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"
46 #include "main/mtypes.h"
47 #include "main/shaderapi.h"
48 #include "main/shaderobj.h"
49 #include "main/uniforms.h"
50 #include "program/hash_table.h"
51 #include "program/prog_instruction.h"
52 #include "program/prog_optimize.h"
53 #include "program/prog_print.h"
54 #include "program/program.h"
55 #include "program/prog_uniform.h"
56 #include "program/prog_parameter.h"
57 #include "program/sampler.h"
59 #include "pipe/p_compiler.h"
60 #include "pipe/p_context.h"
61 #include "pipe/p_screen.h"
62 #include "pipe/p_shader_tokens.h"
63 #include "pipe/p_state.h"
64 #include "util/u_math.h"
65 #include "tgsi/tgsi_ureg.h"
66 #include "tgsi/tgsi_info.h"
67 #include "st_context.h"
68 #include "st_program.h"
69 #include "st_glsl_to_tgsi.h"
70 #include "st_mesa_to_tgsi.h"
73 #define PROGRAM_IMMEDIATE PROGRAM_FILE_MAX
74 #define PROGRAM_ANY_CONST ((1 << PROGRAM_LOCAL_PARAM) | \
75 (1 << PROGRAM_ENV_PARAM) | \
76 (1 << PROGRAM_STATE_VAR) | \
77 (1 << PROGRAM_NAMED_PARAM) | \
78 (1 << PROGRAM_CONSTANT) | \
79 (1 << PROGRAM_UNIFORM))
81 #define MAX_TEMPS 4096
83 /* will be 4 for GLSL 4.00 */
84 #define MAX_GLSL_TEXTURE_OFFSET 1
89 static int swizzle_for_size(int size
);
92 * This struct is a corresponding struct to TGSI ureg_src.
96 st_src_reg(gl_register_file file
, int index
, const glsl_type
*type
)
100 if (type
&& (type
->is_scalar() || type
->is_vector() || type
->is_matrix()))
101 this->swizzle
= swizzle_for_size(type
->vector_elements
);
103 this->swizzle
= SWIZZLE_XYZW
;
105 this->type
= type
? type
->base_type
: GLSL_TYPE_ERROR
;
106 this->reladdr
= NULL
;
109 st_src_reg(gl_register_file file
, int index
, int type
)
114 this->swizzle
= SWIZZLE_XYZW
;
116 this->reladdr
= NULL
;
121 this->type
= GLSL_TYPE_ERROR
;
122 this->file
= PROGRAM_UNDEFINED
;
126 this->reladdr
= NULL
;
129 explicit st_src_reg(st_dst_reg reg
);
131 gl_register_file file
; /**< PROGRAM_* from Mesa */
132 int index
; /**< temporary index, VERT_ATTRIB_*, FRAG_ATTRIB_*, etc. */
133 GLuint swizzle
; /**< SWIZZLE_XYZWONEZERO swizzles from Mesa. */
134 int negate
; /**< NEGATE_XYZW mask from mesa */
135 int type
; /** GLSL_TYPE_* from GLSL IR (enum glsl_base_type) */
136 /** Register index should be offset by the integer in this reg. */
142 st_dst_reg(gl_register_file file
, int writemask
, int type
)
146 this->writemask
= writemask
;
147 this->cond_mask
= COND_TR
;
148 this->reladdr
= NULL
;
154 this->type
= GLSL_TYPE_ERROR
;
155 this->file
= PROGRAM_UNDEFINED
;
158 this->cond_mask
= COND_TR
;
159 this->reladdr
= NULL
;
162 explicit st_dst_reg(st_src_reg reg
);
164 gl_register_file file
; /**< PROGRAM_* from Mesa */
165 int index
; /**< temporary index, VERT_ATTRIB_*, FRAG_ATTRIB_*, etc. */
166 int writemask
; /**< Bitfield of WRITEMASK_[XYZW] */
168 int type
; /** GLSL_TYPE_* from GLSL IR (enum glsl_base_type) */
169 /** Register index should be offset by the integer in this reg. */
173 st_src_reg::st_src_reg(st_dst_reg reg
)
175 this->type
= reg
.type
;
176 this->file
= reg
.file
;
177 this->index
= reg
.index
;
178 this->swizzle
= SWIZZLE_XYZW
;
180 this->reladdr
= reg
.reladdr
;
183 st_dst_reg::st_dst_reg(st_src_reg reg
)
185 this->type
= reg
.type
;
186 this->file
= reg
.file
;
187 this->index
= reg
.index
;
188 this->writemask
= WRITEMASK_XYZW
;
189 this->cond_mask
= COND_TR
;
190 this->reladdr
= reg
.reladdr
;
193 class glsl_to_tgsi_instruction
: public exec_node
{
195 /* Callers of this ralloc-based new need not call delete. It's
196 * easier to just ralloc_free 'ctx' (or any of its ancestors). */
197 static void* operator new(size_t size
, void *ctx
)
201 node
= rzalloc_size(ctx
, size
);
202 assert(node
!= NULL
);
210 /** Pointer to the ir source this tree came from for debugging */
212 GLboolean cond_update
;
214 int sampler
; /**< sampler index */
215 int tex_target
; /**< One of TEXTURE_*_INDEX */
216 GLboolean tex_shadow
;
217 struct tgsi_texture_offset tex_offsets
[MAX_GLSL_TEXTURE_OFFSET
];
218 unsigned tex_offset_num_offset
;
219 int dead_mask
; /**< Used in dead code elimination */
221 class function_entry
*function
; /* Set on TGSI_OPCODE_CAL or TGSI_OPCODE_BGNSUB */
224 class variable_storage
: public exec_node
{
226 variable_storage(ir_variable
*var
, gl_register_file file
, int index
)
227 : file(file
), index(index
), var(var
)
232 gl_register_file file
;
234 ir_variable
*var
; /* variable that maps to this, if any */
237 class immediate_storage
: public exec_node
{
239 immediate_storage(gl_constant_value
*values
, int size
, int type
)
241 memcpy(this->values
, values
, size
* sizeof(gl_constant_value
));
246 gl_constant_value values
[4];
247 int size
; /**< Number of components (1-4) */
248 int type
; /**< GL_FLOAT, GL_INT, GL_BOOL, or GL_UNSIGNED_INT */
251 class function_entry
: public exec_node
{
253 ir_function_signature
*sig
;
256 * identifier of this function signature used by the program.
258 * At the point that TGSI instructions for function calls are
259 * generated, we don't know the address of the first instruction of
260 * the function body. So we make the BranchTarget that is called a
261 * small integer and rewrite them during set_branchtargets().
266 * Pointer to first instruction of the function body.
268 * Set during function body emits after main() is processed.
270 glsl_to_tgsi_instruction
*bgn_inst
;
273 * Index of the first instruction of the function body in actual TGSI.
275 * Set after conversion from glsl_to_tgsi_instruction to TGSI.
279 /** Storage for the return value. */
280 st_src_reg return_reg
;
283 class glsl_to_tgsi_visitor
: public ir_visitor
{
285 glsl_to_tgsi_visitor();
286 ~glsl_to_tgsi_visitor();
288 function_entry
*current_function
;
290 struct gl_context
*ctx
;
291 struct gl_program
*prog
;
292 struct gl_shader_program
*shader_program
;
293 struct gl_shader_compiler_options
*options
;
297 int num_address_regs
;
299 bool indirect_addr_temps
;
300 bool indirect_addr_consts
;
303 bool native_integers
;
305 variable_storage
*find_variable_storage(ir_variable
*var
);
307 int add_constant(gl_register_file file
, gl_constant_value values
[4],
308 int size
, int datatype
, GLuint
*swizzle_out
);
310 function_entry
*get_function_signature(ir_function_signature
*sig
);
312 st_src_reg
get_temp(const glsl_type
*type
);
313 void reladdr_to_temp(ir_instruction
*ir
, st_src_reg
*reg
, int *num_reladdr
);
315 st_src_reg
st_src_reg_for_float(float val
);
316 st_src_reg
st_src_reg_for_int(int val
);
317 st_src_reg
st_src_reg_for_type(int type
, int val
);
320 * \name Visit methods
322 * As typical for the visitor pattern, there must be one \c visit method for
323 * each concrete subclass of \c ir_instruction. Virtual base classes within
324 * the hierarchy should not have \c visit methods.
327 virtual void visit(ir_variable
*);
328 virtual void visit(ir_loop
*);
329 virtual void visit(ir_loop_jump
*);
330 virtual void visit(ir_function_signature
*);
331 virtual void visit(ir_function
*);
332 virtual void visit(ir_expression
*);
333 virtual void visit(ir_swizzle
*);
334 virtual void visit(ir_dereference_variable
*);
335 virtual void visit(ir_dereference_array
*);
336 virtual void visit(ir_dereference_record
*);
337 virtual void visit(ir_assignment
*);
338 virtual void visit(ir_constant
*);
339 virtual void visit(ir_call
*);
340 virtual void visit(ir_return
*);
341 virtual void visit(ir_discard
*);
342 virtual void visit(ir_texture
*);
343 virtual void visit(ir_if
*);
348 /** List of variable_storage */
351 /** List of immediate_storage */
352 exec_list immediates
;
355 /** List of function_entry */
356 exec_list function_signatures
;
357 int next_signature_id
;
359 /** List of glsl_to_tgsi_instruction */
360 exec_list instructions
;
362 glsl_to_tgsi_instruction
*emit(ir_instruction
*ir
, unsigned op
);
364 glsl_to_tgsi_instruction
*emit(ir_instruction
*ir
, unsigned op
,
365 st_dst_reg dst
, st_src_reg src0
);
367 glsl_to_tgsi_instruction
*emit(ir_instruction
*ir
, unsigned op
,
368 st_dst_reg dst
, st_src_reg src0
, st_src_reg src1
);
370 glsl_to_tgsi_instruction
*emit(ir_instruction
*ir
, unsigned op
,
372 st_src_reg src0
, st_src_reg src1
, st_src_reg src2
);
374 unsigned get_opcode(ir_instruction
*ir
, unsigned op
,
376 st_src_reg src0
, st_src_reg src1
);
379 * Emit the correct dot-product instruction for the type of arguments
381 glsl_to_tgsi_instruction
*emit_dp(ir_instruction
*ir
,
387 void emit_scalar(ir_instruction
*ir
, unsigned op
,
388 st_dst_reg dst
, st_src_reg src0
);
390 void emit_scalar(ir_instruction
*ir
, unsigned op
,
391 st_dst_reg dst
, st_src_reg src0
, st_src_reg src1
);
393 void try_emit_float_set(ir_instruction
*ir
, unsigned op
, st_dst_reg dst
);
395 void emit_arl(ir_instruction
*ir
, st_dst_reg dst
, st_src_reg src0
);
397 void emit_scs(ir_instruction
*ir
, unsigned op
,
398 st_dst_reg dst
, const st_src_reg
&src
);
400 bool try_emit_mad(ir_expression
*ir
,
402 bool try_emit_mad_for_and_not(ir_expression
*ir
,
404 bool try_emit_sat(ir_expression
*ir
);
406 void emit_swz(ir_expression
*ir
);
408 bool process_move_condition(ir_rvalue
*ir
);
410 void remove_output_reads(gl_register_file type
);
411 void simplify_cmp(void);
413 void rename_temp_register(int index
, int new_index
);
414 int get_first_temp_read(int index
);
415 int get_first_temp_write(int index
);
416 int get_last_temp_read(int index
);
417 int get_last_temp_write(int index
);
419 void copy_propagate(void);
420 void eliminate_dead_code(void);
421 int eliminate_dead_code_advanced(void);
422 void merge_registers(void);
423 void renumber_registers(void);
428 static st_src_reg undef_src
= st_src_reg(PROGRAM_UNDEFINED
, 0, GLSL_TYPE_ERROR
);
430 static st_dst_reg undef_dst
= st_dst_reg(PROGRAM_UNDEFINED
, SWIZZLE_NOOP
, GLSL_TYPE_ERROR
);
432 static st_dst_reg address_reg
= st_dst_reg(PROGRAM_ADDRESS
, WRITEMASK_X
, GLSL_TYPE_FLOAT
);
435 fail_link(struct gl_shader_program
*prog
, const char *fmt
, ...) PRINTFLIKE(2, 3);
438 fail_link(struct gl_shader_program
*prog
, const char *fmt
, ...)
442 ralloc_vasprintf_append(&prog
->InfoLog
, fmt
, args
);
445 prog
->LinkStatus
= GL_FALSE
;
449 swizzle_for_size(int size
)
451 int size_swizzles
[4] = {
452 MAKE_SWIZZLE4(SWIZZLE_X
, SWIZZLE_X
, SWIZZLE_X
, SWIZZLE_X
),
453 MAKE_SWIZZLE4(SWIZZLE_X
, SWIZZLE_Y
, SWIZZLE_Y
, SWIZZLE_Y
),
454 MAKE_SWIZZLE4(SWIZZLE_X
, SWIZZLE_Y
, SWIZZLE_Z
, SWIZZLE_Z
),
455 MAKE_SWIZZLE4(SWIZZLE_X
, SWIZZLE_Y
, SWIZZLE_Z
, SWIZZLE_W
),
458 assert((size
>= 1) && (size
<= 4));
459 return size_swizzles
[size
- 1];
463 is_tex_instruction(unsigned opcode
)
465 const tgsi_opcode_info
* info
= tgsi_get_opcode_info(opcode
);
470 num_inst_dst_regs(unsigned opcode
)
472 const tgsi_opcode_info
* info
= tgsi_get_opcode_info(opcode
);
473 return info
->num_dst
;
477 num_inst_src_regs(unsigned opcode
)
479 const tgsi_opcode_info
* info
= tgsi_get_opcode_info(opcode
);
480 return info
->is_tex
? info
->num_src
- 1 : info
->num_src
;
483 glsl_to_tgsi_instruction
*
484 glsl_to_tgsi_visitor::emit(ir_instruction
*ir
, unsigned op
,
486 st_src_reg src0
, st_src_reg src1
, st_src_reg src2
)
488 glsl_to_tgsi_instruction
*inst
= new(mem_ctx
) glsl_to_tgsi_instruction();
489 int num_reladdr
= 0, i
;
491 op
= get_opcode(ir
, op
, dst
, src0
, src1
);
493 /* If we have to do relative addressing, we want to load the ARL
494 * reg directly for one of the regs, and preload the other reladdr
495 * sources into temps.
497 num_reladdr
+= dst
.reladdr
!= NULL
;
498 num_reladdr
+= src0
.reladdr
!= NULL
;
499 num_reladdr
+= src1
.reladdr
!= NULL
;
500 num_reladdr
+= src2
.reladdr
!= NULL
;
502 reladdr_to_temp(ir
, &src2
, &num_reladdr
);
503 reladdr_to_temp(ir
, &src1
, &num_reladdr
);
504 reladdr_to_temp(ir
, &src0
, &num_reladdr
);
507 emit_arl(ir
, address_reg
, *dst
.reladdr
);
510 assert(num_reladdr
== 0);
520 inst
->function
= NULL
;
522 if (op
== TGSI_OPCODE_ARL
|| op
== TGSI_OPCODE_UARL
)
523 this->num_address_regs
= 1;
525 /* Update indirect addressing status used by TGSI */
528 case PROGRAM_TEMPORARY
:
529 this->indirect_addr_temps
= true;
531 case PROGRAM_LOCAL_PARAM
:
532 case PROGRAM_ENV_PARAM
:
533 case PROGRAM_STATE_VAR
:
534 case PROGRAM_NAMED_PARAM
:
535 case PROGRAM_CONSTANT
:
536 case PROGRAM_UNIFORM
:
537 this->indirect_addr_consts
= true;
539 case PROGRAM_IMMEDIATE
:
540 assert(!"immediates should not have indirect addressing");
547 for (i
=0; i
<3; i
++) {
548 if(inst
->src
[i
].reladdr
) {
549 switch(inst
->src
[i
].file
) {
550 case PROGRAM_TEMPORARY
:
551 this->indirect_addr_temps
= true;
553 case PROGRAM_LOCAL_PARAM
:
554 case PROGRAM_ENV_PARAM
:
555 case PROGRAM_STATE_VAR
:
556 case PROGRAM_NAMED_PARAM
:
557 case PROGRAM_CONSTANT
:
558 case PROGRAM_UNIFORM
:
559 this->indirect_addr_consts
= true;
561 case PROGRAM_IMMEDIATE
:
562 assert(!"immediates should not have indirect addressing");
571 this->instructions
.push_tail(inst
);
574 try_emit_float_set(ir
, op
, dst
);
580 glsl_to_tgsi_instruction
*
581 glsl_to_tgsi_visitor::emit(ir_instruction
*ir
, unsigned op
,
582 st_dst_reg dst
, st_src_reg src0
, st_src_reg src1
)
584 return emit(ir
, op
, dst
, src0
, src1
, undef_src
);
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
)
591 assert(dst
.writemask
!= 0);
592 return emit(ir
, op
, dst
, src0
, undef_src
, undef_src
);
595 glsl_to_tgsi_instruction
*
596 glsl_to_tgsi_visitor::emit(ir_instruction
*ir
, unsigned op
)
598 return emit(ir
, op
, undef_dst
, undef_src
, undef_src
, undef_src
);
602 * Emits the code to convert the result of float SET instructions to integers.
605 glsl_to_tgsi_visitor::try_emit_float_set(ir_instruction
*ir
, unsigned op
,
608 if ((op
== TGSI_OPCODE_SEQ
||
609 op
== TGSI_OPCODE_SNE
||
610 op
== TGSI_OPCODE_SGE
||
611 op
== TGSI_OPCODE_SLT
))
613 st_src_reg src
= st_src_reg(dst
);
614 src
.negate
= ~src
.negate
;
615 dst
.type
= GLSL_TYPE_FLOAT
;
616 emit(ir
, TGSI_OPCODE_F2I
, dst
, src
);
621 * Determines whether to use an integer, unsigned integer, or float opcode
622 * based on the operands and input opcode, then emits the result.
625 glsl_to_tgsi_visitor::get_opcode(ir_instruction
*ir
, unsigned op
,
627 st_src_reg src0
, st_src_reg src1
)
629 int type
= GLSL_TYPE_FLOAT
;
631 if (src0
.type
== GLSL_TYPE_FLOAT
|| src1
.type
== GLSL_TYPE_FLOAT
)
632 type
= GLSL_TYPE_FLOAT
;
633 else if (native_integers
)
634 type
= src0
.type
== GLSL_TYPE_BOOL
? GLSL_TYPE_INT
: src0
.type
;
636 #define case4(c, f, i, u) \
637 case TGSI_OPCODE_##c: \
638 if (type == GLSL_TYPE_INT) op = TGSI_OPCODE_##i; \
639 else if (type == GLSL_TYPE_UINT) op = TGSI_OPCODE_##u; \
640 else op = TGSI_OPCODE_##f; \
642 #define case3(f, i, u) case4(f, f, i, u)
643 #define case2fi(f, i) case4(f, f, i, i)
644 #define case2iu(i, u) case4(i, LAST, i, u)
650 case3(DIV
, IDIV
, UDIV
);
651 case3(MAX
, IMAX
, UMAX
);
652 case3(MIN
, IMIN
, UMIN
);
657 case3(SGE
, ISGE
, USGE
);
658 case3(SLT
, ISLT
, USLT
);
665 assert(op
!= TGSI_OPCODE_LAST
);
669 glsl_to_tgsi_instruction
*
670 glsl_to_tgsi_visitor::emit_dp(ir_instruction
*ir
,
671 st_dst_reg dst
, st_src_reg src0
, st_src_reg src1
,
674 static const unsigned dot_opcodes
[] = {
675 TGSI_OPCODE_DP2
, TGSI_OPCODE_DP3
, TGSI_OPCODE_DP4
678 return emit(ir
, dot_opcodes
[elements
- 2], dst
, src0
, src1
);
682 * Emits TGSI scalar opcodes to produce unique answers across channels.
684 * Some TGSI opcodes are scalar-only, like ARB_fp/vp. The src X
685 * channel determines the result across all channels. So to do a vec4
686 * of this operation, we want to emit a scalar per source channel used
687 * to produce dest channels.
690 glsl_to_tgsi_visitor::emit_scalar(ir_instruction
*ir
, unsigned op
,
692 st_src_reg orig_src0
, st_src_reg orig_src1
)
695 int done_mask
= ~dst
.writemask
;
697 /* TGSI RCP is a scalar operation splatting results to all channels,
698 * like ARB_fp/vp. So emit as many RCPs as necessary to cover our
701 for (i
= 0; i
< 4; i
++) {
702 GLuint this_mask
= (1 << i
);
703 glsl_to_tgsi_instruction
*inst
;
704 st_src_reg src0
= orig_src0
;
705 st_src_reg src1
= orig_src1
;
707 if (done_mask
& this_mask
)
710 GLuint src0_swiz
= GET_SWZ(src0
.swizzle
, i
);
711 GLuint src1_swiz
= GET_SWZ(src1
.swizzle
, i
);
712 for (j
= i
+ 1; j
< 4; j
++) {
713 /* If there is another enabled component in the destination that is
714 * derived from the same inputs, generate its value on this pass as
717 if (!(done_mask
& (1 << j
)) &&
718 GET_SWZ(src0
.swizzle
, j
) == src0_swiz
&&
719 GET_SWZ(src1
.swizzle
, j
) == src1_swiz
) {
720 this_mask
|= (1 << j
);
723 src0
.swizzle
= MAKE_SWIZZLE4(src0_swiz
, src0_swiz
,
724 src0_swiz
, src0_swiz
);
725 src1
.swizzle
= MAKE_SWIZZLE4(src1_swiz
, src1_swiz
,
726 src1_swiz
, src1_swiz
);
728 inst
= emit(ir
, op
, dst
, src0
, src1
);
729 inst
->dst
.writemask
= this_mask
;
730 done_mask
|= this_mask
;
735 glsl_to_tgsi_visitor::emit_scalar(ir_instruction
*ir
, unsigned op
,
736 st_dst_reg dst
, st_src_reg src0
)
738 st_src_reg undef
= undef_src
;
740 undef
.swizzle
= SWIZZLE_XXXX
;
742 emit_scalar(ir
, op
, dst
, src0
, undef
);
746 glsl_to_tgsi_visitor::emit_arl(ir_instruction
*ir
,
747 st_dst_reg dst
, st_src_reg src0
)
749 int op
= TGSI_OPCODE_ARL
;
751 if (src0
.type
== GLSL_TYPE_INT
|| src0
.type
== GLSL_TYPE_UINT
)
752 op
= TGSI_OPCODE_UARL
;
754 emit(NULL
, op
, dst
, src0
);
758 * Emit an TGSI_OPCODE_SCS instruction
760 * The \c SCS opcode functions a bit differently than the other TGSI opcodes.
761 * Instead of splatting its result across all four components of the
762 * destination, it writes one value to the \c x component and another value to
763 * the \c y component.
765 * \param ir IR instruction being processed
766 * \param op Either \c TGSI_OPCODE_SIN or \c TGSI_OPCODE_COS depending
767 * on which value is desired.
