reference_program(ctx
, &prog
->_LinkedShaders
[i
]->Program
,
5154 _mesa_reference_program(ctx
, &linked_prog
, NULL
);
5159 _mesa_reference_program(ctx
, &linked_prog
, NULL
);