768 * \param dst Destination register
769 * \param src Source register
772 glsl_to_tgsi_visitor::emit_scs(ir_instruction
*ir
, unsigned op
,
774 const st_src_reg
&src
)
776 /* Vertex programs cannot use the SCS opcode.
778 if (this->prog
->Target
== GL_VERTEX_PROGRAM_ARB
) {
779 emit_scalar(ir
, op
, dst
, src
);
783 const unsigned component
= (op
== TGSI_OPCODE_SIN
) ? 0 : 1;
784 const unsigned scs_mask
= (1U << component
);
785 int done_mask
= ~dst
.writemask
;
788 assert(op
== TGSI_OPCODE_SIN
|| op
== TGSI_OPCODE_COS
);
790 /* If there are compnents in the destination that differ from the component
791 * that will be written by the SCS instrution, we'll need a temporary.
793 if (scs_mask
!= unsigned(dst
.writemask
)) {
794 tmp
= get_temp(glsl_type::vec4_type
);
797 for (unsigned i
= 0; i
< 4; i
++) {
798 unsigned this_mask
= (1U << i
);
799 st_src_reg src0
= src
;
801 if ((done_mask
& this_mask
) != 0)
804 /* The source swizzle specified which component of the source generates
805 * sine / cosine for the current component in the destination. The SCS
806 * instruction requires that this value be swizzle to the X component.
807 * Replace the current swizzle with a swizzle that puts the source in
810 unsigned src0_swiz
= GET_SWZ(src
.swizzle
, i
);
812 src0
.swizzle
= MAKE_SWIZZLE4(src0_swiz
, src0_swiz
,
813 src0_swiz
, src0_swiz
);
814 for (unsigned j
= i
+ 1; j
< 4; j
++) {
815 /* If there is another enabled component in the destination that is
816 * derived from the same inputs, generate its value on this pass as
819 if (!(done_mask
& (1 << j
)) &&
820 GET_SWZ(src0
.swizzle
, j
) == src0_swiz
) {
821 this_mask
|= (1 << j
);
825 if (this_mask
!= scs_mask
) {
826 glsl_to_tgsi_instruction
*inst
;
827 st_dst_reg tmp_dst
= st_dst_reg(tmp
);
829 /* Emit the SCS instruction.
831 inst
= emit(ir
, TGSI_OPCODE_SCS
, tmp_dst
, src0
);
832 inst
->dst
.writemask
= scs_mask
;
834 /* Move the result of the SCS instruction to the desired location in
837 tmp
.swizzle
= MAKE_SWIZZLE4(component
, component
,
838 component
, component
);
839 inst
= emit(ir
, TGSI_OPCODE_SCS
, dst
, tmp
);
840 inst
->dst
.writemask
= this_mask
;
842 /* Emit the SCS instruction to write directly to the destination.
844 glsl_to_tgsi_instruction
*inst
= emit(ir
, TGSI_OPCODE_SCS
, dst
, src0
);
845 inst
->dst
.writemask
= scs_mask
;
848 done_mask
|= this_mask
;
853 glsl_to_tgsi_visitor::add_constant(gl_register_file file
,
854 gl_constant_value values
[4], int size
, int datatype
,
857 if (file
== PROGRAM_CONSTANT
) {
858 return _mesa_add_typed_unnamed_constant(this->prog
->Parameters
, values
,
859 size
, datatype
, swizzle_out
);
862 immediate_storage
*entry
;
863 assert(file
== PROGRAM_IMMEDIATE
);
865 /* Search immediate storage to see if we already have an identical
866 * immediate that we can use instead of adding a duplicate entry.
868 foreach_iter(exec_list_iterator
, iter
, this->immediates
) {
869 entry
= (immediate_storage
*)iter
.get();
871 if (entry
->size
== size
&&
872 entry
->type
== datatype
&&
873 !memcmp(entry
->values
, values
, size
* sizeof(gl_constant_value
))) {
879 /* Add this immediate to the list. */
880 entry
= new(mem_ctx
) immediate_storage(values
, size
, datatype
);
881 this->immediates
.push_tail(entry
);
882 this->num_immediates
++;
888 glsl_to_tgsi_visitor::st_src_reg_for_float(float val
)
890 st_src_reg
src(PROGRAM_IMMEDIATE
, -1, GLSL_TYPE_FLOAT
);
891 union gl_constant_value uval
;
894 src
.index
= add_constant(src
.file
, &uval
, 1, GL_FLOAT
, &src
.swizzle
);
900 glsl_to_tgsi_visitor::st_src_reg_for_int(int val
)
902 st_src_reg
src(PROGRAM_IMMEDIATE
, -1, GLSL_TYPE_INT
);
903 union gl_constant_value uval
;
905 assert(native_integers
);
908 src
.index
= add_constant(src
.file
, &uval
, 1, GL_INT
, &src
.swizzle
);
914 glsl_to_tgsi_visitor::st_src_reg_for_type(int type
, int val
)
917 return type
== GLSL_TYPE_FLOAT
? st_src_reg_for_float(val
) :
918 st_src_reg_for_int(val
);
920 return st_src_reg_for_float(val
);
924 type_size(const struct glsl_type
*type
)
929 switch (type
->base_type
) {
932 case GLSL_TYPE_FLOAT
:
934 if (type
->is_matrix()) {
935 return type
->matrix_columns
;
937 /* Regardless of size of vector, it gets a vec4. This is bad
938 * packing for things like floats, but otherwise arrays become a
939 * mess. Hopefully a later pass over the code can pack scalars
940 * down if appropriate.
944 case GLSL_TYPE_ARRAY
:
945 assert(type
->length
> 0);
946 return type_size(type
->fields
.array
) * type
->length
;
947 case GLSL_TYPE_STRUCT
:
949 for (i
= 0; i
< type
->length
; i
++) {
950 size
+= type_size(type
->fields
.structure
[i
].type
);
953 case GLSL_TYPE_SAMPLER
:
954 /* Samplers take up one slot in UNIFORMS[], but they're baked in
965 * In the initial pass of codegen, we assign temporary numbers to
966 * intermediate results. (not SSA -- variable assignments will reuse
970 glsl_to_tgsi_visitor::get_temp(const glsl_type
*type
)
974 src
.type
= native_integers
? type
->base_type
: GLSL_TYPE_FLOAT
;
975 src
.file
= PROGRAM_TEMPORARY
;
976 src
.index
= next_temp
;
978 next_temp
+= type_size(type
);
980 if (type
->is_array() || type
->is_record()) {
981 src
.swizzle
= SWIZZLE_NOOP
;
983 src
.swizzle
= swizzle_for_size(type
->vector_elements
);
991 glsl_to_tgsi_visitor::find_variable_storage(ir_variable
*var
)
994 variable_storage
*entry
;
996 foreach_iter(exec_list_iterator
, iter
, this->variables
) {
997 entry
= (variable_storage
*)iter
.get();
999 if (entry
->var
== var
)
1007 glsl_to_tgsi_visitor::visit(ir_variable
*ir
)
1009 if (strcmp(ir
->name
, "gl_FragCoord") == 0) {
1010 struct gl_fragment_program
*fp
= (struct gl_fragment_program
*)this->prog
;
1012 fp
->OriginUpperLeft
= ir
->origin_upper_left
;
1013 fp
->PixelCenterInteger
= ir
->pixel_center_integer
;
1015 } else if (strcmp(ir
->name
, "gl_FragDepth") == 0) {
1016 struct gl_fragment_program
*fp
= (struct gl_fragment_program
*)this->prog
;
1017 switch (ir
->depth_layout
) {
1018 case ir_depth_layout_none
:
1019 fp
->FragDepthLayout
= FRAG_DEPTH_LAYOUT_NONE
;
1021 case ir_depth_layout_any
:
1022 fp
->FragDepthLayout
= FRAG_DEPTH_LAYOUT_ANY
;
1024 case ir_depth_layout_greater
:
1025 fp
->FragDepthLayout
= FRAG_DEPTH_LAYOUT_GREATER
;
1027 case ir_depth_layout_less
:
1028 fp
->FragDepthLayout
= FRAG_DEPTH_LAYOUT_LESS
;
1030 case ir_depth_layout_unchanged
:
1031 fp
->FragDepthLayout
= FRAG_DEPTH_LAYOUT_UNCHANGED
;
1039 if (ir
->mode
== ir_var_uniform
&& strncmp(ir
->name
, "gl_", 3) == 0) {
1041 const ir_state_slot
*const slots
= ir
->state_slots
;
1042 assert(ir
->state_slots
!= NULL
);
1044 /* Check if this statevar's setup in the STATE file exactly
1045 * matches how we'll want to reference it as a
1046 * struct/array/whatever. If not, then we need to move it into
1047 * temporary storage and hope that it'll get copy-propagated
1050 for (i
= 0; i
< ir
->num_state_slots
; i
++) {
1051 if (slots
[i
].swizzle
!= SWIZZLE_XYZW
) {
1056 variable_storage
*storage
;
1058 if (i
== ir
->num_state_slots
) {
1059 /* We'll set the index later. */
1060 storage
= new(mem_ctx
) variable_storage(ir
, PROGRAM_STATE_VAR
, -1);
1061 this->variables
.push_tail(storage
);
1065 /* The variable_storage constructor allocates slots based on the size
1066 * of the type. However, this had better match the number of state
1067 * elements that we're going to copy into the new temporary.
1069 assert((int) ir
->num_state_slots
== type_size(ir
->type
));
1071 storage
= new(mem_ctx
) variable_storage(ir
, PROGRAM_TEMPORARY
,
1073 this->variables
.push_tail(storage
);
1074 this->next_temp
+= type_size(ir
->type
);
1076 dst
= st_dst_reg(st_src_reg(PROGRAM_TEMPORARY
, storage
->index
,
1077 native_integers
? ir
->type
->base_type
: GLSL_TYPE_FLOAT
));
1081 for (unsigned int i
= 0; i
< ir
->num_state_slots
; i
++) {
1082 int index
= _mesa_add_state_reference(this->prog
->Parameters
,
1083 (gl_state_index
*)slots
[i
].tokens
);
1085 if (storage
->file
== PROGRAM_STATE_VAR
) {
1086 if (storage
->index
== -1) {
1087 storage
->index
= index
;
1089 assert(index
== storage
->index
+ (int)i
);
1092 st_src_reg
src(PROGRAM_STATE_VAR
, index
,
1093 native_integers
? ir
->type
->base_type
: GLSL_TYPE_FLOAT
);
1094 src
.swizzle
= slots
[i
].swizzle
;
1095 emit(ir
, TGSI_OPCODE_MOV
, dst
, src
);
1096 /* even a float takes up a whole vec4 reg in a struct/array. */
1101 if (storage
->file
== PROGRAM_TEMPORARY
&&
1102 dst
.index
!= storage
->index
+ (int) ir
->num_state_slots
) {
1103 fail_link(this->shader_program
,
1104 "failed to load builtin uniform `%s' (%d/%d regs loaded)\n",
1105 ir
->name
, dst
.index
- storage
->index
,
1106 type_size(ir
->type
));
1112 glsl_to_tgsi_visitor::visit(ir_loop
*ir
)
1114 ir_dereference_variable
*counter
= NULL
;
1116 if (ir
->counter
!= NULL
)
1117 counter
= new(ir
) ir_dereference_variable(ir
->counter
);
1119 if (ir
->from
!= NULL
) {
1120 assert(ir
->counter
!= NULL
);
1122 ir_assignment
*a
= new(ir
) ir_assignment(counter
, ir
->from
, NULL
);
1128 emit(NULL
, TGSI_OPCODE_BGNLOOP
);
1132 new(ir
) ir_expression(ir
->cmp
, glsl_type::bool_type
,
1134 ir_if
*if_stmt
= new(ir
) ir_if(e
);
1136 ir_loop_jump
*brk
= new(ir
) ir_loop_jump(ir_loop_jump::jump_break
);
1138 if_stmt
->then_instructions
.push_tail(brk
);
1140 if_stmt
->accept(this);
1147 visit_exec_list(&ir
->body_instructions
, this);
1149 if (ir
->increment
) {
1151 new(ir
) ir_expression(ir_binop_add
, counter
->type
,
1152 counter
, ir
->increment
);
1154 ir_assignment
*a
= new(ir
) ir_assignment(counter
, e
, NULL
);
1161 emit(NULL
, TGSI_OPCODE_ENDLOOP
);
1165 glsl_to_tgsi_visitor::visit(ir_loop_jump
*ir
)
1168 case ir_loop_jump::jump_break
:
1169 emit(NULL
, TGSI_OPCODE_BRK
);
1171 case ir_loop_jump::jump_continue
:
1172 emit(NULL
, TGSI_OPCODE_CONT
);
1179 glsl_to_tgsi_visitor::visit(ir_function_signature
*ir
)
1186 glsl_to_tgsi_visitor::visit(ir_function
*ir
)
1188 /* Ignore function bodies other than main() -- we shouldn't see calls to
1189 * them since they should all be inlined before we get to glsl_to_tgsi.
1191 if (strcmp(ir
->name
, "main") == 0) {
1192 const ir_function_signature
*sig
;
1195 sig
= ir
->matching_signature(&empty
);
1199 foreach_iter(exec_list_iterator
, iter
, sig
->body
) {
1200 ir_instruction
*ir
= (ir_instruction
*)iter
.get();
1208 glsl_to_tgsi_visitor::try_emit_mad(ir_expression
*ir
, int mul_operand
)
1210 int nonmul_operand
= 1 - mul_operand
;
1212 st_dst_reg result_dst
;
1214 ir_expression
*expr
= ir
->operands
[mul_operand
]->as_expression();
1215 if (!expr
|| expr
->operation
!= ir_binop_mul
)
1218 expr
->operands
[0]->accept(this);
1220 expr
->operands
[1]->accept(this);
1222 ir
->operands
[nonmul_operand
]->accept(this);
1225 this->result
= get_temp(ir
->type
);
1226 result_dst
= st_dst_reg(this->result
);
1227 result_dst
.writemask
= (1 << ir
->type
->vector_elements
) - 1;
1228 emit(ir
, TGSI_OPCODE_MAD
, result_dst
, a
, b
, c
);
1234 * Emit MAD(a, -b, a) instead of AND(a, NOT(b))
1236 * The logic values are 1.0 for true and 0.0 for false. Logical-and is
1237 * implemented using multiplication, and logical-or is implemented using
1238 * addition. Logical-not can be implemented as (true - x), or (1.0 - x).
1239 * As result, the logical expression (a & !b) can be rewritten as:
1243 * - (a * 1) - (a * b)
1247 * This final expression can be implemented as a single MAD(a, -b, a)
1251 glsl_to_tgsi_visitor::try_emit_mad_for_and_not(ir_expression
*ir
, int try_operand
)
1253 const int other_operand
= 1 - try_operand
;
1256 ir_expression
*expr
= ir
->operands
[try_operand
]->as_expression();
1257 if (!expr
|| expr
->operation
!= ir_unop_logic_not
)
1260 ir
->operands
[other_operand
]->accept(this);
1262 expr
->operands
[0]->accept(this);
1265 b
.negate
= ~b
.negate
;
1267 this->result
= get_temp(ir
->type
);
1268 emit(ir
, TGSI_OPCODE_MAD
, st_dst_reg(this->result
), a
, b
, a
);
1274 glsl_to_tgsi_visitor::try_emit_sat(ir_expression
*ir
)
1276 /* Saturates were only introduced to vertex programs in
1277 * NV_vertex_program3, so don't give them to drivers in the VP.
1279 if (this->prog
->Target
== GL_VERTEX_PROGRAM_ARB
)
1282 ir_rvalue
*sat_src
= ir
->as_rvalue_to_saturate();
1286 sat_src
->accept(this);
1287 st_src_reg src
= this->result
;
1289 /* If we generated an expression instruction into a temporary in
1290 * processing the saturate's operand, apply the saturate to that
1291 * instruction. Otherwise, generate a MOV to do the saturate.
1293 * Note that we have to be careful to only do this optimization if
1294 * the instruction in question was what generated src->result. For
1295 * example, ir_dereference_array might generate a MUL instruction
1296 * to create the reladdr, and return us a src reg using that
1297 * reladdr. That MUL result is not the value we're trying to
1300 ir_expression
*sat_src_expr
= sat_src
->as_expression();
1301 if (sat_src_expr
&& (sat_src_expr
->operation
== ir_binop_mul
||
1302 sat_src_expr
->operation
== ir_binop_add
||
1303 sat_src_expr
->operation
== ir_binop_dot
)) {
1304 glsl_to_tgsi_instruction
*new_inst
;
1305 new_inst
= (glsl_to_tgsi_instruction
*)this->instructions
.get_tail();
1306 new_inst
->saturate
= true;
1308 this->result
= get_temp(ir
->type
);
1309 st_dst_reg result_dst
= st_dst_reg(this->result
);
1310 result_dst
.writemask
= (1 << ir
->type
->vector_elements
) - 1;
1311 glsl_to_tgsi_instruction
*inst
;
1312 inst
= emit(ir
, TGSI_OPCODE_MOV
, result_dst
, src
);
1313 inst
->saturate
= true;
1320 glsl_to_tgsi_visitor::reladdr_to_temp(ir_instruction
*ir
,
1321 st_src_reg
*reg
, int *num_reladdr
)
1326 emit_arl(ir
, address_reg
, *reg
->reladdr
);
1328 if (*num_reladdr
!= 1) {
1329 st_src_reg temp
= get_temp(glsl_type::vec4_type
);
1331 emit(ir
, TGSI_OPCODE_MOV
, st_dst_reg(temp
), *reg
);
1339 glsl_to_tgsi_visitor::visit(ir_expression
*ir
)
1341 unsigned int operand
;
1342 st_src_reg op
[Elements(ir
->operands
)];
1343 st_src_reg result_src
;
1344 st_dst_reg result_dst
;
1346 /* Quick peephole: Emit MAD(a, b, c) instead of ADD(MUL(a, b), c)
1348 if (ir
->operation
== ir_binop_add
) {
1349 if (try_emit_mad(ir
, 1))
1351 if (try_emit_mad(ir
, 0))
1355 /* Quick peephole: Emit OPCODE_MAD(-a, -b, a) instead of AND(a, NOT(b))
1357 if (ir
->operation
== ir_binop_logic_and
) {
1358 if (try_emit_mad_for_and_not(ir
, 1))
1360 if (try_emit_mad_for_and_not(ir
, 0))
1364 if (try_emit_sat(ir
))
1367 if (ir
->operation
== ir_quadop_vector
)
1368 assert(!"ir_quadop_vector should have been lowered");
1370 for (operand
= 0; operand
< ir
->get_num_operands(); operand
++) {
1371 this->result
.file
= PROGRAM_UNDEFINED
;
1372 ir
->operands
[operand
]->accept(this);
1373 if (this->result
.file
== PROGRAM_UNDEFINED
) {
1375 printf("Failed to get tree for expression operand:\n");
1376 ir
->operands
[operand
]->accept(&v
);
1379 op
[operand
] = this->result
;
1381 /* Matrix expression operands should have been broken down to vector
1382 * operations already.
1384 assert(!ir
->operands
[operand
]->type
->is_matrix());
1387 int vector_elements
= ir
->operands
[0]->type
->vector_elements
;
1388 if (ir
->operands
[1]) {
1389 vector_elements
= MAX2(vector_elements
,
1390 ir
->operands
[1]->type
->vector_elements
);
1393 this->result
.file
= PROGRAM_UNDEFINED
;
1395 /* Storage for our result. Ideally for an assignment we'd be using
1396 * the actual storage for the result here, instead.
1398 result_src
= get_temp(ir
->type
);
1399 /* convenience for the emit functions below. */
1400 result_dst
= st_dst_reg(result_src
);
1401 /* Limit writes to the channels that will be used by result_src later.
1402 * This does limit this temp's use as a temporary for multi-instruction
1405 result_dst
.writemask
= (1 << ir
->type
->vector_elements
) - 1;
1407 switch (ir
->operation
) {
1408 case ir_unop_logic_not
:
1409 if (result_dst
.type
!= GLSL_TYPE_FLOAT
)
1410 emit(ir
, TGSI_OPCODE_NOT
, result_dst
, op
[0]);
1412 /* Previously 'SEQ dst, src, 0.0' was used for this. However, many
1413 * older GPUs implement SEQ using multiple instructions (i915 uses two
1414 * SGE instructions and a MUL instruction). Since our logic values are
1415 * 0.0 and 1.0, 1-x also implements !x.
1417 op
[0].negate
= ~op
[0].negate
;
1418 emit(ir
, TGSI_OPCODE_ADD
, result_dst
, op
[0], st_src_reg_for_float(1.0));
1422 assert(result_dst
.type
== GLSL_TYPE_FLOAT
|| result_dst
.type
== GLSL_TYPE_INT
);
1423 if (result_dst
.type
== GLSL_TYPE_INT
)
1424 emit(ir
, TGSI_OPCODE_INEG
, result_dst
, op
[0]);
1426 op
[0].negate
= ~op
[0].negate
;
1431 assert(result_dst
.type
== GLSL_TYPE_FLOAT
);
1432 emit(ir
, TGSI_OPCODE_ABS
, result_dst
, op
[0]);
1435 emit(ir
, TGSI_OPCODE_SSG
, result_dst
, op
[0]);
1438 emit_scalar(ir
, TGSI_OPCODE_RCP
, result_dst
, op
[0]);
1442 emit_scalar(ir
, TGSI_OPCODE_EX2
, result_dst
, op
[0]);
1446 assert(!"not reached: should be handled by ir_explog_to_explog2");
1449 emit_scalar(ir
, TGSI_OPCODE_LG2
, result_dst
, op
[0]);
1452 emit_scalar(ir
, TGSI_OPCODE_SIN
, result_dst
, op
[0]);
1455 emit_scalar(ir
, TGSI_OPCODE_COS
, result_dst
, op
[0]);
1457 case ir_unop_sin_reduced
:
1458 emit_scs(ir
, TGSI_OPCODE_SIN
, result_dst
, op
[0]);
1460 case ir_unop_cos_reduced
:
1461 emit_scs(ir
, TGSI_OPCODE_COS
, result_dst
, op
[0]);
1465 emit(ir
, TGSI_OPCODE_DDX
, result_dst
, op
[0]);
1468 op
[0].negate
= ~op
[0].negate
;
1469 emit(ir
, TGSI_OPCODE_DDY
, result_dst
, op
[0]);
1472 case ir_unop_noise
: {
1473 /* At some point, a motivated person could add a better
1474 * implementation of noise. Currently not even the nvidia
1475 * binary drivers do anything more than this. In any case, the
1476 * place to do this is in the GL state tracker, not the poor
1479 emit(ir
, TGSI_OPCODE_MOV
, result_dst
, st_src_reg_for_float(0.5));
1484 emit(ir
, TGSI_OPCODE_ADD
, result_dst
, op
[0], op
[1]);
1487 emit(ir
, TGSI_OPCODE_SUB
, result_dst
, op
[0], op
[1]);
1491 emit(ir
, TGSI_OPCODE_MUL
, result_dst
, op
[0], op
[1]);
1494 if (result_dst
.type
== GLSL_TYPE_FLOAT
)
1495 assert(!"not reached: should be handled by ir_div_to_mul_rcp");
1497 emit(ir
, TGSI_OPCODE_DIV
, result_dst
, op
[0], op
[1]);
1500 if (result_dst
.type
== GLSL_TYPE_FLOAT
)
1501 assert(!"ir_binop_mod should have been converted to b * fract(a/b)");
1503 emit(ir
, TGSI_OPCODE_MOD
, result_dst
, op
[0], op
[1]);
1507 emit(ir
, TGSI_OPCODE_SLT
, result_dst
, op
[0], op
[1]);
1509 case ir_binop_greater
:
1510 emit(ir
, TGSI_OPCODE_SLT
, result_dst
, op
[1], op
[0]);
1512 case ir_binop_lequal
:
1513 emit(ir
, TGSI_OPCODE_SGE
, result_dst
, op
[1], op
[0]);
1515 case ir_binop_gequal
:
1516 emit(ir
, TGSI_OPCODE_SGE
, result_dst
, op
[0], op
[1]);
1518 case ir_binop_equal
:
1519 emit(ir
, TGSI_OPCODE_SEQ
, result_dst
, op
[0], op
[1]);
1521 case ir_binop_nequal
:
1522 emit(ir
, TGSI_OPCODE_SNE
, result_dst
, op
[0], op
[1]);
1524 case ir_binop_all_equal
:
1525 /* "==" operator producing a scalar boolean. */
1526 if (ir
->operands
[0]->type
->is_vector() ||
1527 ir
->operands
[1]->type
->is_vector()) {
1528 st_src_reg temp
= get_temp(native_integers
?
1529 glsl_type::get_instance(ir
->operands
[0]->type
->base_type
, 4, 1) :
1530 glsl_type::vec4_type
);
1532 if (native_integers
) {
1533 st_dst_reg temp_dst
= st_dst_reg(temp
);
1534 st_src_reg temp1
= st_src_reg(temp
), temp2
= st_src_reg(temp
);
1536 emit(ir
, TGSI_OPCODE_SEQ
, st_dst_reg(temp
), op
[0], op
[1]);
1538 /* Emit 1-3 AND operations to combine the SEQ results. */
1539 switch (ir
->operands
[0]->type
->vector_elements
) {
1543 temp_dst
.writemask
= WRITEMASK_Y
;
1544 temp1
.swizzle
= SWIZZLE_YYYY
;
1545 temp2
.swizzle
= SWIZZLE_ZZZZ
;
1546 emit(ir
, TGSI_OPCODE_AND
, temp_dst
, temp1
, temp2
);
1549 temp_dst
.writemask
= WRITEMASK_X
;
1550 temp1
.swizzle
= SWIZZLE_XXXX
;
1551 temp2
.swizzle
= SWIZZLE_YYYY
;
1552 emit(ir
, TGSI_OPCODE_AND
, temp_dst
, temp1
, temp2
);
1553 temp_dst
.writemask
= WRITEMASK_Y
;
1554 temp1
.swizzle
= SWIZZLE_ZZZZ
;
1555 temp2
.swizzle
= SWIZZLE_WWWW
;
1556 emit(ir
, TGSI_OPCODE_AND
, temp_dst
, temp1
, temp2
);
1559 temp1
.swizzle
= SWIZZLE_XXXX
;
1560 temp2
.swizzle
= SWIZZLE_YYYY
;
1561 emit(ir
, TGSI_OPCODE_AND
, result_dst
, temp1
, temp2
);
1563 emit(ir
, TGSI_OPCODE_SNE
, st_dst_reg(temp
), op
[0], op
[1]);
1565 /* After the dot-product, the value will be an integer on the
1566 * range [0,4]. Zero becomes 1.0, and positive values become zero.
1568 emit_dp(ir
, result_dst
, temp
, temp
, vector_elements
);
1570 /* Negating the result of the dot-product gives values on the range
1571 * [-4, 0]. Zero becomes 1.0, and negative values become zero.
1572 * This is achieved using SGE.
1574 st_src_reg sge_src
= result_src
;
1575 sge_src
.negate
= ~sge_src
.negate
;
1576 emit(ir
, TGSI_OPCODE_SGE
, result_dst
, sge_src
, st_src_reg_for_float(0.0));
1579 emit(ir
, TGSI_OPCODE_SEQ
, result_dst
, op
[0], op
[1]);
1582 case ir_binop_any_nequal
:
1583 /* "!=" operator producing a scalar boolean. */
1584 if (ir
->operands
[0]->type
->is_vector() ||
1585 ir
->operands
[1]->type
->is_vector()) {
1586 st_src_reg temp
= get_temp(native_integers
?
1587 glsl_type::get_instance(ir
->operands
[0]->type
->base_type
, 4, 1) :
1588 glsl_type::vec4_type
);
1589 emit(ir
, TGSI_OPCODE_SNE
, st_dst_reg(temp
), op
[0], op
[1]);
1591 if (native_integers
) {
1592 st_dst_reg temp_dst
= st_dst_reg(temp
);
1593 st_src_reg temp1
= st_src_reg(temp
), temp2
= st_src_reg(temp
);
1595 /* Emit 1-3 OR operations to combine the SNE results. */
1596 switch (ir
->operands
[0]->type
->vector_elements
) {
1600 temp_dst
.writemask
= WRITEMASK_Y
;
1601 temp1
.swizzle
= SWIZZLE_YYYY
;
1602 temp2
.swizzle
= SWIZZLE_ZZZZ
;
1603 emit(ir
, TGSI_OPCODE_OR
, temp_dst
, temp1
, temp2
);
1606 temp_dst
.writemask
= WRITEMASK_X
;
1607 temp1
.swizzle
= SWIZZLE_XXXX
;
1608 temp2
.swizzle
= SWIZZLE_YYYY
;
1609 emit(ir
, TGSI_OPCODE_OR
, temp_dst
, temp1
, temp2
);
1610 temp_dst
.writemask
= WRITEMASK_Y
;
1611 temp1
.swizzle
= SWIZZLE_ZZZZ
;
1612 temp2
.swizzle
= SWIZZLE_WWWW
;
1613 emit(ir
, TGSI_OPCODE_OR
, temp_dst
, temp1
, temp2
);
1616 temp1
.swizzle
= SWIZZLE_XXXX
;
1617 temp2
.swizzle
= SWIZZLE_YYYY
;
1618 emit(ir
, TGSI_OPCODE_OR
, result_dst
, temp1
, temp2
);
1620 /* After the dot-product, the value will be an integer on the
1621 * range [0,4]. Zero stays zero, and positive values become 1.0.
1623 glsl_to_tgsi_instruction
*const dp
=
1624 emit_dp(ir
, result_dst
, temp
, temp
, vector_elements
);
1625 if (this->prog
->Target
== GL_FRAGMENT_PROGRAM_ARB
) {
1626 /* The clamping to [0,1] can be done for free in the fragment
1627 * shader with a saturate.
1629 dp
->saturate
= true;
1631 /* Negating the result of the dot-product gives values on the range
1632 * [-4, 0]. Zero stays zero, and negative values become 1.0. This
1633 * achieved using SLT.
1635 st_src_reg slt_src
= result_src
;
1636 slt_src
.negate
= ~slt_src
.negate
;
1637 emit(ir
, TGSI_OPCODE_SLT
, result_dst
, slt_src
, st_src_reg_for_float(0.0));
1641 emit(ir
, TGSI_OPCODE_SNE
, result_dst
, op
[0], op
[1]);
1646 assert(ir
->operands
[0]->type
->is_vector());
1648 /* After the dot-product, the value will be an integer on the
1649 * range [0,4]. Zero stays zero, and positive values become 1.0.
1651 glsl_to_tgsi_instruction
*const dp
=
1652 emit_dp(ir
, result_dst
, op
[0], op
[0],
1653 ir
->operands
[0]->type
->vector_elements
);
1654 if (this->prog
->Target
== GL_FRAGMENT_PROGRAM_ARB
&&
1655 result_dst
.type
== GLSL_TYPE_FLOAT
) {
1656 /* The clamping to [0,1] can be done for free in the fragment
1657 * shader with a saturate.
1659 dp
->saturate
= true;
1660 } else if (result_dst
.type
== GLSL_TYPE_FLOAT
) {
1661 /* Negating the result of the dot-product gives values on the range
1662 * [-4, 0]. Zero stays zero, and negative values become 1.0. This
1663 * is achieved using SLT.
1665 st_src_reg slt_src
= result_src
;
1666 slt_src
.negate
= ~slt_src
.negate
;
1667 emit(ir
, TGSI_OPCODE_SLT
, result_dst
, slt_src
, st_src_reg_for_float(0.0));
1670 /* Use SNE 0 if integers are being used as boolean values. */
1671 emit(ir
, TGSI_OPCODE_SNE
, result_dst
, result_src
, st_src_reg_for_int(0));
1676 case ir_binop_logic_xor
:
1677 if (native_integers
)
1678 emit(ir
, TGSI_OPCODE_XOR
, result_dst
, op
[0], op
[1]);
1680 emit(ir
, TGSI_OPCODE_SNE
, result_dst
, op
[0], op
[1]);
1683 case ir_binop_logic_or
: {
1684 if (native_integers
) {
1685 /* If integers are used as booleans, we can use an actual "or"
1688 assert(native_integers
);
1689 emit(ir
, TGSI_OPCODE_OR
, result_dst
, op
[0], op
[1]);
1691 /* After the addition, the value will be an integer on the
1692 * range [0,2]. Zero stays zero, and positive values become 1.0.
1694 glsl_to_tgsi_instruction
*add
=
1695 emit(ir
, TGSI_OPCODE_ADD
, result_dst
, op
[0], op
[1]);
1696 if (this->prog
->Target
== GL_FRAGMENT_PROGRAM_ARB
) {
1697 /* The clamping to [0,1] can be done for free in the fragment
1698 * shader with a saturate if floats are being used as boolean values.
1700 add
->saturate
= true;
1702 /* Negating the result of the addition gives values on the range
1703 * [-2, 0]. Zero stays zero, and negative values become 1.0. This
1704 * is achieved using SLT.
1706 st_src_reg slt_src
= result_src
;
1707 slt_src
.negate
= ~slt_src
.negate
;
1708 emit(ir
, TGSI_OPCODE_SLT
, result_dst
, slt_src
, st_src_reg_for_float(0.0));
1714 case ir_binop_logic_and
:
1715 /* If native integers are disabled, the bool args are stored as float 0.0
1716 * or 1.0, so "mul" gives us "and". If they're enabled, just use the
1717 * actual AND opcode.
1719 if (native_integers
)
1720 emit(ir
, TGSI_OPCODE_AND
, result_dst
, op
[0], op
[1]);
1722 emit(ir
, TGSI_OPCODE_MUL
, result_dst
, op
[0], op
[1]);
1726 assert(ir
->operands
[0]->type
->is_vector());
1727 assert(ir
->operands
[0]->type
== ir
->operands
[1]->type
);
1728 emit_dp(ir
, result_dst
, op
[0], op
[1],
1729 ir
->operands
[0]->type
->vector_elements
);
1733 /* sqrt(x) = x * rsq(x). */
1734 emit_scalar(ir
, TGSI_OPCODE_RSQ
, result_dst
, op
[0]);
1735 emit(ir
, TGSI_OPCODE_MUL
, result_dst
, result_src
, op
[0]);
1736 /* For incoming channels <= 0, set the result to 0. */
1737 op
[0].negate
= ~op
[0].negate
;
1738 emit(ir
, TGSI_OPCODE_CMP
, result_dst
,
1739 op
[0], result_src
, st_src_reg_for_float(0.0));
1742 emit_scalar(ir
, TGSI_OPCODE_RSQ
, result_dst
, op
[0]);
1745 if (native_integers
) {
1746 emit(ir
, TGSI_OPCODE_I2F
, result_dst
, op
[0]);
1749 /* fallthrough to next case otherwise */
1751 if (native_integers
) {
1752 emit(ir
, TGSI_OPCODE_AND
, result_dst
, op
[0], st_src_reg_for_float(1.0));
1755 /* fallthrough to next case otherwise */
1758 /* Converting between signed and unsigned integers is a no-op. */
1762 if (native_integers
) {
1763 /* Booleans are stored as integers using ~0 for true and 0 for false.
1764 * GLSL requires that int(bool) return 1 for true and 0 for false.
1765 * This conversion is done with AND, but it could be done with NEG.
1767 emit(ir
, TGSI_OPCODE_AND
, result_dst
, op
[0], st_src_reg_for_int(1));
1769 /* Booleans and integers are both stored as floats when native
1770 * integers are disabled.
1776 if (native_integers
)
1777 emit(ir
, TGSI_OPCODE_F2I
, result_dst
, op
[0]);
1779 emit(ir
, TGSI_OPCODE_TRUNC
, result_dst
, op
[0]);
1782 emit(ir
, TGSI_OPCODE_SNE
, result_dst
, op
[0], st_src_reg_for_float(0.0));
1785 if (native_integers
)
1786 emit(ir
, TGSI_OPCODE_INEG
, result_dst
, op
[0]);
1788 emit(ir
, TGSI_OPCODE_SNE
, result_dst
, op
[0], st_src_reg_for_float(0.0));
1791 emit(ir
, TGSI_OPCODE_TRUNC
, result_dst
, op
[0]);
1794 op
[0].negate
= ~op
[0].negate
;
1795 emit(ir
, TGSI_OPCODE_FLR
, result_dst
, op
[0]);
1796 result_src
.negate
= ~result_src
.negate
;
1799 emit(ir
, TGSI_OPCODE_FLR
, result_dst
, op
[0]);
1802 emit(ir
, TGSI_OPCODE_FRC
, result_dst
, op
[0]);
1806 emit(ir
, TGSI_OPCODE_MIN
, result_dst
, op
[0], op
[1]);
1809 emit(ir
, TGSI_OPCODE_MAX
, result_dst
, op
[0], op
[1]);
1812 emit_scalar(ir
, TGSI_OPCODE_POW
, result_dst
, op
[0], op
[1]);
1815 case ir_unop_bit_not
:
1816 if (native_integers
) {
1817 emit(ir
, TGSI_OPCODE_NOT
, result_dst
, op
[0]);
1821 if (native_integers
) {
1822 emit(ir
, TGSI_OPCODE_U2F
, result_dst
, op
[0]);
1825 case ir_binop_lshift
:
1826 if (native_integers
) {
1827 emit(ir
, TGSI_OPCODE_SHL
, result_dst
, op
[0]);
1830 case ir_binop_rshift
:
1831 if (native_integers
) {
1832 emit(ir
, TGSI_OPCODE_ISHR
, result_dst
, op
[0]);
1835 case ir_binop_bit_and
:
1836 if (native_integers
) {
1837 emit(ir
, TGSI_OPCODE_AND
, result_dst
, op
[0]);
1840 case ir_binop_bit_xor
:
1841 if (native_integers
) {
1842 emit(ir
, TGSI_OPCODE_XOR
, result_dst
, op
[0]);
1845 case ir_binop_bit_or
:
1846 if (native_integers
) {
1847 emit(ir
, TGSI_OPCODE_OR
, result_dst
, op
[0]);
1850 case ir_unop_round_even
:
1851 assert(!"GLSL 1.30 features unsupported");
1854 case ir_quadop_vector
:
1855 /* This operation should have already been handled.
1857 assert(!"Should not get here.");
1861 this->result
= result_src
;
1866 glsl_to_tgsi_visitor::visit(ir_swizzle
*ir
)
1872 /* Note that this is only swizzles in expressions, not those on the left
1873 * hand side of an assignment, which do write masking. See ir_assignment
1877 ir
->val
->accept(this);
1879 assert(src
.file
!= PROGRAM_UNDEFINED
);
1881 for (i
= 0; i
< 4; i
++) {
1882 if (i
< ir
->type
->vector_elements
) {
1885 swizzle
[i
] = GET_SWZ(src
.swizzle
, ir
->mask
.x
);
1888 swizzle
[i
] = GET_SWZ(src
.swizzle
, ir
->mask
.y
);
1891 swizzle
[i
] = GET_SWZ(src
.swizzle
, ir
->mask
.z
);
1894 swizzle
[i
] = GET_SWZ(src
.swizzle
, ir
->mask
.w
);
1898 /* If the type is smaller than a vec4, replicate the last
1901 swizzle
[i
] = swizzle
[ir
->type
->vector_elements
- 1];
1905 src
.swizzle
= MAKE_SWIZZLE4(swizzle
[0], swizzle
[1], swizzle
[2], swizzle
[3]);
1911 glsl_to_tgsi_visitor::visit(ir_dereference_variable
*ir
)
1913 variable_storage
*entry
= find_variable_storage(ir
->var
);
1914 ir_variable
*var
= ir
->var
;
1917 switch (var
->mode
) {
1918 case ir_var_uniform
:
1919 entry
= new(mem_ctx
) variable_storage(var
, PROGRAM_UNIFORM
,
1921 this->variables
.push_tail(entry
);
1925 /* The linker assigns locations for varyings and attributes,
1926 * including deprecated builtins (like gl_Color), user-assign
1927 * generic attributes (glBindVertexLocation), and
1928 * user-defined varyings.
1930 * FINISHME: We would hit this path for function arguments. Fix!
1932 assert(var
->location
!= -1);
1933 entry
= new(mem_ctx
) variable_storage(var
,
1938 assert(var
->location
!= -1);
1939 entry
= new(mem_ctx
) variable_storage(var
,
1943 case ir_var_system_value
:
1944 entry
= new(mem_ctx
) variable_storage(var
,
1945 PROGRAM_SYSTEM_VALUE
,
1949 case ir_var_temporary
:
1950 entry
= new(mem_ctx
) variable_storage(var
, PROGRAM_TEMPORARY
,
1952 this->variables
.push_tail(entry
);
1954 next_temp
+= type_size(var
->type
);
1959 printf("Failed to make storage for %s\n", var
->name
);
1964 this->result
= st_src_reg(entry
->file
, entry
->index
, var
->type
);
1965 if (!native_integers
)
1966 this->result
.type
= GLSL_TYPE_FLOAT
;
1970 glsl_to_tgsi_visitor::visit(ir_dereference_array
*ir
)
1974 int element_size
= type_size(ir
->type
);
1976 index
= ir
->array_index
->constant_expression_value();
1978 ir
->array
->accept(this);
1982 src
.index
+= index
->value
.i
[0] * element_size
;
1984 /* Variable index array dereference. It eats the "vec4" of the
1985 * base of the array and an index that offsets the TGSI register
1988 ir
->array_index
->accept(this);
1990 st_src_reg index_reg
;
1992 if (element_size
== 1) {
1993 index_reg
= this->result
;
1995 index_reg
= get_temp(native_integers
?
1996 glsl_type::int_type
: glsl_type::float_type
);
1998 emit(ir
, TGSI_OPCODE_MUL
, st_dst_reg(index_reg
),
1999 this->result
, st_src_reg_for_type(index_reg
.type
, element_size
));
2002 /* If there was already a relative address register involved, add the
2003 * new and the old together to get the new offset.
2005 if (src
.reladdr
!= NULL
) {
2006 st_src_reg accum_reg
= get_temp(native_integers
?
2007 glsl_type::int_type
: glsl_type::float_type
);
2009 emit(ir
, TGSI_OPCODE_ADD
, st_dst_reg(accum_reg
),
2010 index_reg
, *src
.reladdr
);
2012 index_reg
= accum_reg
;
2015 src
.reladdr
= ralloc(mem_ctx
, st_src_reg
);
2016 memcpy(src
.reladdr
, &index_reg
, sizeof(index_reg
));
2019 /* If the type is smaller than a vec4, replicate the last channel out. */
2020 if (ir
->type
->is_scalar() || ir
->type
->is_vector())
2021 src
.swizzle
= swizzle_for_size(ir
->type
->vector_elements
);
2023 src
.swizzle
= SWIZZLE_NOOP
;
2029 glsl_to_tgsi_visitor::visit(ir_dereference_record
*ir
)
2032 const glsl_type
*struct_type
= ir
->record
->type
;
2035 ir
->record
->accept(this);
2037 for (i
= 0; i
< struct_type
->length
; i
++) {
2038 if (strcmp(struct_type
->fields
.structure
[i
].name
, ir
->field
) == 0)
2040 offset
+= type_size(struct_type
->fields
.structure
[i
].type
);
2043 /* If the type is smaller than a vec4, replicate the last channel out. */
2044 if (ir
->type
->is_scalar() || ir
->type
->is_vector())
2045 this->result
.swizzle
= swizzle_for_size(ir
->type
->vector_elements
);
2047 this->result
.swizzle
= SWIZZLE_NOOP
;
2049 this->result
.index
+= offset
;
2053 * We want to be careful in assignment setup to hit the actual storage
2054 * instead of potentially using a temporary like we might with the
2055 * ir_dereference handler.
2058 get_assignment_lhs(ir_dereference
*ir
, glsl_to_tgsi_visitor
*v
)
2060 /* The LHS must be a dereference. If the LHS is a variable indexed array
2061 * access of a vector, it must be separated into a series conditional moves
2062 * before reaching this point (see ir_vec_index_to_cond_assign).
2064 assert(ir
->as_dereference());
2065 ir_dereference_array
*deref_array
= ir
->as_dereference_array();
2067 assert(!deref_array
->array
->type
->is_vector());
2070 /* Use the rvalue deref handler for the most part. We'll ignore
2071 * swizzles in it and write swizzles using writemask, though.
2074 return st_dst_reg(v
->result
);
2078 * Process the condition of a conditional assignment
2080 * Examines the condition of a conditional assignment to generate the optimal
2081 * first operand of a \c CMP instruction. If the condition is a relational
2082 * operator with 0 (e.g., \c ir_binop_less), the value being compared will be
2083 * used as the source for the \c CMP instruction. Otherwise the comparison
2084 * is processed to a boolean result, and the boolean result is used as the
2085 * operand to the CMP instruction.
2088 glsl_to_tgsi_visitor::process_move_condition(ir_rvalue
*ir
)
2090 ir_rvalue
*src_ir
= ir
;
2092 bool switch_order
= false;
2094 ir_expression
*const expr
= ir
->as_expression();
2095 if ((expr
!= NULL
) && (expr
->get_num_operands() == 2)) {
2096 bool zero_on_left
= false;
2098 if (expr
->operands
[0]->is_zero()) {
2099 src_ir
= expr
->operands
[1];
2100 zero_on_left
= true;
2101 } else if (expr
->operands
[1]->is_zero()) {
2102 src_ir
= expr
->operands
[0];
2103 zero_on_left
= false;
2107 * (a < 0) T F F ( a < 0) T F F
2108 * (0 < a) F F T (-a < 0) F F T
2109 * (a <= 0) T T F (-a < 0) F F T (swap order of other operands)
2110 * (0 <= a) F T T ( a < 0) T F F (swap order of other operands)
2111 * (a > 0) F F T (-a < 0) F F T
2112 * (0 > a) T F F ( a < 0) T F F
2113 * (a >= 0) F T T ( a < 0) T F F (swap order of other operands)
2114 * (0 >= a) T T F (-a < 0) F F T (swap order of other operands)
2116 * Note that exchanging the order of 0 and 'a' in the comparison simply
2117 * means that the value of 'a' should be negated.
2120 switch (expr
->operation
) {
2122 switch_order
= false;
2123 negate
= zero_on_left
;
2126 case ir_binop_greater
:
2127 switch_order
= false;
2128 negate
= !zero_on_left
;
2131 case ir_binop_lequal
:
2132 switch_order
= true;
2133 negate
= !zero_on_left
;
2136 case ir_binop_gequal
:
2137 switch_order
= true;
2138 negate
= zero_on_left
;
2142 /* This isn't the right kind of comparison afterall, so make sure
2143 * the whole condition is visited.
2151 src_ir
->accept(this);
2153 /* We use the TGSI_OPCODE_CMP (a < 0 ? b : c) for conditional moves, and the
2154 * condition we produced is 0.0 or 1.0. By flipping the sign, we can
2155 * choose which value TGSI_OPCODE_CMP produces without an extra instruction
2156 * computing the condition.
2159 this->result
.negate
= ~this->result
.negate
;
2161 return switch_order
;
2165 glsl_to_tgsi_visitor::visit(ir_assignment
*ir
)
2171 ir
->rhs
->accept(this);
2174 l
= get_assignment_lhs(ir
->lhs
, this);
2176 /* FINISHME: This should really set to the correct maximal writemask for each
2177 * FINISHME: component written (in the loops below). This case can only
2178 * FINISHME: occur for matrices, arrays, and structures.
2180 if (ir
->write_mask
== 0) {
2181 assert(!ir
->lhs
->type
->is_scalar() && !ir
->lhs
->type
->is_vector());
2182 l
.writemask
= WRITEMASK_XYZW
;
2183 } else if (ir
->lhs
->type
->is_scalar() &&
2184 ir
->lhs
->variable_referenced()->mode
== ir_var_out
) {
2185 /* FINISHME: This hack makes writing to gl_FragDepth, which lives in the
2186 * FINISHME: W component of fragment shader output zero, work correctly.
2188 l
.writemask
= WRITEMASK_XYZW
;
2191 int first_enabled_chan
= 0;
2194 l
.writemask
= ir
->write_mask
;
2196 for (int i
= 0; i
< 4; i
++) {
2197 if (l
.writemask
& (1 << i
)) {
2198 first_enabled_chan
= GET_SWZ(r
.swizzle
, i
);
2203 /* Swizzle a small RHS vector into the channels being written.
2205 * glsl ir treats write_mask as dictating how many channels are
2206 * present on the RHS while TGSI treats write_mask as just
2207 * showing which channels of the vec4 RHS get written.
2209 for (int i
= 0; i
< 4; i
++) {
2210 if (l
.writemask
& (1 << i
))
2211 swizzles
[i
] = GET_SWZ(r
.swizzle
, rhs_chan
++);
2213 swizzles
[i
] = first_enabled_chan
;
2215 r
.swizzle
= MAKE_SWIZZLE4(swizzles
[0], swizzles
[1],
2216 swizzles
[2], swizzles
[3]);
2219 assert(l
.file
!= PROGRAM_UNDEFINED
);
2220 assert(r
.file
!= PROGRAM_UNDEFINED
);
2222 if (ir
->condition
) {
2223 const bool switch_order
= this->process_move_condition(ir
->condition
);
2224 st_src_reg condition
= this->result
;
2226 for (i
= 0; i
< type_size(ir
->lhs
->type
); i
++) {
2227 st_src_reg l_src
= st_src_reg(l
);
2228 st_src_reg condition_temp
= condition
;
2229 l_src
.swizzle
= swizzle_for_size(ir
->lhs
->type
->vector_elements
);
2231 if (native_integers
) {
2232 /* This is necessary because TGSI's CMP instruction expects the
2233 * condition to be a float, and we store booleans as integers.
2234 * If TGSI had a UCMP instruction or similar, this extra
2235 * instruction would not be necessary.
2237 condition_temp
= get_temp(glsl_type::vec4_type
);
2238 condition
.negate
= 0;
2239 emit(ir
, TGSI_OPCODE_I2F
, st_dst_reg(condition_temp
), condition
);
2240 condition_temp
.swizzle
= condition
.swizzle
;
2244 emit(ir
, TGSI_OPCODE_CMP
, l
, condition_temp
, l_src
, r
);
2246 emit(ir
, TGSI_OPCODE_CMP
, l
, condition_temp
, r
, l_src
);
2252 } else if (ir
->rhs
->as_expression() &&
2253 this->instructions
.get_tail() &&
2254 ir
->rhs
== ((glsl_to_tgsi_instruction
*)this->instructions
.get_tail())->ir
&&
2255 type_size(ir
->lhs
->type
) == 1 &&
2256 l
.writemask
== ((glsl_to_tgsi_instruction
*)this->instructions
.get_tail())->dst
.writemask
) {
2257 /* To avoid emitting an extra MOV when assigning an expression to a
2258 * variable, emit the last instruction of the expression again, but
2259 * replace the destination register with the target of the assignment.
2260 * Dead code elimination will remove the original instruction.
2262 glsl_to_tgsi_instruction
*inst
, *new_inst
;
2263 inst
= (glsl_to_tgsi_instruction
*)this->instructions
.get_tail();
2264 new_inst
= emit(ir
, inst
->op
, l
, inst
->src
[0], inst
->src
[1], inst
->src
[2]);
2265 new_inst
->saturate
= inst
->saturate
;
2266 inst
->dead_mask
= inst
->dst
.writemask
;
2268 for (i
= 0; i
< type_size(ir
->lhs
->type
); i
++) {
2269 emit(ir
, TGSI_OPCODE_MOV
, l
, r
);
2278 glsl_to_tgsi_visitor::visit(ir_constant
*ir
)
2281 GLfloat stack_vals
[4] = { 0 };
2282 gl_constant_value
*values
= (gl_constant_value
*) stack_vals
;
2283 GLenum gl_type
= GL_NONE
;
2285 static int in_array
= 0;
2286 gl_register_file file
= in_array
? PROGRAM_CONSTANT
: PROGRAM_IMMEDIATE
;
2288 /* Unfortunately, 4 floats is all we can get into
2289 * _mesa_add_typed_unnamed_constant. So, make a temp to store an
2290 * aggregate constant and move each constant value into it. If we
2291 * get lucky, copy propagation will eliminate the extra moves.
2293 if (ir
->type
->base_type
== GLSL_TYPE_STRUCT
) {
2294 st_src_reg temp_base
= get_temp(ir
->type
);
2295 st_dst_reg temp
= st_dst_reg(temp_base
);
2297 foreach_iter(exec_list_iterator
, iter
, ir
->components
) {
2298 ir_constant
*field_value
= (ir_constant
*)iter
.get();
2299 int size
= type_size(field_value
->type
);
2303 field_value
->accept(this);
2306 for (i
= 0; i
< (unsigned int)size
; i
++) {
2307 emit(ir
, TGSI_OPCODE_MOV
, temp
, src
);
2313 this->result
= temp_base
;
2317 if (ir
->type
->is_array()) {
2318 st_src_reg temp_base
= get_temp(ir
->type
);
2319 st_dst_reg temp
= st_dst_reg(temp_base
);
2320 int size
= type_size(ir
->type
->fields
.array
);
2325 for (i
= 0; i
< ir
->type
->length
; i
++) {
2326 ir
->array_elements
[i
]->accept(this);
2328 for (int j
= 0; j
< size
; j
++) {
2329 emit(ir
, TGSI_OPCODE_MOV
, temp
, src
);
2335 this->result
= temp_base
;
2340 if (ir
->type
->is_matrix()) {
2341 st_src_reg mat
= get_temp(ir
->type
);
2342 st_dst_reg mat_column
= st_dst_reg(mat
);
2344 for (i
= 0; i
< ir
->type
->matrix_columns
; i
++) {
2345 assert(ir
->type
->base_type
== GLSL_TYPE_FLOAT
);
2346 values
= (gl_constant_value
*) &ir
->value
.f
[i
* ir
->type
->vector_elements
];
2348 src
= st_src_reg(file
, -1, ir
->type
->base_type
);
2349 src
.index
= add_constant(file
,
2351 ir
->type
->vector_elements
,
2354 emit(ir
, TGSI_OPCODE_MOV
, mat_column
, src
);
2363 switch (ir
->type
->base_type
) {
2364 case GLSL_TYPE_FLOAT
:
2366 for (i
= 0; i
< ir
->type
->vector_elements
; i
++) {
2367 values
[i
].f
= ir
->value
.f
[i
];
2370 case GLSL_TYPE_UINT
:
2371 gl_type
= native_integers
? GL_UNSIGNED_INT
: GL_FLOAT
;
2372 for (i
= 0; i
< ir
->type
->vector_elements
; i
++) {
2373 if (native_integers
)
2374 values
[i
].u
= ir
->value
.u
[i
];
2376 values
[i
].f
= ir
->value
.u
[i
];
2380 gl_type
= native_integers
? GL_INT
: GL_FLOAT
;
2381 for (i
= 0; i
< ir
->type
->vector_elements
; i
++) {
2382 if (native_integers
)
2383 values
[i
].i
= ir
->value
.i
[i
];
2385 values
[i
].f
= ir
->value
.i
[i
];
2388 case GLSL_TYPE_BOOL
:
2389 gl_type
= native_integers
? GL_BOOL
: GL_FLOAT
;
2390 for (i
= 0; i
< ir
->type
->vector_elements
; i
++) {
2391 if (native_integers
)
2392 values
[i
].b
= ir
->value
.b
[i
];
2394 values
[i
].f
= ir
->value
.b
[i
];
2398 assert(!"Non-float/uint/int/bool constant");
2401 this->result
= st_src_reg(file
, -1, ir
->type
);
2402 this->result
.index
= add_constant(file
,
2404 ir
->type
->vector_elements
,
2406 &this->result
.swizzle
);
2410 glsl_to_tgsi_visitor::get_function_signature(ir_function_signature
*sig
)
2412 function_entry
*entry
;
2414 foreach_iter(exec_list_iterator
, iter
, this->function_signatures
) {
2415 entry
= (function_entry
*)iter
.get();
2417 if (entry
->sig
== sig
)
2421 entry
= ralloc(mem_ctx
, function_entry
);
2423 entry
->sig_id
= this->next_signature_id
++;
2424 entry
->bgn_inst
= NULL
;
2426 /* Allocate storage for all the parameters. */
2427 foreach_iter(exec_list_iterator
, iter
, sig
->parameters
) {
2428 ir_variable
*param
= (ir_variable
*)iter
.get();
2429 variable_storage
*storage
;
2431 storage
= find_variable_storage(param
);
2434 storage
= new(mem_ctx
) variable_storage(param
, PROGRAM_TEMPORARY
,
2436 this->variables
.push_tail(storage
);
2438 this->next_temp
+= type_size(param
->type
);
2441 if (!sig
->return_type
->is_void()) {
2442 entry
->return_reg
= get_temp(sig
->return_type
);
2444 entry
->return_reg
= undef_src
;
2447 this->function_signatures
.push_tail(entry
);
2452 glsl_to_tgsi_visitor::visit(ir_call
*ir
)
2454 glsl_to_tgsi_instruction
*call_inst
;
2455 ir_function_signature
*sig
= ir
->get_callee();
2456 function_entry
*entry
= get_function_signature(sig
);
2459 /* Process in parameters. */
2460 exec_list_iterator sig_iter
= sig
->parameters
.iterator();
2461 foreach_iter(exec_list_iterator
, iter
, *ir
) {
2462 ir_rvalue
*param_rval
= (ir_rvalue
*)iter
.get();
2463 ir_variable
*param
= (ir_variable
*)sig_iter
.get();
2465 if (param
->mode
== ir_var_in
||
2466 param
->mode
== ir_var_inout
) {
2467 variable_storage
*storage
= find_variable_storage(param
);
2470 param_rval
->accept(this);
2471 st_src_reg r
= this->result
;
2474 l
.file
= storage
->file
;
2475 l
.index
= storage
->index
;
2477 l
.writemask
= WRITEMASK_XYZW
;
2478 l
.cond_mask
= COND_TR
;
2480 for (i
= 0; i
< type_size(param
->type
); i
++) {
2481 emit(ir
, TGSI_OPCODE_MOV
, l
, r
);
2489 assert(!sig_iter
.has_next());
2491 /* Emit call instruction */
2492 call_inst
= emit(ir
, TGSI_OPCODE_CAL
);
2493 call_inst
->function
= entry
;
2495 /* Process out parameters. */
2496 sig_iter
= sig
->parameters
.iterator();
2497 foreach_iter(exec_list_iterator
, iter
, *ir
) {
2498 ir_rvalue
*param_rval
= (ir_rvalue
*)iter
.get();
2499 ir_variable
*param
= (ir_variable
*)sig_iter
.get();
2501 if (param
->mode
== ir_var_out
||
2502 param
->mode
== ir_var_inout
) {
2503 variable_storage
*storage
= find_variable_storage(param
);
2507 r
.file
= storage
->file
;
2508 r
.index
= storage
->index
;
2510 r
.swizzle
= SWIZZLE_NOOP
;
2513 param_rval
->accept(this);
2514 st_dst_reg l
= st_dst_reg(this->result
);
2516 for (i
= 0; i
< type_size(param
->type
); i
++) {
2517 emit(ir
, TGSI_OPCODE_MOV
, l
, r
);
2525 assert(!sig_iter
.has_next());
2527 /* Process return value. */
2528 this->result
= entry
->return_reg
;
2532 glsl_to_tgsi_visitor::visit(ir_texture
*ir
)
2534 st_src_reg result_src
, coord
, lod_info
, projector
, dx
, dy
, offset
;
2535 st_dst_reg result_dst
, coord_dst
;
2536 glsl_to_tgsi_instruction
*inst
= NULL
;
2537 unsigned opcode
= TGSI_OPCODE_NOP
;
2539 if (ir
->coordinate
) {
2540 ir
->coordinate
->accept(this);
2542 /* Put our coords in a temp. We'll need to modify them for shadow,
2543 * projection, or LOD, so the only case we'd use it as is is if
2544 * we're doing plain old texturing. The optimization passes on
2545 * glsl_to_tgsi_visitor should handle cleaning up our mess in that case.
2547 coord
= get_temp(glsl_type::vec4_type
);
2548 coord_dst
= st_dst_reg(coord
);
2549 emit(ir
, TGSI_OPCODE_MOV
, coord_dst
, this->result
);
2552 if (ir
->projector
) {
2553 ir
->projector
->accept(this);
2554 projector
= this->result
;
2557 /* Storage for our result. Ideally for an assignment we'd be using
2558 * the actual storage for the result here, instead.
2560 result_src
= get_temp(glsl_type::vec4_type
);
2561 result_dst
= st_dst_reg(result_src
);
2565 opcode
= TGSI_OPCODE_TEX
;
2568 opcode
= TGSI_OPCODE_TXB
;
2569 ir
->lod_info
.bias
->accept(this);
2570 lod_info
= this->result
;
2573 opcode
= TGSI_OPCODE_TXL
;
2574 ir
->lod_info
.lod
->accept(this);
2575 lod_info
= this->result
;
2578 opcode
= TGSI_OPCODE_TXD
;
2579 ir
->lod_info
.grad
.dPdx
->accept(this);
2581 ir
->lod_info
.grad
.dPdy
->accept(this);
2585 opcode
= TGSI_OPCODE_TXQ
;
2586 ir
->lod_info
.lod
->accept(this);
2587 lod_info
= this->result
;
2590 opcode
= TGSI_OPCODE_TXF
;
2591 ir
->lod_info
.lod
->accept(this);
2592 lod_info
= this->result
;
2594 ir
->offset
->accept(this);
2595 offset
= this->result
;
2600 const glsl_type
*sampler_type
= ir
->sampler
->type
;
2602 if (ir
->projector
) {
2603 if (opcode
== TGSI_OPCODE_TEX
) {
2604 /* Slot the projector in as the last component of the coord. */
2605 coord_dst
.writemask
= WRITEMASK_W
;
2606 emit(ir
, TGSI_OPCODE_MOV
, coord_dst
, projector
);
2607 coord_dst
.writemask
= WRITEMASK_XYZW
;
2608 opcode
= TGSI_OPCODE_TXP
;
2610 st_src_reg coord_w
= coord
;
2611 coord_w
.swizzle
= SWIZZLE_WWWW
;
2613 /* For the other TEX opcodes there's no projective version
2614 * since the last slot is taken up by LOD info. Do the
2615 * projective divide now.
2617 coord_dst
.writemask
= WRITEMASK_W
;
2618 emit(ir
, TGSI_OPCODE_RCP
, coord_dst
, projector
);
2620 /* In the case where we have to project the coordinates "by hand,"
2621 * the shadow comparator value must also be projected.
2623 st_src_reg tmp_src
= coord
;
2624 if (ir
->shadow_comparitor
) {
2625 /* Slot the shadow value in as the second to last component of the
2628 ir
->shadow_comparitor
->accept(this);
2630 tmp_src
= get_temp(glsl_type::vec4_type
);
2631 st_dst_reg tmp_dst
= st_dst_reg(tmp_src
);
2633 /* Projective division not allowed for array samplers. */
2634 assert(!sampler_type
->sampler_array
);
2636 tmp_dst
.writemask
= WRITEMASK_Z
;
2637 emit(ir
, TGSI_OPCODE_MOV
, tmp_dst
, this->result
);
2639 tmp_dst
.writemask
= WRITEMASK_XY
;
2640 emit(ir
, TGSI_OPCODE_MOV
, tmp_dst
, coord
);
2643 coord_dst
.writemask
= WRITEMASK_XYZ
;
2644 emit(ir
, TGSI_OPCODE_MUL
, coord_dst
, tmp_src
, coord_w
);
2646 coord_dst
.writemask
= WRITEMASK_XYZW
;
2647 coord
.swizzle
= SWIZZLE_XYZW
;
2651 /* If projection is done and the opcode is not TGSI_OPCODE_TXP, then the shadow
2652 * comparator was put in the correct place (and projected) by the code,
2653 * above, that handles by-hand projection.
2655 if (ir
->shadow_comparitor
&& (!ir
->projector
|| opcode
== TGSI_OPCODE_TXP
)) {
2656 /* Slot the shadow value in as the second to last component of the
2659 ir
->shadow_comparitor
->accept(this);
2661 /* XXX This will need to be updated for cubemap array samplers. */
2662 if (sampler_type
->sampler_dimensionality
== GLSL_SAMPLER_DIM_2D
&&
2663 sampler_type
->sampler_array
) {
2664 coord_dst
.writemask
= WRITEMASK_W
;
2666 coord_dst
.writemask
= WRITEMASK_Z
;
2669 emit(ir
, TGSI_OPCODE_MOV
, coord_dst
, this->result
);
2670 coord_dst
.writemask
= WRITEMASK_XYZW
;
2673 if (opcode
== TGSI_OPCODE_TXL
|| opcode
== TGSI_OPCODE_TXB
||
2674 opcode
== TGSI_OPCODE_TXF
) {
2675 /* TGSI stores LOD or LOD bias in the last channel of the coords. */
2676 coord_dst
.writemask
= WRITEMASK_W
;
2677 emit(ir
, TGSI_OPCODE_MOV
, coord_dst
, lod_info
);
2678 coord_dst
.writemask
= WRITEMASK_XYZW
;
2681 if (opcode
== TGSI_OPCODE_TXD
)
2682 inst
= emit(ir
, opcode
, result_dst
, coord
, dx
, dy
);
2683 else if (opcode
== TGSI_OPCODE_TXQ
)
2684 inst
= emit(ir
, opcode
, result_dst
, lod_info
);
2685 else if (opcode
== TGSI_OPCODE_TXF
) {
2686 inst
= emit(ir
, opcode
, result_dst
, coord
);
2688 inst
= emit(ir
, opcode
, result_dst
, coord
);
2690 if (ir
->shadow_comparitor
)
2691 inst
->tex_shadow
= GL_TRUE
;
2693 inst
->sampler
= _mesa_get_sampler_uniform_value(ir
->sampler
,
2694 this->shader_program
,
2698 inst
->tex_offset_num_offset
= 1;
2699 inst
->tex_offsets
[0].Index
= offset
.index
;
2700 inst
->tex_offsets
[0].File
= offset
.file
;
2701 inst
->tex_offsets
[0].SwizzleX
= GET_SWZ(offset
.swizzle
, 0);
2702 inst
->tex_offsets
[0].SwizzleY
= GET_SWZ(offset
.swizzle
, 1);
2703 inst
->tex_offsets
[0].SwizzleZ
= GET_SWZ(offset
.swizzle
, 2);
2706 switch (sampler_type
->sampler_dimensionality
) {
2707 case GLSL_SAMPLER_DIM_1D
:
2708 inst
->tex_target
= (sampler_type
->sampler_array
)
2709 ? TEXTURE_1D_ARRAY_INDEX
: TEXTURE_1D_INDEX
;
2711 case GLSL_SAMPLER_DIM_2D
:
2712 inst
->tex_target
= (sampler_type
->sampler_array
)
2713 ? TEXTURE_2D_ARRAY_INDEX
: TEXTURE_2D_INDEX
;
2715 case GLSL_SAMPLER_DIM_3D
:
2716 inst
->tex_target
= TEXTURE_3D_INDEX
;
2718 case GLSL_SAMPLER_DIM_CUBE
:
2719 inst
->tex_target
= TEXTURE_CUBE_INDEX
;
2721 case GLSL_SAMPLER_DIM_RECT
:
2722 inst
->tex_target
= TEXTURE_RECT_INDEX
;
2724 case GLSL_SAMPLER_DIM_BUF
:
2725 assert(!"FINISHME: Implement ARB_texture_buffer_object");
2728 assert(!"Should not get here.");
2731 this->result
= result_src
;
2735 glsl_to_tgsi_visitor::visit(ir_return
*ir
)
2737 if (ir
->get_value()) {
2741 assert(current_function
);
2743 ir
->get_value()->accept(this);
2744 st_src_reg r
= this->result
;
2746 l
= st_dst_reg(current_function
->return_reg
);
2748 for (i
= 0; i
< type_size(current_function
->sig
->return_type
); i
++) {
2749 emit(ir
, TGSI_OPCODE_MOV
, l
, r
);
2755 emit(ir
, TGSI_OPCODE_RET
);
2759 glsl_to_tgsi_visitor::visit(ir_discard
*ir
)
2761 struct gl_fragment_program
*fp
= (struct gl_fragment_program
*)this->prog
;
2763 if (ir
->condition
) {
2764 ir
->condition
->accept(this);
2765 this->result
.negate
= ~this->result
.negate
;
2766 emit(ir
, TGSI_OPCODE_KIL
, undef_dst
, this->result
);
2768 emit(ir
, TGSI_OPCODE_KILP
);
2771 fp
->UsesKill
= GL_TRUE
;
2775 glsl_to_tgsi_visitor::visit(ir_if
*ir
)
2777 glsl_to_tgsi_instruction
*cond_inst
, *if_inst
;
2778 glsl_to_tgsi_instruction
*prev_inst
;
2780 prev_inst
= (glsl_to_tgsi_instruction
*)this->instructions
.get_tail();
2782 ir
->condition
->accept(this);
2783 assert(this->result
.file
!= PROGRAM_UNDEFINED
);
2785 if (this->options
->EmitCondCodes
) {
2786 cond_inst
= (glsl_to_tgsi_instruction
*)this->instructions
.get_tail();
2788 /* See if we actually generated any instruction for generating
2789 * the condition. If not, then cook up a move to a temp so we
2790 * have something to set cond_update on.
2792 if (cond_inst
== prev_inst
) {
2793 st_src_reg temp
= get_temp(glsl_type::bool_type
);
2794 cond_inst
= emit(ir
->condition
, TGSI_OPCODE_MOV
, st_dst_reg(temp
), result
);
2796 cond_inst
->cond_update
= GL_TRUE
;
2798 if_inst
= emit(ir
->condition
, TGSI_OPCODE_IF
);
2799 if_inst
->dst
.cond_mask
= COND_NE
;
2801 if_inst
= emit(ir
->condition
, TGSI_OPCODE_IF
, undef_dst
, this->result
);
2804 this->instructions
.push_tail(if_inst
);
2806 visit_exec_list(&ir
->then_instructions
, this);
2808 if (!ir
->else_instructions
.is_empty()) {
2809 emit(ir
->condition
, TGSI_OPCODE_ELSE
);
2810 visit_exec_list(&ir
->else_instructions
, this);
2813 if_inst
= emit(ir
->condition
, TGSI_OPCODE_ENDIF
);
2816 glsl_to_tgsi_visitor::glsl_to_tgsi_visitor()
2818 result
.file
= PROGRAM_UNDEFINED
;
2820 next_signature_id
= 1;
2822 current_function
= NULL
;
2823 num_address_regs
= 0;
2824 indirect_addr_temps
= false;
2825 indirect_addr_consts
= false;
2826 mem_ctx
= ralloc_context(NULL
);
2829 glsl_to_tgsi_visitor::~glsl_to_tgsi_visitor()
2831 ralloc_free(mem_ctx
);
2834 extern "C" void free_glsl_to_tgsi_visitor(glsl_to_tgsi_visitor
*v
)
2841 * Count resources used by the given gpu program (number of texture
2845 count_resources(glsl_to_tgsi_visitor
*v
, gl_program
*prog
)
2847 v
->samplers_used
= 0;
2849 foreach_iter(exec_list_iterator
, iter
, v
->instructions
) {
2850 glsl_to_tgsi_instruction
*inst
= (glsl_to_tgsi_instruction
*)iter
.get();
2852 if (is_tex_instruction(inst
->op
)) {
2853 v
->samplers_used
|= 1 << inst
->sampler
;
2855 prog
->SamplerTargets
[inst
->sampler
] =
2856 (gl_texture_index
)inst
->tex_target
;
2857 if (inst
->tex_shadow
) {
2858 prog
->ShadowSamplers
|= 1 << inst
->sampler
;
2863 prog
->SamplersUsed
= v
->samplers_used
;
2864 _mesa_update_shader_textures_used(prog
);
2869 * Check if the given vertex/fragment/shader program is within the
2870 * resource limits of the context (number of texture units, etc).
2871 * If any of those checks fail, record a linker error.
2873 * XXX more checks are needed...
2876 check_resources(const struct gl_context
*ctx
,
2877 struct gl_shader_program
*shader_program
,
2878 glsl_to_tgsi_visitor
*prog
,
2879 struct gl_program
*proginfo
)
2881 switch (proginfo
->Target
) {
2882 case GL_VERTEX_PROGRAM_ARB
:
2883 if (_mesa_bitcount(prog
->samplers_used
) >
2884 ctx
->Const
.MaxVertexTextureImageUnits
) {
2885 fail_link(shader_program
, "Too many vertex shader texture samplers");
2887 if (proginfo
->Parameters
->NumParameters
> MAX_UNIFORMS
) {
2888 fail_link(shader_program
, "Too many vertex shader constants");
2891 case MESA_GEOMETRY_PROGRAM
:
2892 if (_mesa_bitcount(prog
->samplers_used
) >
2893 ctx
->Const
.MaxGeometryTextureImageUnits
) {
2894 fail_link(shader_program
, "Too many geometry shader texture samplers");
2896 if (proginfo
->Parameters
->NumParameters
>
2897 MAX_GEOMETRY_UNIFORM_COMPONENTS
/ 4) {
2898 fail_link(shader_program
, "Too many geometry shader constants");
2901 case GL_FRAGMENT_PROGRAM_ARB
:
2902 if (_mesa_bitcount(prog
->samplers_used
) >
2903 ctx
->Const
.MaxTextureImageUnits
) {
2904 fail_link(shader_program
, "Too many fragment shader texture samplers");
2906 if (proginfo
->Parameters
->NumParameters
> MAX_UNIFORMS
) {
2907 fail_link(shader_program
, "Too many fragment shader constants");
2911 _mesa_problem(ctx
, "unexpected program type in check_resources()");
2917 struct uniform_sort
{
2918 struct gl_uniform
*u
;
2922 /* The shader_program->Uniforms list is almost sorted in increasing
2923 * uniform->{Frag,Vert}Pos locations, but not quite when there are
2924 * uniforms shared between targets. We need to add parameters in
2925 * increasing order for the targets.
2928 sort_uniforms(const void *a
, const void *b
)
2930 struct uniform_sort
*u1
= (struct uniform_sort
*)a
;
2931 struct uniform_sort
*u2
= (struct uniform_sort
*)b
;
2933 return u1
->pos
- u2
->pos
;
2936 /* Add the uniforms to the parameters. The linker chose locations
2937 * in our parameters lists (which weren't created yet), which the
2938 * uniforms code will use to poke values into our parameters list
2939 * when uniforms are updated.
2942 add_uniforms_to_parameters_list(struct gl_shader_program
*shader_program
,
2943 struct gl_shader
*shader
,
2944 struct gl_program
*prog
)
2947 unsigned int next_sampler
= 0, num_uniforms
= 0;
2948 struct uniform_sort
*sorted_uniforms
;
2950 sorted_uniforms
= ralloc_array(NULL
, struct uniform_sort
,
2951 shader_program
->Uniforms
->NumUniforms
);
2953 for (i
= 0; i
< shader_program
->Uniforms
->NumUniforms
; i
++) {
2954 struct gl_uniform
*uniform
= shader_program
->Uniforms
->Uniforms
+ i
;
2955 int parameter_index
= -1;
2957 switch (shader
->Type
) {
2958 case GL_VERTEX_SHADER
:
2959 parameter_index
= uniform
->VertPos
;
2961 case GL_FRAGMENT_SHADER
:
2962 parameter_index
= uniform
->FragPos
;
2964 case GL_GEOMETRY_SHADER
:
2965 parameter_index
= uniform
->GeomPos
;
2969 /* Only add uniforms used in our target. */
2970 if (parameter_index
!= -1) {
2971 sorted_uniforms
[num_uniforms
].pos
= parameter_index
;
2972 sorted_uniforms
[num_uniforms
].u
= uniform
;
2977 qsort(sorted_uniforms
, num_uniforms
, sizeof(struct uniform_sort
),
2980 for (i
= 0; i
< num_uniforms
; i
++) {
2981 struct gl_uniform
*uniform
= sorted_uniforms
[i
].u
;
2982 int parameter_index
= sorted_uniforms
[i
].pos
;
2983 const glsl_type
*type
= uniform
->Type
;
2986 if (type
->is_vector() ||
2987 type
->is_scalar()) {
2988 size
= type
->vector_elements
;
2990 size
= type_size(type
) * 4;
2993 gl_register_file file
;
2994 if (type
->is_sampler() ||
2995 (type
->is_array() && type
->fields
.array
->is_sampler())) {
2996 file
= PROGRAM_SAMPLER
;
2998 file
= PROGRAM_UNIFORM
;
3001 GLint index
= _mesa_lookup_parameter_index(prog
->Parameters
, -1,
3005 index
= _mesa_add_parameter(prog
->Parameters
, file
,
3006 uniform
->Name
, size
, type
->gl_type
,
3009 /* Sampler uniform values are stored in prog->SamplerUnits,
3010 * and the entry in that array is selected by this index we
3011 * store in ParameterValues[].
3013 if (file
== PROGRAM_SAMPLER
) {
3014 for (unsigned int j
= 0; j
< size
/ 4; j
++)
3015 prog
->Parameters
->ParameterValues
[index
+ j
][0].f
= next_sampler
++;
3018 /* The location chosen in the Parameters list here (returned
3019 * from _mesa_add_uniform) has to match what the linker chose.
3021 if (index
!= parameter_index
) {
3022 fail_link(shader_program
, "Allocation of uniform `%s' to target "
3023 "failed (%d vs %d)\n",
3024 uniform
->Name
, index
, parameter_index
);
3029 ralloc_free(sorted_uniforms
);
3033 set_uniform_initializer(struct gl_context
*ctx
, void *mem_ctx
,
3034 struct gl_shader_program
*shader_program
,
3035 const char *name
, const glsl_type
*type
,
3038 if (type
->is_record()) {
3039 ir_constant
*field_constant
;
3041 field_constant
= (ir_constant
*)val
->components
.get_head();
3043 for (unsigned int i
= 0; i
< type
->length
; i
++) {
3044 const glsl_type
*field_type
= type
->fields
.structure
[i
].type
;
3045 const char *field_name
= ralloc_asprintf(mem_ctx
, "%s.%s", name
,
3046 type
->fields
.structure
[i
].name
);
3047 set_uniform_initializer(ctx
, mem_ctx
, shader_program
, field_name
,
3048 field_type
, field_constant
);
3049 field_constant
= (ir_constant
*)field_constant
->next
;
3054 int loc
= _mesa_get_uniform_location(ctx
, shader_program
, name
);
3057 fail_link(shader_program
,
3058 "Couldn't find uniform for initializer %s\n", name
);
3062 for (unsigned int i
= 0; i
< (type
->is_array() ? type
->length
: 1); i
++) {
3063 ir_constant
*element
;
3064 const glsl_type
*element_type
;
3065 if (type
->is_array()) {
3066 element
= val
->array_elements
[i
];
3067 element_type
= type
->fields
.array
;
3070 element_type
= type
;
3075 if (element_type
->base_type
== GLSL_TYPE_BOOL
) {
3076 int *conv
= ralloc_array(mem_ctx
, int, element_type
->components());
3077 for (unsigned int j
= 0; j
< element_type
->components(); j
++) {
3078 conv
[j
] = element
->value
.b
[j
];
3080 values
= (void *)conv
;
3081 element_type
= glsl_type::get_instance(GLSL_TYPE_INT
,
3082 element_type
->vector_elements
,
3085 values
= &element
->value
;
3088 if (element_type
->is_matrix()) {
3089 _mesa_uniform_matrix(ctx
, shader_program
,
3090 element_type
->matrix_columns
,
3091 element_type
->vector_elements
,
3092 loc
, 1, GL_FALSE
, (GLfloat
*)values
);
3093 loc
+= element_type
->matrix_columns
;
3095 _mesa_uniform(ctx
, shader_program
, loc
, element_type
->matrix_columns
,
3096 values
, element_type
->gl_type
);
3097 loc
+= type_size(element_type
);
3103 * Scan/rewrite program to remove reads of custom (output) registers.
3104 * The passed type has to be either PROGRAM_OUTPUT or PROGRAM_VARYING
3105 * (for vertex shaders).
3106 * In GLSL shaders, varying vars can be read and written.
3107 * On some hardware, trying to read an output register causes trouble.
3108 * So, rewrite the program to use a temporary register in this case.
3110 * Based on _mesa_remove_output_reads from programopt.c.
3113 glsl_to_tgsi_visitor::remove_output_reads(gl_register_file type
)
3116 GLint outputMap
[VERT_RESULT_MAX
];
3117 GLint outputTypes
[VERT_RESULT_MAX
];
3118 GLuint numVaryingReads
= 0;
3119 GLboolean usedTemps
[MAX_TEMPS
];
3120 GLuint firstTemp
= 0;
3122 _mesa_find_used_registers(prog
, PROGRAM_TEMPORARY
,
3123 usedTemps
, MAX_TEMPS
);
3125 assert(type
== PROGRAM_VARYING
|| type
== PROGRAM_OUTPUT
);
3126 assert(prog
->Target
== GL_VERTEX_PROGRAM_ARB
|| type
!= PROGRAM_VARYING
);
3128 for (i
= 0; i
< VERT_RESULT_MAX
; i
++)
3131 /* look for instructions which read from varying vars */
3132 foreach_iter(exec_list_iterator
, iter
, this->instructions
) {
3133 glsl_to_tgsi_instruction
*inst
= (glsl_to_tgsi_instruction
*)iter
.get();
3134 const GLuint numSrc
= num_inst_src_regs(inst
->op
);
3136 for (j
= 0; j
< numSrc
; j
++) {
3137 if (inst
->src
[j
].file
== type
) {
3138 /* replace the read with a temp reg */
3139 const GLuint var
= inst
->src
[j
].index
;
3140 if (outputMap
[var
] == -1) {
3142 outputMap
[var
] = _mesa_find_free_register(usedTemps
,
3145 outputTypes
[var
] = inst
->src
[j
].type
;
3146 firstTemp
= outputMap
[var
] + 1;
3148 inst
->src
[j
].file
= PROGRAM_TEMPORARY
;
3149 inst
->src
[j
].index
= outputMap
[var
];
3154 if (numVaryingReads
== 0)
3155 return; /* nothing to be done */
3157 /* look for instructions which write to the varying vars identified above */
3158 foreach_iter(exec_list_iterator
, iter
, this->instructions
) {
3159 glsl_to_tgsi_instruction
*inst
= (glsl_to_tgsi_instruction
*)iter
.get();
3160 if (inst
->dst
.file
== type
&& outputMap
[inst
->dst
.index
] >= 0) {
3161 /* change inst to write to the temp reg, instead of the varying */
3162 inst
->dst
.file
= PROGRAM_TEMPORARY
;
3163 inst
->dst
.index
= outputMap
[inst
->dst
.index
];
3167 /* insert new MOV instructions at the end */
3168 for (i
= 0; i
< VERT_RESULT_MAX
; i
++) {
3169 if (outputMap
[i
] >= 0) {
3170 /* MOV VAR[i], TEMP[tmp]; */
3171 st_src_reg src
= st_src_reg(PROGRAM_TEMPORARY
, outputMap
[i
], outputTypes
[i
]);
3172 st_dst_reg dst
= st_dst_reg(type
, WRITEMASK_XYZW
, outputTypes
[i
]);
3174 this->emit(NULL
, TGSI_OPCODE_MOV
, dst
, src
);
3180 * Returns the mask of channels (bitmask of WRITEMASK_X,Y,Z,W) which
3181 * are read from the given src in this instruction
3184 get_src_arg_mask(st_dst_reg dst
, st_src_reg src
)
3186 int read_mask
= 0, comp
;
3188 /* Now, given the src swizzle and the written channels, find which
3189 * components are actually read
3191 for (comp
= 0; comp
< 4; ++comp
) {
3192 const unsigned coord
= GET_SWZ(src
.swizzle
, comp
);
3194 if (dst
.writemask
& (1 << comp
) && coord
<= SWIZZLE_W
)
3195 read_mask
|= 1 << coord
;
3202 * This pass replaces CMP T0, T1 T2 T0 with MOV T0, T2 when the CMP
3203 * instruction is the first instruction to write to register T0. There are
3204 * several lowering passes done in GLSL IR (e.g. branches and
3205 * relative addressing) that create a large number of conditional assignments
3206 * that ir_to_mesa converts to CMP instructions like the one mentioned above.
3208 * Here is why this conversion is safe:
3209 * CMP T0, T1 T2 T0 can be expanded to:
3215 * If (T1 < 0.0) evaluates to true then our replacement MOV T0, T2 is the same
3216 * as the original program. If (T1 < 0.0) evaluates to false, executing
3217 * MOV T0, T0 will store a garbage value in T0 since T0 is uninitialized.
3218 * Therefore, it doesn't matter that we are replacing MOV T0, T0 with MOV T0, T2
3219 * because any instruction that was going to read from T0 after this was going
3220 * to read a garbage value anyway.
3223 glsl_to_tgsi_visitor::simplify_cmp(void)
3225 unsigned tempWrites
[MAX_TEMPS
];
3226 unsigned outputWrites
[MAX_PROGRAM_OUTPUTS
];
3228 memset(tempWrites
, 0, sizeof(tempWrites
));
3229 memset(outputWrites
, 0, sizeof(outputWrites
));
3231 foreach_iter(exec_list_iterator
, iter
, this->instructions
) {
3232 glsl_to_tgsi_instruction
*inst
= (glsl_to_tgsi_instruction
*)iter
.get();
3233 unsigned prevWriteMask
= 0;
3235 /* Give up if we encounter relative addressing or flow control. */
3236 if (inst
->dst
.reladdr
||
3237 tgsi_get_opcode_info(inst
->op
)->is_branch
||
3238 inst
->op
== TGSI_OPCODE_BGNSUB
||
3239 inst
->op
== TGSI_OPCODE_CONT
||
3240 inst
->op
== TGSI_OPCODE_END
||
3241 inst
->op
== TGSI_OPCODE_ENDSUB
||
3242 inst
->op
== TGSI_OPCODE_RET
) {
3246 if (inst
->dst
.file
== PROGRAM_OUTPUT
) {
3247 assert(inst
->dst
.index
< MAX_PROGRAM_OUTPUTS
);
3248 prevWriteMask
= outputWrites
[inst
->dst
.index
];
3249 outputWrites
[inst
->dst
.index
] |= inst
->dst
.writemask
;
3250 } else if (inst
->dst
.file
== PROGRAM_TEMPORARY
) {
3251 assert(inst
->dst
.index
< MAX_TEMPS
);
3252 prevWriteMask
= tempWrites
[inst
->dst
.index
];
3253 tempWrites
[inst
->dst
.index
] |= inst
->dst
.writemask
;
3256 /* For a CMP to be considered a conditional write, the destination
3257 * register and source register two must be the same. */
3258 if (inst
->op
== TGSI_OPCODE_CMP
3259 && !(inst
->dst
.writemask
& prevWriteMask
)
3260 && inst
->src
[2].file
== inst
->dst
.file
3261 && inst
->src
[2].index
== inst
->dst
.index
3262 && inst
->dst
.writemask
== get_src_arg_mask(inst
->dst
, inst
->src
[2])) {
3264 inst
->op
= TGSI_OPCODE_MOV
;
3265 inst
->src
[0] = inst
->src
[1];
3270 /* Replaces all references to a temporary register index with another index. */
3272 glsl_to_tgsi_visitor::rename_temp_register(int index
, int new_index
)
3274 foreach_iter(exec_list_iterator
, iter
, this->instructions
) {
3275 glsl_to_tgsi_instruction
*inst
= (glsl_to_tgsi_instruction
*)iter
.get();
3278 for (j
=0; j
< num_inst_src_regs(inst
->op
); j
++) {
3279 if (inst
->src
[j
].file
== PROGRAM_TEMPORARY
&&
3280 inst
->src
[j
].index
== index
) {
3281 inst
->src
[j
].index
= new_index
;
3285 if (inst
->dst
.file
== PROGRAM_TEMPORARY
&& inst
->dst
.index
== index
) {
3286 inst
->dst
.index
= new_index
;
3292 glsl_to_tgsi_visitor::get_first_temp_read(int index
)
3294 int depth
= 0; /* loop depth */
3295 int loop_start
= -1; /* index of the first active BGNLOOP (if any) */
3298 foreach_iter(exec_list_iterator
, iter
, this->instructions
) {
3299 glsl_to_tgsi_instruction
*inst
= (glsl_to_tgsi_instruction
*)iter
.get();
3301 for (j
=0; j
< num_inst_src_regs(inst
->op
); j
++) {
3302 if (inst
->src
[j
].file
== PROGRAM_TEMPORARY
&&
3303 inst
->src
[j
].index
== index
) {
3304 return (depth
== 0) ? i
: loop_start
;
3308 if (inst
->op
== TGSI_OPCODE_BGNLOOP
) {
3311 } else if (inst
->op
== TGSI_OPCODE_ENDLOOP
) {
3324 glsl_to_tgsi_visitor::get_first_temp_write(int index
)
3326 int depth
= 0; /* loop depth */
3327 int loop_start
= -1; /* index of the first active BGNLOOP (if any) */
3330 foreach_iter(exec_list_iterator
, iter
, this->instructions
) {
3331 glsl_to_tgsi_instruction
*inst
= (glsl_to_tgsi_instruction
*)iter
.get();
3333 if (inst
->dst
.file
== PROGRAM_TEMPORARY
&& inst
->dst
.index
== index
) {
3334 return (depth
== 0) ? i
: loop_start
;
3337 if (inst
->op
== TGSI_OPCODE_BGNLOOP
) {
3340 } else if (inst
->op
== TGSI_OPCODE_ENDLOOP
) {
3353 glsl_to_tgsi_visitor::get_last_temp_read(int index
)
3355 int depth
= 0; /* loop depth */
3356 int last
= -1; /* index of last instruction that reads the temporary */
3359 foreach_iter(exec_list_iterator
, iter
, this->instructions
) {
3360 glsl_to_tgsi_instruction
*inst
= (glsl_to_tgsi_instruction
*)iter
.get();
3362 for (j
=0; j
< num_inst_src_regs(inst
->op
); j
++) {
3363 if (inst
->src
[j
].file
== PROGRAM_TEMPORARY
&&
3364 inst
->src
[j
].index
== index
) {
3365 last
= (depth
== 0) ? i
: -2;
3369 if (inst
->op
== TGSI_OPCODE_BGNLOOP
)
3371 else if (inst
->op
== TGSI_OPCODE_ENDLOOP
)
3372 if (--depth
== 0 && last
== -2)
3384 glsl_to_tgsi_visitor::get_last_temp_write(int index
)
3386 int depth
= 0; /* loop depth */
3387 int last
= -1; /* index of last instruction that writes to the temporary */
3390 foreach_iter(exec_list_iterator
, iter
, this->instructions
) {
3391 glsl_to_tgsi_instruction
*inst
= (glsl_to_tgsi_instruction
*)iter
.get();
3393 if (inst
->dst
.file
== PROGRAM_TEMPORARY
&& inst
->dst
.index
== index
)
3394 last
= (depth
== 0) ? i
: -2;
3396 if (inst
->op
== TGSI_OPCODE_BGNLOOP
)
3398 else if (inst
->op
== TGSI_OPCODE_ENDLOOP
)
3399 if (--depth
== 0 && last
== -2)
3411 * On a basic block basis, tracks available PROGRAM_TEMPORARY register
3412 * channels for copy propagation and updates following instructions to
3413 * use the original versions.
3415 * The glsl_to_tgsi_visitor lazily produces code assuming that this pass
3416 * will occur. As an example, a TXP production before this pass:
3418 * 0: MOV TEMP[1], INPUT[4].xyyy;
3419 * 1: MOV TEMP[1].w, INPUT[4].wwww;
3420 * 2: TXP TEMP[2], TEMP[1], texture[0], 2D;
3424 * 0: MOV TEMP[1], INPUT[4].xyyy;
3425 * 1: MOV TEMP[1].w, INPUT[4].wwww;
3426 * 2: TXP TEMP[2], INPUT[4].xyyw, texture[0], 2D;
3428 * which allows for dead code elimination on TEMP[1]'s writes.
3431 glsl_to_tgsi_visitor::copy_propagate(void)
3433 glsl_to_tgsi_instruction
**acp
= rzalloc_array(mem_ctx
,
3434 glsl_to_tgsi_instruction
*,
3435 this->next_temp
* 4);
3436 int *acp_level
= rzalloc_array(mem_ctx
, int, this->next_temp
* 4);
3439 foreach_iter(exec_list_iterator
, iter
, this->instructions
) {
3440 glsl_to_tgsi_instruction
*inst
= (glsl_to_tgsi_instruction
*)iter
.get();
3442 assert(inst
->dst
.file
!= PROGRAM_TEMPORARY
3443 || inst
->dst
.index
< this->next_temp
);
3445 /* First, do any copy propagation possible into the src regs. */
3446 for (int r
= 0; r
< 3; r
++) {
3447 glsl_to_tgsi_instruction
*first
= NULL
;
3449 int acp_base
= inst
->src
[r
].index
* 4;
3451 if (inst
->src
[r
].file
!= PROGRAM_TEMPORARY
||
3452 inst
->src
[r
].reladdr
)
3455 /* See if we can find entries in the ACP consisting of MOVs
3456 * from the same src register for all the swizzled channels
3457 * of this src register reference.
3459 for (int i
= 0; i
< 4; i
++) {
3460 int src_chan
= GET_SWZ(inst
->src
[r
].swizzle
, i
);
3461 glsl_to_tgsi_instruction
*copy_chan
= acp
[acp_base
+ src_chan
];
3468 assert(acp_level
[acp_base
+ src_chan
] <= level
);
3473 if (first
->src
[0].file
!= copy_chan
->src
[0].file
||
3474 first
->src
[0].index
!= copy_chan
->src
[0].index
) {
3482 /* We've now validated that we can copy-propagate to
3483 * replace this src register reference. Do it.
3485 inst
->src
[r
].file
= first
->src
[0].file
;
3486 inst
->src
[r
].index
= first
->src
[0].index
;
3489 for (int i
= 0; i
< 4; i
++) {
3490 int src_chan
= GET_SWZ(inst
->src
[r
].swizzle
, i
);
3491 glsl_to_tgsi_instruction
*copy_inst
= acp
[acp_base
+ src_chan
];
3492 swizzle
|= (GET_SWZ(copy_inst
->src
[0].swizzle
, src_chan
) <<
3495 inst
->src
[r
].swizzle
= swizzle
;
3500 case TGSI_OPCODE_BGNLOOP
:
3501 case TGSI_OPCODE_ENDLOOP
:
3502 /* End of a basic block, clear the ACP entirely. */
3503 memset(acp
, 0, sizeof(*acp
) * this->next_temp
* 4);
3506 case TGSI_OPCODE_IF
:
3510 case TGSI_OPCODE_ENDIF
:
3511 case TGSI_OPCODE_ELSE
:
3512 /* Clear all channels written inside the block from the ACP, but
3513 * leaving those that were not touched.
3515 for (int r
= 0; r
< this->next_temp
; r
++) {
3516 for (int c
= 0; c
< 4; c
++) {
3517 if (!acp
[4 * r
+ c
])
3520 if (acp_level
[4 * r
+ c
] >= level
)
3521 acp
[4 * r
+ c
] = NULL
;
3524 if (inst
->op
== TGSI_OPCODE_ENDIF
)
3529 /* Continuing the block, clear any written channels from
3532 if (inst
->dst
.file
== PROGRAM_TEMPORARY
&& inst
->dst
.reladdr
) {
3533 /* Any temporary might be written, so no copy propagation
3534 * across this instruction.
3536 memset(acp
, 0, sizeof(*acp
) * this->next_temp
* 4);
3537 } else if (inst
->dst
.file
== PROGRAM_OUTPUT
&&
3538 inst
->dst
.reladdr
) {
3539 /* Any output might be written, so no copy propagation
3540 * from outputs across this instruction.
3542 for (int r
= 0; r
< this->next_temp
; r
++) {
3543 for (int c
= 0; c
< 4; c
++) {
3544 if (!acp
[4 * r
+ c
])
3547 if (acp
[4 * r
+ c
]->src
[0].file
== PROGRAM_OUTPUT
)
3548 acp
[4 * r
+ c
] = NULL
;
3551 } else if (inst
->dst
.file
== PROGRAM_TEMPORARY
||
3552 inst
->dst
.file
== PROGRAM_OUTPUT
) {
3553 /* Clear where it's used as dst. */
3554 if (inst
->dst
.file
== PROGRAM_TEMPORARY
) {
3555 for (int c
= 0; c
< 4; c
++) {
3556 if (inst
->dst
.writemask
& (1 << c
)) {
3557 acp
[4 * inst
->dst
.index
+ c
] = NULL
;
3562 /* Clear where it's used as src. */
3563 for (int r
= 0; r
< this->next_temp
; r
++) {
3564 for (int c
= 0; c
< 4; c
++) {
3565 if (!acp
[4 * r
+ c
])
3568 int src_chan
= GET_SWZ(acp
[4 * r
+ c
]->src
[0].swizzle
, c
);
3570 if (acp
[4 * r
+ c
]->src
[0].file
== inst
->dst
.file
&&
3571 acp
[4 * r
+ c
]->src
[0].index
== inst
->dst
.index
&&
3572 inst
->dst
.writemask
& (1 << src_chan
))
3574 acp
[4 * r
+ c
] = NULL
;
3582 /* If this is a copy, add it to the ACP. */
3583 if (inst
->op
== TGSI_OPCODE_MOV
&&
3584 inst
->dst
.file
== PROGRAM_TEMPORARY
&&
3585 !inst
->dst
.reladdr
&&
3587 !inst
->src
[0].reladdr
&&
3588 !inst
->src
[0].negate
) {
3589 for (int i
= 0; i
< 4; i
++) {
3590 if (inst
->dst
.writemask
& (1 << i
)) {
3591 acp
[4 * inst
->dst
.index
+ i
] = inst
;
3592 acp_level
[4 * inst
->dst
.index
+ i
] = level
;
3598 ralloc_free(acp_level
);
3603 * Tracks available PROGRAM_TEMPORARY registers for dead code elimination.
3605 * The glsl_to_tgsi_visitor lazily produces code assuming that this pass
3606 * will occur. As an example, a TXP production after copy propagation but
3609 * 0: MOV TEMP[1], INPUT[4].xyyy;
3610 * 1: MOV TEMP[1].w, INPUT[4].wwww;
3611 * 2: TXP TEMP[2], INPUT[4].xyyw, texture[0], 2D;
3613 * and after this pass:
3615 * 0: TXP TEMP[2], INPUT[4].xyyw, texture[0], 2D;
3617 * FIXME: assumes that all functions are inlined (no support for BGNSUB/ENDSUB)
3618 * FIXME: doesn't eliminate all dead code inside of loops; it steps around them
3621 glsl_to_tgsi_visitor::eliminate_dead_code(void)
3625 for (i
=0; i
< this->next_temp
; i
++) {
3626 int last_read
= get_last_temp_read(i
);
3629 foreach_iter(exec_list_iterator
, iter
, this->instructions
) {
3630 glsl_to_tgsi_instruction
*inst
= (glsl_to_tgsi_instruction
*)iter
.get();
3632 if (inst
->dst
.file
== PROGRAM_TEMPORARY
&& inst
->dst
.index
== i
&&
3645 * On a basic block basis, tracks available PROGRAM_TEMPORARY registers for dead
3646 * code elimination. This is less primitive than eliminate_dead_code(), as it
3647 * is per-channel and can detect consecutive writes without a read between them
3648 * as dead code. However, there is some dead code that can be eliminated by
3649 * eliminate_dead_code() but not this function - for example, this function
3650 * cannot eliminate an instruction writing to a register that is never read and
3651 * is the only instruction writing to that register.
3653 * The glsl_to_tgsi_visitor lazily produces code assuming that this pass
3657 glsl_to_tgsi_visitor::eliminate_dead_code_advanced(void)
3659 glsl_to_tgsi_instruction
**writes
= rzalloc_array(mem_ctx
,
3660 glsl_to_tgsi_instruction
*,
3661 this->next_temp
* 4);
3662 int *write_level
= rzalloc_array(mem_ctx
, int, this->next_temp
* 4);
3666 foreach_iter(exec_list_iterator
, iter
, this->instructions
) {
3667 glsl_to_tgsi_instruction
*inst
= (glsl_to_tgsi_instruction
*)iter
.get();
3669 assert(inst
->dst
.file
!= PROGRAM_TEMPORARY
3670 || inst
->dst
.index
< this->next_temp
);
3673 case TGSI_OPCODE_BGNLOOP
:
3674 case TGSI_OPCODE_ENDLOOP
:
3675 /* End of a basic block, clear the write array entirely.
3676 * FIXME: This keeps us from killing dead code when the writes are
3677 * on either side of a loop, even when the register isn't touched
3680 memset(writes
, 0, sizeof(*writes
) * this->next_temp
* 4);
3683 case TGSI_OPCODE_ENDIF
:
3687 case TGSI_OPCODE_ELSE
:
3688 /* Clear all channels written inside the preceding if block from the
3689 * write array, but leave those that were not touched.
3691 * FIXME: This destroys opportunities to remove dead code inside of
3692 * IF blocks that are followed by an ELSE block.
3694 for (int r
= 0; r
< this->next_temp
; r
++) {
3695 for (int c
= 0; c
< 4; c
++) {
3696 if (!writes
[4 * r
+ c
])
3699 if (write_level
[4 * r
+ c
] >= level
)
3700 writes
[4 * r
+ c
] = NULL
;
3705 case TGSI_OPCODE_IF
:
3707 /* fallthrough to default case to mark the condition as read */
3710 /* Continuing the block, clear any channels from the write array that
3711 * are read by this instruction.
3713 for (unsigned i
= 0; i
< Elements(inst
->src
); i
++) {
3714 if (inst
->src
[i
].file
== PROGRAM_TEMPORARY
&& inst
->src
[i
].reladdr
){
3715 /* Any temporary might be read, so no dead code elimination
3716 * across this instruction.
3718 memset(writes
, 0, sizeof(*writes
) * this->next_temp
* 4);
3719 } else if (inst
->src
[i
].file
== PROGRAM_TEMPORARY
) {
3720 /* Clear where it's used as src. */
3721 int src_chans
= 1 << GET_SWZ(inst
->src
[i
].swizzle
, 0);
3722 src_chans
|= 1 << GET_SWZ(inst
->src
[i
].swizzle
, 1);
3723 src_chans
|= 1 << GET_SWZ(inst
->src
[i
].swizzle
, 2);
3724 src_chans
|= 1 << GET_SWZ(inst
->src
[i
].swizzle
, 3);
3726 for (int c
= 0; c
< 4; c
++) {
3727 if (src_chans
& (1 << c
)) {
3728 writes
[4 * inst
->src
[i
].index
+ c
] = NULL
;
3736 /* If this instruction writes to a temporary, add it to the write array.
3737 * If there is already an instruction in the write array for one or more
3738 * of the channels, flag that channel write as dead.
3740 if (inst
->dst
.file
== PROGRAM_TEMPORARY
&&
3741 !inst
->dst
.reladdr
&&
3743 for (int c
= 0; c
< 4; c
++) {
3744 if (inst
->dst
.writemask
& (1 << c
)) {
3745 if (writes
[4 * inst
->dst
.index
+ c
]) {
3746 if (write_level
[4 * inst
->dst
.index
+ c
] < level
)
3749 writes
[4 * inst
->dst
.index
+ c
]->dead_mask
|= (1 << c
);
3751 writes
[4 * inst
->dst
.index
+ c
] = inst
;
3752 write_level
[4 * inst
->dst
.index
+ c
] = level
;
3758 /* Anything still in the write array at this point is dead code. */
3759 for (int r
= 0; r
< this->next_temp
; r
++) {
3760 for (int c
= 0; c
< 4; c
++) {
3761 glsl_to_tgsi_instruction
*inst
= writes
[4 * r
+ c
];
3763 inst
->dead_mask
|= (1 << c
);
3767 /* Now actually remove the instructions that are completely dead and update
3768 * the writemask of other instructions with dead channels.
3770 foreach_iter(exec_list_iterator
, iter
, this->instructions
) {
3771 glsl_to_tgsi_instruction
*inst
= (glsl_to_tgsi_instruction
*)iter
.get();
3773 if (!inst
->dead_mask
|| !inst
->dst
.writemask
)
3775 else if (inst
->dead_mask
== inst
->dst
.writemask
) {
3780 inst
->dst
.writemask
&= ~(inst
->dead_mask
);
3783 ralloc_free(write_level
);
3784 ralloc_free(writes
);
3789 /* Merges temporary registers together where possible to reduce the number of
3790 * registers needed to run a program.
3792 * Produces optimal code only after copy propagation and dead code elimination
3795 glsl_to_tgsi_visitor::merge_registers(void)
3797 int *last_reads
= rzalloc_array(mem_ctx
, int, this->next_temp
);
3798 int *first_writes
= rzalloc_array(mem_ctx
, int, this->next_temp
);
3801 /* Read the indices of the last read and first write to each temp register
3802 * into an array so that we don't have to traverse the instruction list as
3804 for (i
=0; i
< this->next_temp
; i
++) {
3805 last_reads
[i
] = get_last_temp_read(i
);
3806 first_writes
[i
] = get_first_temp_write(i
);
3809 /* Start looking for registers with non-overlapping usages that can be
3810 * merged together. */
3811 for (i
=0; i
< this->next_temp
; i
++) {
3812 /* Don't touch unused registers. */
3813 if (last_reads
[i
] < 0 || first_writes
[i
] < 0) continue;
3815 for (j
=0; j
< this->next_temp
; j
++) {
3816 /* Don't touch unused registers. */
3817 if (last_reads
[j
] < 0 || first_writes
[j
] < 0) continue;
3819 /* We can merge the two registers if the first write to j is after or
3820 * in the same instruction as the last read from i. Note that the
3821 * register at index i will always be used earlier or at the same time
3822 * as the register at index j. */
3823 if (first_writes
[i
] <= first_writes
[j
] &&
3824 last_reads
[i
] <= first_writes
[j
])
3826 rename_temp_register(j
, i
); /* Replace all references to j with i.*/
3828 /* Update the first_writes and last_reads arrays with the new
3829 * values for the merged register index, and mark the newly unused
3830 * register index as such. */
3831 last_reads
[i
] = last_reads
[j
];
3832 first_writes
[j
] = -1;
3838 ralloc_free(last_reads
);
3839 ralloc_free(first_writes
);
3842 /* Reassign indices to temporary registers by reusing unused indices created
3843 * by optimization passes. */
3845 glsl_to_tgsi_visitor::renumber_registers(void)
3850 for (i
=0; i
< this->next_temp
; i
++) {
3851 if (get_first_temp_read(i
) < 0) continue;
3853 rename_temp_register(i
, new_index
);
3857 this->next_temp
= new_index
;
3861 * Returns a fragment program which implements the current pixel transfer ops.
3862 * Based on get_pixel_transfer_program in st_atom_pixeltransfer.c.
3865 get_pixel_transfer_visitor(struct st_fragment_program
*fp
,
3866 glsl_to_tgsi_visitor
*original
,
3867 int scale_and_bias
, int pixel_maps
)
3869 glsl_to_tgsi_visitor
*v
= new glsl_to_tgsi_visitor();
3870 struct st_context
*st
= st_context(original
->ctx
);
3871 struct gl_program
*prog
= &fp
->Base
.Base
;
3872 struct gl_program_parameter_list
*params
= _mesa_new_parameter_list();
3873 st_src_reg coord
, src0
;
3875 glsl_to_tgsi_instruction
*inst
;
3877 /* Copy attributes of the glsl_to_tgsi_visitor in the original shader. */
3878 v
->ctx
= original
->ctx
;
3880 v
->glsl_version
= original
->glsl_version
;
3881 v
->native_integers
= original
->native_integers
;
3882 v
->options
= original
->options
;
3883 v
->next_temp
= original
->next_temp
;
3884 v
->num_address_regs
= original
->num_address_regs
;
3885 v
->samplers_used
= prog
->SamplersUsed
= original
->samplers_used
;
3886 v
->indirect_addr_temps
= original
->indirect_addr_temps
;
3887 v
->indirect_addr_consts
= original
->indirect_addr_consts
;
3888 memcpy(&v
->immediates
, &original
->immediates
, sizeof(v
->immediates
));
3891 * Get initial pixel color from the texture.
3892 * TEX colorTemp, fragment.texcoord[0], texture[0], 2D;
3894 coord
= st_src_reg(PROGRAM_INPUT
, FRAG_ATTRIB_TEX0
, glsl_type::vec2_type
);
3895 src0
= v
->get_temp(glsl_type::vec4_type
);
3896 dst0
= st_dst_reg(src0
);
3897 inst
= v
->emit(NULL
, TGSI_OPCODE_TEX
, dst0
, coord
);
3899 inst
->tex_target
= TEXTURE_2D_INDEX
;
3901 prog
->InputsRead
|= (1 << FRAG_ATTRIB_TEX0
);
3902 prog
->SamplersUsed
|= (1 << 0); /* mark sampler 0 as used */
3903 v
->samplers_used
|= (1 << 0);
3905 if (scale_and_bias
) {
3906 static const gl_state_index scale_state
[STATE_LENGTH
] =
3907 { STATE_INTERNAL
, STATE_PT_SCALE
,
3908 (gl_state_index
) 0, (gl_state_index
) 0, (gl_state_index
) 0 };
3909 static const gl_state_index bias_state
[STATE_LENGTH
] =
3910 { STATE_INTERNAL
, STATE_PT_BIAS
,
3911 (gl_state_index
) 0, (gl_state_index
) 0, (gl_state_index
) 0 };
3912 GLint scale_p
, bias_p
;
3913 st_src_reg scale
, bias
;
3915 scale_p
= _mesa_add_state_reference(params
, scale_state
);
3916 bias_p
= _mesa_add_state_reference(params
, bias_state
);
3918 /* MAD colorTemp, colorTemp, scale, bias; */
3919 scale
= st_src_reg(PROGRAM_STATE_VAR
, scale_p
, GLSL_TYPE_FLOAT
);
3920 bias
= st_src_reg(PROGRAM_STATE_VAR
, bias_p
, GLSL_TYPE_FLOAT
);
3921 inst
= v
->emit(NULL
, TGSI_OPCODE_MAD
, dst0
, src0
, scale
, bias
);
3925 st_src_reg temp
= v
->get_temp(glsl_type::vec4_type
);
3926 st_dst_reg temp_dst
= st_dst_reg(temp
);
3928 assert(st
->pixel_xfer
.pixelmap_texture
);
3930 /* With a little effort, we can do four pixel map look-ups with
3931 * two TEX instructions:
3934 /* TEX temp.rg, colorTemp.rgba, texture[1], 2D; */
3935 temp_dst
.writemask
= WRITEMASK_XY
; /* write R,G */
3936 inst
= v
->emit(NULL
, TGSI_OPCODE_TEX
, temp_dst
, src0
);
3938 inst
->tex_target
= TEXTURE_2D_INDEX
;
3940 /* TEX temp.ba, colorTemp.baba, texture[1], 2D; */
3941 src0
.swizzle
= MAKE_SWIZZLE4(SWIZZLE_Z
, SWIZZLE_W
, SWIZZLE_Z
, SWIZZLE_W
);
3942 temp_dst
.writemask
= WRITEMASK_ZW
; /* write B,A */
3943 inst
= v
->emit(NULL
, TGSI_OPCODE_TEX
, temp_dst
, src0
);
3945 inst
->tex_target
= TEXTURE_2D_INDEX
;
3947 prog
->SamplersUsed
|= (1 << 1); /* mark sampler 1 as used */
3948 v
->samplers_used
|= (1 << 1);
3950 /* MOV colorTemp, temp; */
3951 inst
= v
->emit(NULL
, TGSI_OPCODE_MOV
, dst0
, temp
);
3954 /* Now copy the instructions from the original glsl_to_tgsi_visitor into the
3956 foreach_iter(exec_list_iterator
, iter
, original
->instructions
) {
3957 glsl_to_tgsi_instruction
*inst
= (glsl_to_tgsi_instruction
*)iter
.get();
3958 st_src_reg src_regs
[3];
3960 if (inst
->dst
.file
== PROGRAM_OUTPUT
)
3961 prog
->OutputsWritten
|= BITFIELD64_BIT(inst
->dst
.index
);
3963 for (int i
=0; i
<3; i
++) {
3964 src_regs
[i
] = inst
->src
[i
];
3965 if (src_regs
[i
].file
== PROGRAM_INPUT
&&
3966 src_regs
[i
].index
== FRAG_ATTRIB_COL0
)
3968 src_regs
[i
].file
= PROGRAM_TEMPORARY
;
3969 src_regs
[i
].index
= src0
.index
;
3971 else if (src_regs
[i
].file
== PROGRAM_INPUT
)
3972 prog
->InputsRead
|= (1 << src_regs
[i
].index
);
3975 v
->emit(NULL
, inst
->op
, inst
->dst
, src_regs
[0], src_regs
[1], src_regs
[2]);
3978 /* Make modifications to fragment program info. */
3979 prog
->Parameters
= _mesa_combine_parameter_lists(params
,
3980 original
->prog
->Parameters
);
3981 _mesa_free_parameter_list(params
);
3982 count_resources(v
, prog
);
3983 fp
->glsl_to_tgsi
= v
;
3987 * Make fragment program for glBitmap:
3988 * Sample the texture and kill the fragment if the bit is 0.
3989 * This program will be combined with the user's fragment program.
3991 * Based on make_bitmap_fragment_program in st_cb_bitmap.c.
3994 get_bitmap_visitor(struct st_fragment_program
*fp
,
3995 glsl_to_tgsi_visitor
*original
, int samplerIndex
)
3997 glsl_to_tgsi_visitor
*v
= new glsl_to_tgsi_visitor();
3998 struct st_context
*st
= st_context(original
->ctx
);
3999 struct gl_program
*prog
= &fp
->Base
.Base
;
4000 st_src_reg coord
, src0
;
4002 glsl_to_tgsi_instruction
*inst
;
4004 /* Copy attributes of the glsl_to_tgsi_visitor in the original shader. */
4005 v
->ctx
= original
->ctx
;
4007 v
->glsl_version
= original
->glsl_version
;
4008 v
->native_integers
= original
->native_integers
;
4009 v
->options
= original
->options
;
4010 v
->next_temp
= original
->next_temp
;
4011 v
->num_address_regs
= original
->num_address_regs
;
4012 v
->samplers_used
= prog
->SamplersUsed
= original
->samplers_used
;
4013 v
->indirect_addr_temps
= original
->indirect_addr_temps
;
4014 v
->indirect_addr_consts
= original
->indirect_addr_consts
;
4015 memcpy(&v
->immediates
, &original
->immediates
, sizeof(v
->immediates
));
4017 /* TEX tmp0, fragment.texcoord[0], texture[0], 2D; */
4018 coord
= st_src_reg(PROGRAM_INPUT
, FRAG_ATTRIB_TEX0
, glsl_type::vec2_type
);
4019 src0
= v
->get_temp(glsl_type::vec4_type
);
4020 dst0
= st_dst_reg(src0
);
4021 inst
= v
->emit(NULL
, TGSI_OPCODE_TEX
, dst0
, coord
);
4022 inst
->sampler
= samplerIndex
;
4023 inst
->tex_target
= TEXTURE_2D_INDEX
;
4025 prog
->InputsRead
|= (1 << FRAG_ATTRIB_TEX0
);
4026 prog
->SamplersUsed
|= (1 << samplerIndex
); /* mark sampler as used */
4027 v
->samplers_used
|= (1 << samplerIndex
);
4029 /* KIL if -tmp0 < 0 # texel=0 -> keep / texel=0 -> discard */
4030 src0
.negate
= NEGATE_XYZW
;
4031 if (st
->bitmap
.tex_format
== PIPE_FORMAT_L8_UNORM
)
4032 src0
.swizzle
= SWIZZLE_XXXX
;
4033 inst
= v
->emit(NULL
, TGSI_OPCODE_KIL
, undef_dst
, src0
);
4035 /* Now copy the instructions from the original glsl_to_tgsi_visitor into the
4037 foreach_iter(exec_list_iterator
, iter
, original
->instructions
) {
4038 glsl_to_tgsi_instruction
*inst
= (glsl_to_tgsi_instruction
*)iter
.get();
4039 st_src_reg src_regs
[3];
4041 if (inst
->dst
.file
== PROGRAM_OUTPUT
)
4042 prog
->OutputsWritten
|= BITFIELD64_BIT(inst
->dst
.index
);
4044 for (int i
=0; i
<3; i
++) {
4045 src_regs
[i
] = inst
->src
[i
];
4046 if (src_regs
[i
].file
== PROGRAM_INPUT
)
4047 prog
->InputsRead
|= (1 << src_regs
[i
].index
);
4050 v
->emit(NULL
, inst
->op
, inst
->dst
, src_regs
[0], src_regs
[1], src_regs
[2]);
4053 /* Make modifications to fragment program info. */
4054 prog
->Parameters
= _mesa_clone_parameter_list(original
->prog
->Parameters
);
4055 count_resources(v
, prog
);
4056 fp
->glsl_to_tgsi
= v
;
4059 /* ------------------------- TGSI conversion stuff -------------------------- */
4061 unsigned branch_target
;
4066 * Intermediate state used during shader translation.
4068 struct st_translate
{
4069 struct ureg_program
*ureg
;
4071 struct ureg_dst temps
[MAX_TEMPS
];
4072 struct ureg_src
*constants
;
4073 struct ureg_src
*immediates
;
4074 struct ureg_dst outputs
[PIPE_MAX_SHADER_OUTPUTS
];
4075 struct ureg_src inputs
[PIPE_MAX_SHADER_INPUTS
];
4076 struct ureg_dst address
[1];
4077 struct ureg_src samplers
[PIPE_MAX_SAMPLERS
];
4078 struct ureg_src systemValues
[SYSTEM_VALUE_MAX
];
4080 /* Extra info for handling point size clamping in vertex shader */
4081 struct ureg_dst pointSizeResult
; /**< Actual point size output register */
4082 struct ureg_src pointSizeConst
; /**< Point size range constant register */
4083 GLint pointSizeOutIndex
; /**< Temp point size output register */
4084 GLboolean prevInstWrotePointSize
;
4086 const GLuint
*inputMapping
;
4087 const GLuint
*outputMapping
;
4089 /* For every instruction that contains a label (eg CALL), keep
4090 * details so that we can go back afterwards and emit the correct
4091 * tgsi instruction number for each label.
4093 struct label
*labels
;
4094 unsigned labels_size
;
4095 unsigned labels_count
;
4097 /* Keep a record of the tgsi instruction number that each mesa
4098 * instruction starts at, will be used to fix up labels after
4103 unsigned insn_count
;
4105 unsigned procType
; /**< TGSI_PROCESSOR_VERTEX/FRAGMENT */
4110 /** Map Mesa's SYSTEM_VALUE_x to TGSI_SEMANTIC_x */
4111 static unsigned mesa_sysval_to_semantic
[SYSTEM_VALUE_MAX
] = {
4113 TGSI_SEMANTIC_INSTANCEID
4117 * Make note of a branch to a label in the TGSI code.
4118 * After we've emitted all instructions, we'll go over the list
4119 * of labels built here and patch the TGSI code with the actual
4120 * location of each label.
4122 static unsigned *get_label(struct st_translate
*t
, unsigned branch_target
)
4126 if (t
->labels_count
+ 1 >= t
->labels_size
) {
4127 t
->labels_size
= 1 << (util_logbase2(t
->labels_size
) + 1);
4128 t
->labels
= (struct label
*)realloc(t
->labels
,
4129 t
->labels_size
* sizeof(struct label
));
4130 if (t
->labels
== NULL
) {
4131 static unsigned dummy
;
4137 i
= t
->labels_count
++;
4138 t
->labels
[i
].branch_target
= branch_target
;
4139 return &t
->labels
[i
].token
;
4143 * Called prior to emitting the TGSI code for each instruction.
4144 * Allocate additional space for instructions if needed.
4145 * Update the insn[] array so the next glsl_to_tgsi_instruction points to
4146 * the next TGSI instruction.
4148 static void set_insn_start(struct st_translate
*t
, unsigned start
)
4150 if (t
->insn_count
+ 1 >= t
->insn_size
) {
4151 t
->insn_size
= 1 << (util_logbase2(t
->insn_size
) + 1);
4152 t
->insn
= (unsigned *)realloc(t
->insn
, t
->insn_size
* sizeof(t
->insn
[0]));
4153 if (t
->insn
== NULL
) {
4159 t
->insn
[t
->insn_count
++] = start
;
4163 * Map a glsl_to_tgsi constant/immediate to a TGSI immediate.
4165 static struct ureg_src
4166 emit_immediate(struct st_translate
*t
,
4167 gl_constant_value values
[4],
4170 struct ureg_program
*ureg
= t
->ureg
;
4175 return ureg_DECL_immediate(ureg
, &values
[0].f
, size
);
4177 return ureg_DECL_immediate_int(ureg
, &values
[0].i
, size
);
4178 case GL_UNSIGNED_INT
:
4180 return ureg_DECL_immediate_uint(ureg
, &values
[0].u
, size
);
4182 assert(!"should not get here - type must be float, int, uint, or bool");
4183 return ureg_src_undef();
4188 * Map a glsl_to_tgsi dst register to a TGSI ureg_dst register.
4190 static struct ureg_dst
4191 dst_register(struct st_translate
*t
,
4192 gl_register_file file
,
4196 case PROGRAM_UNDEFINED
:
4197 return ureg_dst_undef();
4199 case PROGRAM_TEMPORARY
:
4200 if (ureg_dst_is_undef(t
->temps
[index
]))
4201 t
->temps
[index
] = ureg_DECL_temporary(t
->ureg
);
4203 return t
->temps
[index
];
4205 case PROGRAM_OUTPUT
:
4206 if (t
->procType
== TGSI_PROCESSOR_VERTEX
&& index
== VERT_RESULT_PSIZ
)
4207 t
->prevInstWrotePointSize
= GL_TRUE
;
4209 if (t
->procType
== TGSI_PROCESSOR_VERTEX
)
4210 assert(index
< VERT_RESULT_MAX
);
4211 else if (t
->procType
== TGSI_PROCESSOR_FRAGMENT
)
4212 assert(index
< FRAG_RESULT_MAX
);
4214 assert(index
< GEOM_RESULT_MAX
);
4216 assert(t
->outputMapping
[index
] < Elements(t
->outputs
));
4218 return t
->outputs
[t
->outputMapping
[index
]];
4220 case PROGRAM_ADDRESS
:
4221 return t
->address
[index
];
4224 assert(!"unknown dst register file");
4225 return ureg_dst_undef();
4230 * Map a glsl_to_tgsi src register to a TGSI ureg_src register.
4232 static struct ureg_src
4233 src_register(struct st_translate
*t
,
4234 gl_register_file file
,
4238 case PROGRAM_UNDEFINED
:
4239 return ureg_src_undef();
4241 case PROGRAM_TEMPORARY
:
4243 assert(index
< Elements(t
->temps
));
4244 if (ureg_dst_is_undef(t
->temps
[index
]))
4245 t
->temps
[index
] = ureg_DECL_temporary(t
->ureg
);
4246 return ureg_src(t
->temps
[index
]);
4248 case PROGRAM_NAMED_PARAM
:
4249 case PROGRAM_ENV_PARAM
:
4250 case PROGRAM_LOCAL_PARAM
:
4251 case PROGRAM_UNIFORM
:
4253 return t
->constants
[index
];
4254 case PROGRAM_STATE_VAR
:
4255 case PROGRAM_CONSTANT
: /* ie, immediate */
4257 return ureg_DECL_constant(t
->ureg
, 0);
4259 return t
->constants
[index
];
4261 case PROGRAM_IMMEDIATE
:
4262 return t
->immediates
[index
];
4265 assert(t
->inputMapping
[index
] < Elements(t
->inputs
));
4266 return t
->inputs
[t
->inputMapping
[index
]];
4268 case PROGRAM_OUTPUT
:
4269 assert(t
->outputMapping
[index
] < Elements(t
->outputs
));
4270 return ureg_src(t
->outputs
[t
->outputMapping
[index
]]); /* not needed? */
4272 case PROGRAM_ADDRESS
:
4273 return ureg_src(t
->address
[index
]);
4275 case PROGRAM_SYSTEM_VALUE
:
4276 assert(index
< Elements(t
->systemValues
));
4277 return t
->systemValues
[index
];
4280 assert(!"unknown src register file");
4281 return ureg_src_undef();
4286 * Create a TGSI ureg_dst register from an st_dst_reg.
4288 static struct ureg_dst
4289 translate_dst(struct st_translate
*t
,
4290 const st_dst_reg
*dst_reg
,
4293 struct ureg_dst dst
= dst_register(t
,
4297 dst
= ureg_writemask(dst
, dst_reg
->writemask
);
4300 dst
= ureg_saturate(dst
);
4302 if (dst_reg
->reladdr
!= NULL
)
4303 dst
= ureg_dst_indirect(dst
, ureg_src(t
->address
[0]));
4309 * Create a TGSI ureg_src register from an st_src_reg.
4311 static struct ureg_src
4312 translate_src(struct st_translate
*t
, const st_src_reg
*src_reg
)
4314 struct ureg_src src
= src_register(t
, src_reg
->file
, src_reg
->index
);
4316 src
= ureg_swizzle(src
,
4317 GET_SWZ(src_reg
->swizzle
, 0) & 0x3,
4318 GET_SWZ(src_reg
->swizzle
, 1) & 0x3,
4319 GET_SWZ(src_reg
->swizzle
, 2) & 0x3,
4320 GET_SWZ(src_reg
->swizzle
, 3) & 0x3);
4322 if ((src_reg
->negate
& 0xf) == NEGATE_XYZW
)
4323 src
= ureg_negate(src
);
4325 if (src_reg
->reladdr
!= NULL
) {
4326 /* Normally ureg_src_indirect() would be used here, but a stupid compiler
4327 * bug in g++ makes ureg_src_indirect (an inline C function) erroneously
4328 * set the bit for src.Negate. So we have to do the operation manually
4329 * here to work around the compiler's problems. */
4330 /*src = ureg_src_indirect(src, ureg_src(t->address[0]));*/
4331 struct ureg_src addr
= ureg_src(t
->address
[0]);
4333 src
.IndirectFile
= addr
.File
;
4334 src
.IndirectIndex
= addr
.Index
;
4335 src
.IndirectSwizzle
= addr
.SwizzleX
;
4337 if (src_reg
->file
!= PROGRAM_INPUT
&&
4338 src_reg
->file
!= PROGRAM_OUTPUT
) {
4339 /* If src_reg->index was negative, it was set to zero in
4340 * src_register(). Reassign it now. But don't do this
4341 * for input/output regs since they get remapped while
4342 * const buffers don't.
4344 src
.Index
= src_reg
->index
;
4351 static struct tgsi_texture_offset
4352 translate_tex_offset(struct st_translate
*t
,
4353 const struct tgsi_texture_offset
*in_offset
)
4355 struct tgsi_texture_offset offset
;
4357 assert(in_offset
->File
== PROGRAM_IMMEDIATE
);
4359 offset
.File
= TGSI_FILE_IMMEDIATE
;
4360 offset
.Index
= in_offset
->Index
;
4361 offset
.SwizzleX
= in_offset
->SwizzleX
;
4362 offset
.SwizzleY
= in_offset
->SwizzleY
;
4363 offset
.SwizzleZ
= in_offset
->SwizzleZ
;
4369 compile_tgsi_instruction(struct st_translate
*t
,
4370 const glsl_to_tgsi_instruction
*inst
)
4372 struct ureg_program
*ureg
= t
->ureg
;
4374 struct ureg_dst dst
[1];
4375 struct ureg_src src
[4];
4376 struct tgsi_texture_offset texoffsets
[MAX_GLSL_TEXTURE_OFFSET
];
4381 num_dst
= num_inst_dst_regs(inst
->op
);
4382 num_src
= num_inst_src_regs(inst
->op
);
4385 dst
[0] = translate_dst(t
,
4389 for (i
= 0; i
< num_src
; i
++)
4390 src
[i
] = translate_src(t
, &inst
->src
[i
]);
4393 case TGSI_OPCODE_BGNLOOP
:
4394 case TGSI_OPCODE_CAL
:
4395 case TGSI_OPCODE_ELSE
:
4396 case TGSI_OPCODE_ENDLOOP
:
4397 case TGSI_OPCODE_IF
:
4398 assert(num_dst
== 0);
4399 ureg_label_insn(ureg
,
4403 inst
->op
== TGSI_OPCODE_CAL
? inst
->function
->sig_id
: 0));
4406 case TGSI_OPCODE_TEX
:
4407 case TGSI_OPCODE_TXB
:
4408 case TGSI_OPCODE_TXD
:
4409 case TGSI_OPCODE_TXL
:
4410 case TGSI_OPCODE_TXP
:
4411 case TGSI_OPCODE_TXQ
:
4412 case TGSI_OPCODE_TXF
:
4413 src
[num_src
++] = t
->samplers
[inst
->sampler
];
4414 for (i
= 0; i
< inst
->tex_offset_num_offset
; i
++) {
4415 texoffsets
[i
] = translate_tex_offset(t
, &inst
->tex_offsets
[i
]);
4420 translate_texture_target(inst
->tex_target
, inst
->tex_shadow
),
4421 texoffsets
, inst
->tex_offset_num_offset
,
4425 case TGSI_OPCODE_SCS
:
4426 dst
[0] = ureg_writemask(dst
[0], TGSI_WRITEMASK_XY
);
4427 ureg_insn(ureg
, inst
->op
, dst
, num_dst
, src
, num_src
);
4440 * Emit the TGSI instructions to adjust the WPOS pixel center convention
4441 * Basically, add (adjX, adjY) to the fragment position.
4444 emit_adjusted_wpos(struct st_translate
*t
,
4445 const struct gl_program
*program
,
4446 float adjX
, float adjY
)
4448 struct ureg_program
*ureg
= t
->ureg
;
4449 struct ureg_dst wpos_temp
= ureg_DECL_temporary(ureg
);
4450 struct ureg_src wpos_input
= t
->inputs
[t
->inputMapping
[FRAG_ATTRIB_WPOS
]];
4452 /* Note that we bias X and Y and pass Z and W through unchanged.
4453 * The shader might also use gl_FragCoord.w and .z.
4455 ureg_ADD(ureg
, wpos_temp
, wpos_input
,
4456 ureg_imm4f(ureg
, adjX
, adjY
, 0.0f
, 0.0f
));
4458 t
->inputs
[t
->inputMapping
[FRAG_ATTRIB_WPOS
]] = ureg_src(wpos_temp
);
4463 * Emit the TGSI instructions for inverting the WPOS y coordinate.
4464 * This code is unavoidable because it also depends on whether
4465 * a FBO is bound (STATE_FB_WPOS_Y_TRANSFORM).
4468 emit_wpos_inversion(struct st_translate
*t
,
4469 const struct gl_program
*program
,
4472 struct ureg_program
*ureg
= t
->ureg
;
4474 /* Fragment program uses fragment position input.
4475 * Need to replace instances of INPUT[WPOS] with temp T
4476 * where T = INPUT[WPOS] by y is inverted.
4478 static const gl_state_index wposTransformState
[STATE_LENGTH
]
4479 = { STATE_INTERNAL
, STATE_FB_WPOS_Y_TRANSFORM
,
4480 (gl_state_index
)0, (gl_state_index
)0, (gl_state_index
)0 };
4482 /* XXX: note we are modifying the incoming shader here! Need to
4483 * do this before emitting the constant decls below, or this
4486 unsigned wposTransConst
= _mesa_add_state_reference(program
->Parameters
,
4487 wposTransformState
);
4489 struct ureg_src wpostrans
= ureg_DECL_constant(ureg
, wposTransConst
);
4490 struct ureg_dst wpos_temp
;
4491 struct ureg_src wpos_input
= t
->inputs
[t
->inputMapping
[FRAG_ATTRIB_WPOS
]];
4493 /* MOV wpos_temp, input[wpos]
4495 if (wpos_input
.File
== TGSI_FILE_TEMPORARY
)
4496 wpos_temp
= ureg_dst(wpos_input
);
4498 wpos_temp
= ureg_DECL_temporary(ureg
);
4499 ureg_MOV(ureg
, wpos_temp
, wpos_input
);
4503 /* MAD wpos_temp.y, wpos_input, wpostrans.xxxx, wpostrans.yyyy
4506 ureg_writemask(wpos_temp
, TGSI_WRITEMASK_Y
),
4508 ureg_scalar(wpostrans
, 0),
4509 ureg_scalar(wpostrans
, 1));
4511 /* MAD wpos_temp.y, wpos_input, wpostrans.zzzz, wpostrans.wwww
4514 ureg_writemask(wpos_temp
, TGSI_WRITEMASK_Y
),
4516 ureg_scalar(wpostrans
, 2),
4517 ureg_scalar(wpostrans
, 3));
4520 /* Use wpos_temp as position input from here on:
4522 t
->inputs
[t
->inputMapping
[FRAG_ATTRIB_WPOS
]] = ureg_src(wpos_temp
);
4527 * Emit fragment position/ooordinate code.
4530 emit_wpos(struct st_context
*st
,
4531 struct st_translate
*t
,
4532 const struct gl_program
*program
,
4533 struct ureg_program
*ureg
)
4535 const struct gl_fragment_program
*fp
=
4536 (const struct gl_fragment_program
*) program
;
4537 struct pipe_screen
*pscreen
= st
->pipe
->screen
;
4538 boolean invert
= FALSE
;
4540 if (fp
->OriginUpperLeft
) {
4541 /* Fragment shader wants origin in upper-left */
4542 if (pscreen
->get_param(pscreen
, PIPE_CAP_TGSI_FS_COORD_ORIGIN_UPPER_LEFT
)) {
4543 /* the driver supports upper-left origin */
4545 else if (pscreen
->get_param(pscreen
, PIPE_CAP_TGSI_FS_COORD_ORIGIN_LOWER_LEFT
)) {
4546 /* the driver supports lower-left origin, need to invert Y */
4547 ureg_property_fs_coord_origin(ureg
, TGSI_FS_COORD_ORIGIN_LOWER_LEFT
);
4554 /* Fragment shader wants origin in lower-left */
4555 if (pscreen
->get_param(pscreen
, PIPE_CAP_TGSI_FS_COORD_ORIGIN_LOWER_LEFT
))
4556 /* the driver supports lower-left origin */
4557 ureg_property_fs_coord_origin(ureg
, TGSI_FS_COORD_ORIGIN_LOWER_LEFT
);
4558 else if (pscreen
->get_param(pscreen
, PIPE_CAP_TGSI_FS_COORD_ORIGIN_UPPER_LEFT
))
4559 /* the driver supports upper-left origin, need to invert Y */
4565 if (fp
->PixelCenterInteger
) {
4566 /* Fragment shader wants pixel center integer */
4567 if (pscreen
->get_param(pscreen
, PIPE_CAP_TGSI_FS_COORD_PIXEL_CENTER_INTEGER
))
4568 /* the driver supports pixel center integer */
4569 ureg_property_fs_coord_pixel_center(ureg
, TGSI_FS_COORD_PIXEL_CENTER_INTEGER
);
4570 else if (pscreen
->get_param(pscreen
, PIPE_CAP_TGSI_FS_COORD_PIXEL_CENTER_HALF_INTEGER
))
4571 /* the driver supports pixel center half integer, need to bias X,Y */
4572 emit_adjusted_wpos(t
, program
, 0.5f
, invert
? 0.5f
: -0.5f
);
4577 /* Fragment shader wants pixel center half integer */
4578 if (pscreen
->get_param(pscreen
, PIPE_CAP_TGSI_FS_COORD_PIXEL_CENTER_HALF_INTEGER
)) {
4579 /* the driver supports pixel center half integer */
4581 else if (pscreen
->get_param(pscreen
, PIPE_CAP_TGSI_FS_COORD_PIXEL_CENTER_INTEGER
)) {
4582 /* the driver supports pixel center integer, need to bias X,Y */
4583 ureg_property_fs_coord_pixel_center(ureg
, TGSI_FS_COORD_PIXEL_CENTER_INTEGER
);
4584 emit_adjusted_wpos(t
, program
, 0.5f
, invert
? -0.5f
: 0.5f
);
4590 /* we invert after adjustment so that we avoid the MOV to temporary,
4591 * and reuse the adjustment ADD instead */
4592 emit_wpos_inversion(t
, program
, invert
);
4596 * OpenGL's fragment gl_FrontFace input is 1 for front-facing, 0 for back.
4597 * TGSI uses +1 for front, -1 for back.
4598 * This function converts the TGSI value to the GL value. Simply clamping/
4599 * saturating the value to [0,1] does the job.
4602 emit_face_var(struct st_translate
*t
)
4604 struct ureg_program
*ureg
= t
->ureg
;
4605 struct ureg_dst face_temp
= ureg_DECL_temporary(ureg
);
4606 struct ureg_src face_input
= t
->inputs
[t
->inputMapping
[FRAG_ATTRIB_FACE
]];
4608 /* MOV_SAT face_temp, input[face] */
4609 face_temp
= ureg_saturate(face_temp
);
4610 ureg_MOV(ureg
, face_temp
, face_input
);
4612 /* Use face_temp as face input from here on: */
4613 t
->inputs
[t
->inputMapping
[FRAG_ATTRIB_FACE
]] = ureg_src(face_temp
);
4617 emit_edgeflags(struct st_translate
*t
)
4619 struct ureg_program
*ureg
= t
->ureg
;
4620 struct ureg_dst edge_dst
= t
->outputs
[t
->outputMapping
[VERT_RESULT_EDGE
]];
4621 struct ureg_src edge_src
= t
->inputs
[t
->inputMapping
[VERT_ATTRIB_EDGEFLAG
]];
4623 ureg_MOV(ureg
, edge_dst
, edge_src
);
4627 * Translate intermediate IR (glsl_to_tgsi_instruction) to TGSI format.
4628 * \param program the program to translate
4629 * \param numInputs number of input registers used
4630 * \param inputMapping maps Mesa fragment program inputs to TGSI generic
4632 * \param inputSemanticName the TGSI_SEMANTIC flag for each input
4633 * \param inputSemanticIndex the semantic index (ex: which texcoord) for
4635 * \param interpMode the TGSI_INTERPOLATE_LINEAR/PERSP mode for each input
4636 * \param numOutputs number of output registers used
4637 * \param outputMapping maps Mesa fragment program outputs to TGSI
4639 * \param outputSemanticName the TGSI_SEMANTIC flag for each output
4640 * \param outputSemanticIndex the semantic index (ex: which texcoord) for
4643 * \return PIPE_OK or PIPE_ERROR_OUT_OF_MEMORY
4645 extern "C" enum pipe_error
4646 st_translate_program(
4647 struct gl_context
*ctx
,
4649 struct ureg_program
*ureg
,
4650 glsl_to_tgsi_visitor
*program
,
4651 const struct gl_program
*proginfo
,
4653 const GLuint inputMapping
[],
4654 const ubyte inputSemanticName
[],
4655 const ubyte inputSemanticIndex
[],
4656 const GLuint interpMode
[],
4658 const GLuint outputMapping
[],
4659 const ubyte outputSemanticName
[],
4660 const ubyte outputSemanticIndex
[],
4661 boolean passthrough_edgeflags
)
4663 struct st_translate translate
, *t
;
4665 enum pipe_error ret
= PIPE_OK
;
4667 assert(numInputs
<= Elements(t
->inputs
));
4668 assert(numOutputs
<= Elements(t
->outputs
));
4671 memset(t
, 0, sizeof *t
);
4673 t
->procType
= procType
;
4674 t
->inputMapping
= inputMapping
;
4675 t
->outputMapping
= outputMapping
;
4677 t
->pointSizeOutIndex
= -1;
4678 t
->prevInstWrotePointSize
= GL_FALSE
;
4681 * Declare input attributes.
4683 if (procType
== TGSI_PROCESSOR_FRAGMENT
) {
4684 for (i
= 0; i
< numInputs
; i
++) {
4685 t
->inputs
[i
] = ureg_DECL_fs_input(ureg
,
4686 inputSemanticName
[i
],
4687 inputSemanticIndex
[i
],
4691 if (proginfo
->InputsRead
& FRAG_BIT_WPOS
) {
4692 /* Must do this after setting up t->inputs, and before
4693 * emitting constant references, below:
4695 emit_wpos(st_context(ctx
), t
, proginfo
, ureg
);
4698 if (proginfo
->InputsRead
& FRAG_BIT_FACE
)
4702 * Declare output attributes.
4704 for (i
= 0; i
< numOutputs
; i
++) {
4705 switch (outputSemanticName
[i
]) {
4706 case TGSI_SEMANTIC_POSITION
:
4707 t
->outputs
[i
] = ureg_DECL_output(ureg
,
4708 TGSI_SEMANTIC_POSITION
, /* Z/Depth */
4709 outputSemanticIndex
[i
]);
4710 t
->outputs
[i
] = ureg_writemask(t
->outputs
[i
], TGSI_WRITEMASK_Z
);
4712 case TGSI_SEMANTIC_STENCIL
:
4713 t
->outputs
[i
] = ureg_DECL_output(ureg
,
4714 TGSI_SEMANTIC_STENCIL
, /* Stencil */
4715 outputSemanticIndex
[i
]);
4716 t
->outputs
[i
] = ureg_writemask(t
->outputs
[i
], TGSI_WRITEMASK_Y
);
4718 case TGSI_SEMANTIC_COLOR
:
4719 t
->outputs
[i
] = ureg_DECL_output(ureg
,
4720 TGSI_SEMANTIC_COLOR
,
4721 outputSemanticIndex
[i
]);
4724 assert(!"fragment shader outputs must be POSITION/STENCIL/COLOR");
4725 return PIPE_ERROR_BAD_INPUT
;
4729 else if (procType
== TGSI_PROCESSOR_GEOMETRY
) {
4730 for (i
= 0; i
< numInputs
; i
++) {
4731 t
->inputs
[i
] = ureg_DECL_gs_input(ureg
,
4733 inputSemanticName
[i
],
4734 inputSemanticIndex
[i
]);
4737 for (i
= 0; i
< numOutputs
; i
++) {
4738 t
->outputs
[i
] = ureg_DECL_output(ureg
,
4739 outputSemanticName
[i
],
4740 outputSemanticIndex
[i
]);
4744 assert(procType
== TGSI_PROCESSOR_VERTEX
);
4746 for (i
= 0; i
< numInputs
; i
++) {
4747 t
->inputs
[i
] = ureg_DECL_vs_input(ureg
, i
);
4750 for (i
= 0; i
< numOutputs
; i
++) {
4751 t
->outputs
[i
] = ureg_DECL_output(ureg
,
4752 outputSemanticName
[i
],
4753 outputSemanticIndex
[i
]);
4754 if ((outputSemanticName
[i
] == TGSI_SEMANTIC_PSIZE
) && proginfo
->Id
) {
4755 /* Writing to the point size result register requires special
4756 * handling to implement clamping.
4758 static const gl_state_index pointSizeClampState
[STATE_LENGTH
]
4759 = { STATE_INTERNAL
, STATE_POINT_SIZE_IMPL_CLAMP
, (gl_state_index
)0, (gl_state_index
)0, (gl_state_index
)0 };
4760 /* XXX: note we are modifying the incoming shader here! Need to
4761 * do this before emitting the constant decls below, or this
4764 unsigned pointSizeClampConst
=
4765 _mesa_add_state_reference(proginfo
->Parameters
,
4766 pointSizeClampState
);
4767 struct ureg_dst psizregtemp
= ureg_DECL_temporary(ureg
);
4768 t
->pointSizeConst
= ureg_DECL_constant(ureg
, pointSizeClampConst
);
4769 t
->pointSizeResult
= t
->outputs
[i
];
4770 t
->pointSizeOutIndex
= i
;
4771 t
->outputs
[i
] = psizregtemp
;
4774 if (passthrough_edgeflags
)
4778 /* Declare address register.
4780 if (program
->num_address_regs
> 0) {
4781 assert(program
->num_address_regs
== 1);
4782 t
->address
[0] = ureg_DECL_address(ureg
);
4785 /* Declare misc input registers
4788 GLbitfield sysInputs
= proginfo
->SystemValuesRead
;
4789 unsigned numSys
= 0;
4790 for (i
= 0; sysInputs
; i
++) {
4791 if (sysInputs
& (1 << i
)) {
4792 unsigned semName
= mesa_sysval_to_semantic
[i
];
4793 t
->systemValues
[i
] = ureg_DECL_system_value(ureg
, numSys
, semName
, 0);
4795 sysInputs
&= ~(1 << i
);
4800 if (program
->indirect_addr_temps
) {
4801 /* If temps are accessed with indirect addressing, declare temporaries
4802 * in sequential order. Else, we declare them on demand elsewhere.
4803 * (Note: the number of temporaries is equal to program->next_temp)
4805 for (i
= 0; i
< (unsigned)program
->next_temp
; i
++) {
4806 /* XXX use TGSI_FILE_TEMPORARY_ARRAY when it's supported by ureg */
4807 t
->temps
[i
] = ureg_DECL_temporary(t
->ureg
);
4811 /* Emit constants and uniforms. TGSI uses a single index space for these,
4812 * so we put all the translated regs in t->constants.
4814 if (proginfo
->Parameters
) {
4815 t
->constants
= (struct ureg_src
*)CALLOC(proginfo
->Parameters
->NumParameters
* sizeof(t
->constants
[0]));
4816 if (t
->constants
== NULL
) {
4817 ret
= PIPE_ERROR_OUT_OF_MEMORY
;
4821 for (i
= 0; i
< proginfo
->Parameters
->NumParameters
; i
++) {
4822 switch (proginfo
->Parameters
->Parameters
[i
].Type
) {
4823 case PROGRAM_ENV_PARAM
:
4824 case PROGRAM_LOCAL_PARAM
:
4825 case PROGRAM_STATE_VAR
:
4826 case PROGRAM_NAMED_PARAM
:
4827 case PROGRAM_UNIFORM
:
4828 t
->constants
[i
] = ureg_DECL_constant(ureg
, i
);
4831 /* Emit immediates for PROGRAM_CONSTANT only when there's no indirect
4832 * addressing of the const buffer.
4833 * FIXME: Be smarter and recognize param arrays:
4834 * indirect addressing is only valid within the referenced
4837 case PROGRAM_CONSTANT
:
4838 if (program
->indirect_addr_consts
)
4839 t
->constants
[i
] = ureg_DECL_constant(ureg
, i
);
4841 t
->constants
[i
] = emit_immediate(t
,
4842 proginfo
->Parameters
->ParameterValues
[i
],
4843 proginfo
->Parameters
->Parameters
[i
].DataType
,
4852 /* Emit immediate values.
4854 t
->immediates
= (struct ureg_src
*)CALLOC(program
->num_immediates
* sizeof(struct ureg_src
));
4855 if (t
->immediates
== NULL
) {
4856 ret
= PIPE_ERROR_OUT_OF_MEMORY
;
4860 foreach_iter(exec_list_iterator
, iter
, program
->immediates
) {
4861 immediate_storage
*imm
= (immediate_storage
*)iter
.get();
4862 t
->immediates
[i
++] = emit_immediate(t
, imm
->values
, imm
->type
, imm
->size
);
4865 /* texture samplers */
4866 for (i
= 0; i
< ctx
->Const
.MaxTextureImageUnits
; i
++) {
4867 if (program
->samplers_used
& (1 << i
)) {
4868 t
->samplers
[i
] = ureg_DECL_sampler(ureg
, i
);
4872 /* Emit each instruction in turn:
4874 foreach_iter(exec_list_iterator
, iter
, program
->instructions
) {
4875 set_insn_start(t
, ureg_get_instruction_number(ureg
));
4876 compile_tgsi_instruction(t
, (glsl_to_tgsi_instruction
*)iter
.get());
4878 if (t
->prevInstWrotePointSize
&& proginfo
->Id
) {
4879 /* The previous instruction wrote to the (fake) vertex point size
4880 * result register. Now we need to clamp that value to the min/max
4881 * point size range, putting the result into the real point size
4883 * Note that we can't do this easily at the end of program due to
4884 * possible early return.
4886 set_insn_start(t
, ureg_get_instruction_number(ureg
));
4888 ureg_writemask(t
->outputs
[t
->pointSizeOutIndex
], WRITEMASK_X
),
4889 ureg_src(t
->outputs
[t
->pointSizeOutIndex
]),
4890 ureg_swizzle(t
->pointSizeConst
, 1,1,1,1));
4891 ureg_MIN(t
->ureg
, ureg_writemask(t
->pointSizeResult
, WRITEMASK_X
),
4892 ureg_src(t
->outputs
[t
->pointSizeOutIndex
]),
4893 ureg_swizzle(t
->pointSizeConst
, 2,2,2,2));
4895 t
->prevInstWrotePointSize
= GL_FALSE
;
4898 /* Fix up all emitted labels:
4900 for (i
= 0; i
< t
->labels_count
; i
++) {
4901 ureg_fixup_label(ureg
, t
->labels
[i
].token
,
4902 t
->insn
[t
->labels
[i
].branch_target
]);
4909 FREE(t
->immediates
);
4912 debug_printf("%s: translate error flag set\n", __FUNCTION__
);
4917 /* ----------------------------- End TGSI code ------------------------------ */
4920 * Convert a shader's GLSL IR into a Mesa gl_program, although without
4921 * generating Mesa IR.
4923 static struct gl_program
*
4924 get_mesa_program(struct gl_context
*ctx
,
4925 struct gl_shader_program
*shader_program
,
4926 struct gl_shader
*shader
)
4928 glsl_to_tgsi_visitor
* v
= new glsl_to_tgsi_visitor();
4929 struct gl_program
*prog
;
4931 const char *target_string
;
4933 struct gl_shader_compiler_options
*options
=
4934 &ctx
->ShaderCompilerOptions
[_mesa_shader_type_to_index(shader
->Type
)];
4936 switch (shader
->Type
) {
4937 case GL_VERTEX_SHADER
:
4938 target
= GL_VERTEX_PROGRAM_ARB
;
4939 target_string
= "vertex";
4941 case GL_FRAGMENT_SHADER
:
4942 target
= GL_FRAGMENT_PROGRAM_ARB
;
4943 target_string
= "fragment";
4945 case GL_GEOMETRY_SHADER
:
4946 target
= GL_GEOMETRY_PROGRAM_NV
;
4947 target_string
= "geometry";
4950 assert(!"should not be reached");
4954 validate_ir_tree(shader
->ir
);
4956 prog
= ctx
->Driver
.NewProgram(ctx
, target
, shader_program
->Name
);
4959 prog
->Parameters
= _mesa_new_parameter_list();
4962 v
->shader_program
= shader_program
;
4963 v
->options
= options
;
4964 v
->glsl_version
= ctx
->Const
.GLSLVersion
;
4965 v
->native_integers
= ctx
->Const
.NativeIntegers
;
4967 add_uniforms_to_parameters_list(shader_program
, shader
, prog
);
4969 /* Emit intermediate IR for main(). */
4970 visit_exec_list(shader
->ir
, v
);
4972 /* Now emit bodies for any functions that were used. */
4974 progress
= GL_FALSE
;
4976 foreach_iter(exec_list_iterator
, iter
, v
->function_signatures
) {
4977 function_entry
*entry
= (function_entry
*)iter
.get();
4979 if (!entry
->bgn_inst
) {
4980 v
->current_function
= entry
;
4982 entry
->bgn_inst
= v
->emit(NULL
, TGSI_OPCODE_BGNSUB
);
4983 entry
->bgn_inst
->function
= entry
;
4985 visit_exec_list(&entry
->sig
->body
, v
);
4987 glsl_to_tgsi_instruction
*last
;
4988 last
= (glsl_to_tgsi_instruction
*)v
->instructions
.get_tail();
4989 if (last
->op
!= TGSI_OPCODE_RET
)
4990 v
->emit(NULL
, TGSI_OPCODE_RET
);
4992 glsl_to_tgsi_instruction
*end
;
4993 end
= v
->emit(NULL
, TGSI_OPCODE_ENDSUB
);
4994 end
->function
= entry
;
5002 /* Print out some information (for debugging purposes) used by the
5003 * optimization passes. */
5004 for (i
=0; i
< v
->next_temp
; i
++) {
5005 int fr
= v
->get_first_temp_read(i
);
5006 int fw
= v
->get_first_temp_write(i
);
5007 int lr
= v
->get_last_temp_read(i
);
5008 int lw
= v
->get_last_temp_write(i
);
5010 printf("Temp %d: FR=%3d FW=%3d LR=%3d LW=%3d\n", i
, fr
, fw
, lr
, lw
);
5015 /* Remove reads to output registers, and to varyings in vertex shaders. */
5016 v
->remove_output_reads(PROGRAM_OUTPUT
);
5017 if (target
== GL_VERTEX_PROGRAM_ARB
)
5018 v
->remove_output_reads(PROGRAM_VARYING
);
5020 /* Perform optimizations on the instructions in the glsl_to_tgsi_visitor. */
5022 v
->copy_propagate();
5023 while (v
->eliminate_dead_code_advanced());
5025 /* FIXME: These passes to optimize temporary registers don't work when there
5026 * is indirect addressing of the temporary register space. We need proper
5027 * array support so that we don't have to give up these passes in every
5028 * shader that uses arrays.
5030 if (!v
->indirect_addr_temps
) {
5031 v
->eliminate_dead_code();
5032 v
->merge_registers();
5033 v
->renumber_registers();
5036 /* Write the END instruction. */
5037 v
->emit(NULL
, TGSI_OPCODE_END
);
5039 if (ctx
->Shader
.Flags
& GLSL_DUMP
) {
5041 printf("GLSL IR for linked %s program %d:\n", target_string
,
5042 shader_program
->Name
);
5043 _mesa_print_ir(shader
->ir
, NULL
);
5048 prog
->Instructions
= NULL
;
5049 prog
->NumInstructions
= 0;
5051 do_set_program_inouts(shader
->ir
, prog
);
5052 count_resources(v
, prog
);
5054 check_resources(ctx
, shader_program
, v
, prog
);
5056 _mesa_reference_program(ctx
, &shader
->Program
, prog
);
5058 struct st_vertex_program
*stvp
;
5059 struct st_fragment_program
*stfp
;
5060 struct st_geometry_program
*stgp
;
5062 switch (shader
->Type
) {
5063 case GL_VERTEX_SHADER
:
5064 stvp
= (struct st_vertex_program
*)prog
;
5065 stvp
->glsl_to_tgsi
= v
;
5067 case GL_FRAGMENT_SHADER
:
5068 stfp
= (struct st_fragment_program
*)prog
;
5069 stfp
->glsl_to_tgsi
= v
;
5071 case GL_GEOMETRY_SHADER
:
5072 stgp
= (struct st_geometry_program
*)prog
;
5073 stgp
->glsl_to_tgsi
= v
;
5076 assert(!"should not be reached");
5086 st_new_shader(struct gl_context
*ctx
, GLuint name
, GLuint type
)
5088 struct gl_shader
*shader
;
5089 assert(type
== GL_FRAGMENT_SHADER
|| type
== GL_VERTEX_SHADER
||
5090 type
== GL_GEOMETRY_SHADER_ARB
);
5091 shader
= rzalloc(NULL
, struct gl_shader
);
5093 shader
->Type
= type
;
5094 shader
->Name
= name
;
5095 _mesa_init_shader(ctx
, shader
);
5100 struct gl_shader_program
*
5101 st_new_shader_program(struct gl_context
*ctx
, GLuint name
)
5103 struct gl_shader_program
*shProg
;
5104 shProg
= rzalloc(NULL
, struct gl_shader_program
);
5106 shProg
->Name
= name
;
5107 _mesa_init_shader_program(ctx
, shProg
);
5114 * Called via ctx->Driver.LinkShader()
5115 * This actually involves converting GLSL IR into an intermediate TGSI-like IR
5116 * with code lowering and other optimizations.
5119 st_link_shader(struct gl_context
*ctx
, struct gl_shader_program
*prog
)
5121 assert(prog
->LinkStatus
);
5123 for (unsigned i
= 0; i
< MESA_SHADER_TYPES
; i
++) {
5124 if (prog
->_LinkedShaders
[i
] == NULL
)
5128 exec_list
*ir
= prog
->_LinkedShaders
[i
]->ir
;
5129 const struct gl_shader_compiler_options
*options
=
5130 &ctx
->ShaderCompilerOptions
[_mesa_shader_type_to_index(prog
->_LinkedShaders
[i
]->Type
)];
5136 do_mat_op_to_vec(ir
);
5137 lower_instructions(ir
, (MOD_TO_FRACT
| DIV_TO_MUL_RCP
| EXP_TO_EXP2
5138 | LOG_TO_LOG2
| INT_DIV_TO_MUL_RCP
5139 | ((options
->EmitNoPow
) ? POW_TO_EXP2
: 0)));
5141 progress
= do_lower_jumps(ir
, true, true, options
->EmitNoMainReturn
, options
->EmitNoCont
, options
->EmitNoLoops
) || progress
;
5143 progress
= do_common_optimization(ir
, true, options
->MaxUnrollIterations
) || progress
;
5145 progress
= lower_quadop_vector(ir
, false) || progress
;
5147 if (options
->MaxIfDepth
== 0)
5148 progress
= lower_discard(ir
) || progress
;
5150 progress
= lower_if_to_cond_assign(ir
, options
->MaxIfDepth
) || progress
;
5152 if (options
->EmitNoNoise
)
5153 progress
= lower_noise(ir
) || progress
;
5155 /* If there are forms of indirect addressing that the driver
5156 * cannot handle, perform the lowering pass.
5158 if (options
->EmitNoIndirectInput
|| options
->EmitNoIndirectOutput
5159 || options
->EmitNoIndirectTemp
|| options
->EmitNoIndirectUniform
)
5161 lower_variable_index_to_cond_assign(ir
,
5162 options
->EmitNoIndirectInput
,
5163 options
->EmitNoIndirectOutput
,
5164 options
->EmitNoIndirectTemp
,
5165 options
->EmitNoIndirectUniform
)
5168 progress
= do_vec_index_to_cond_assign(ir
) || progress
;
5171 validate_ir_tree(ir
);
5174 for (unsigned i
= 0; i
< MESA_SHADER_TYPES
; i
++) {
5175 struct gl_program
*linked_prog
;
5177 if (prog
->_LinkedShaders
[i
] == NULL
)
5180 linked_prog
= get_mesa_program(ctx
, prog
, prog
->_LinkedShaders
[i
]);
5185 switch (prog
->_LinkedShaders
[i
]->Type
) {
5186 case GL_VERTEX_SHADER
:
5187 _mesa_reference_vertprog(ctx
, &prog
->VertexProgram
,
5188 (struct gl_vertex_program
*)linked_prog
);
5189 ok
= ctx
->Driver
.ProgramStringNotify(ctx
, GL_VERTEX_PROGRAM_ARB
,
5192 _mesa_reference_vertprog(ctx
, &prog
->VertexProgram
, NULL
);
5195 case GL_FRAGMENT_SHADER
:
5196 _mesa_reference_fragprog(ctx
, &prog
->FragmentProgram
,
5197 (struct gl_fragment_program
*)linked_prog
);
5198 ok
= ctx
->Driver
.ProgramStringNotify(ctx
, GL_FRAGMENT_PROGRAM_ARB
,
5201 _mesa_reference_fragprog(ctx
, &prog
->FragmentProgram
, NULL
);
5204 case GL_GEOMETRY_SHADER
:
5205 _mesa_reference_geomprog(ctx
, &prog
->GeometryProgram
,
5206 (struct gl_geometry_program
*)linked_prog
);
5207 ok
= ctx
->Driver
.ProgramStringNotify(ctx
, GL_GEOMETRY_PROGRAM_NV
,
5210 _mesa_reference_geomprog(ctx
, &prog
->GeometryProgram
, NULL
);
5215 _mesa_reference_program(ctx
, &prog
->_LinkedShaders
[i
]->Program
, NULL
);
5216 _mesa_reference_program(ctx
, &linked_prog
, NULL
);
5221 _mesa_reference_program(ctx
, &linked_prog
, NULL
);