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
86 static int swizzle_for_size(int size
);
89 * This struct is a corresponding struct to TGSI ureg_src.
93 st_src_reg(gl_register_file file
, int index
, const glsl_type
*type
)
97 if (type
&& (type
->is_scalar() || type
->is_vector() || type
->is_matrix()))
98 this->swizzle
= swizzle_for_size(type
->vector_elements
);
100 this->swizzle
= SWIZZLE_XYZW
;
102 this->type
= type
? type
->base_type
: GLSL_TYPE_ERROR
;
103 this->reladdr
= NULL
;
106 st_src_reg(gl_register_file file
, int index
, int type
)
111 this->swizzle
= SWIZZLE_XYZW
;
113 this->reladdr
= NULL
;
118 this->type
= GLSL_TYPE_ERROR
;
119 this->file
= PROGRAM_UNDEFINED
;
123 this->reladdr
= NULL
;
126 explicit st_src_reg(st_dst_reg reg
);
128 gl_register_file file
; /**< PROGRAM_* from Mesa */
129 int index
; /**< temporary index, VERT_ATTRIB_*, FRAG_ATTRIB_*, etc. */
130 GLuint swizzle
; /**< SWIZZLE_XYZWONEZERO swizzles from Mesa. */
131 int negate
; /**< NEGATE_XYZW mask from mesa */
132 int type
; /** GLSL_TYPE_* from GLSL IR (enum glsl_base_type) */
133 /** Register index should be offset by the integer in this reg. */
139 st_dst_reg(gl_register_file file
, int writemask
, int type
)
143 this->writemask
= writemask
;
144 this->cond_mask
= COND_TR
;
145 this->reladdr
= NULL
;
151 this->type
= GLSL_TYPE_ERROR
;
152 this->file
= PROGRAM_UNDEFINED
;
155 this->cond_mask
= COND_TR
;
156 this->reladdr
= NULL
;
159 explicit st_dst_reg(st_src_reg reg
);
161 gl_register_file file
; /**< PROGRAM_* from Mesa */
162 int index
; /**< temporary index, VERT_ATTRIB_*, FRAG_ATTRIB_*, etc. */
163 int writemask
; /**< Bitfield of WRITEMASK_[XYZW] */
165 int type
; /** GLSL_TYPE_* from GLSL IR (enum glsl_base_type) */
166 /** Register index should be offset by the integer in this reg. */
170 st_src_reg::st_src_reg(st_dst_reg reg
)
172 this->type
= reg
.type
;
173 this->file
= reg
.file
;
174 this->index
= reg
.index
;
175 this->swizzle
= SWIZZLE_XYZW
;
177 this->reladdr
= reg
.reladdr
;
180 st_dst_reg::st_dst_reg(st_src_reg reg
)
182 this->type
= reg
.type
;
183 this->file
= reg
.file
;
184 this->index
= reg
.index
;
185 this->writemask
= WRITEMASK_XYZW
;
186 this->cond_mask
= COND_TR
;
187 this->reladdr
= reg
.reladdr
;
190 class glsl_to_tgsi_instruction
: public exec_node
{
192 /* Callers of this ralloc-based new need not call delete. It's
193 * easier to just ralloc_free 'ctx' (or any of its ancestors). */
194 static void* operator new(size_t size
, void *ctx
)
198 node
= rzalloc_size(ctx
, size
);
199 assert(node
!= NULL
);
207 /** Pointer to the ir source this tree came from for debugging */
209 GLboolean cond_update
;
211 int sampler
; /**< sampler index */
212 int tex_target
; /**< One of TEXTURE_*_INDEX */
213 GLboolean tex_shadow
;
214 int dead_mask
; /**< Used in dead code elimination */
216 class function_entry
*function
; /* Set on TGSI_OPCODE_CAL or TGSI_OPCODE_BGNSUB */
219 class variable_storage
: public exec_node
{
221 variable_storage(ir_variable
*var
, gl_register_file file
, int index
)
222 : file(file
), index(index
), var(var
)
227 gl_register_file file
;
229 ir_variable
*var
; /* variable that maps to this, if any */
232 class immediate_storage
: public exec_node
{
234 immediate_storage(gl_constant_value
*values
, int size
, int type
)
236 memcpy(this->values
, values
, size
* sizeof(gl_constant_value
));
241 gl_constant_value values
[4];
242 int size
; /**< Number of components (1-4) */
243 int type
; /**< GL_FLOAT, GL_INT, GL_BOOL, or GL_UNSIGNED_INT */
246 class function_entry
: public exec_node
{
248 ir_function_signature
*sig
;
251 * identifier of this function signature used by the program.
253 * At the point that TGSI instructions for function calls are
254 * generated, we don't know the address of the first instruction of
255 * the function body. So we make the BranchTarget that is called a
256 * small integer and rewrite them during set_branchtargets().
261 * Pointer to first instruction of the function body.
263 * Set during function body emits after main() is processed.
265 glsl_to_tgsi_instruction
*bgn_inst
;
268 * Index of the first instruction of the function body in actual TGSI.
270 * Set after conversion from glsl_to_tgsi_instruction to TGSI.
274 /** Storage for the return value. */
275 st_src_reg return_reg
;
278 class glsl_to_tgsi_visitor
: public ir_visitor
{
280 glsl_to_tgsi_visitor();
281 ~glsl_to_tgsi_visitor();
283 function_entry
*current_function
;
285 struct gl_context
*ctx
;
286 struct gl_program
*prog
;
287 struct gl_shader_program
*shader_program
;
288 struct gl_shader_compiler_options
*options
;
292 int num_address_regs
;
294 bool indirect_addr_temps
;
295 bool indirect_addr_consts
;
298 bool native_integers
;
300 variable_storage
*find_variable_storage(ir_variable
*var
);
302 int add_constant(gl_register_file file
, gl_constant_value values
[4],
303 int size
, int datatype
, GLuint
*swizzle_out
);
305 function_entry
*get_function_signature(ir_function_signature
*sig
);
307 st_src_reg
get_temp(const glsl_type
*type
);
308 void reladdr_to_temp(ir_instruction
*ir
, st_src_reg
*reg
, int *num_reladdr
);
310 st_src_reg
st_src_reg_for_float(float val
);
311 st_src_reg
st_src_reg_for_int(int val
);
312 st_src_reg
st_src_reg_for_type(int type
, int val
);
315 * \name Visit methods
317 * As typical for the visitor pattern, there must be one \c visit method for
318 * each concrete subclass of \c ir_instruction. Virtual base classes within
319 * the hierarchy should not have \c visit methods.
322 virtual void visit(ir_variable
*);
323 virtual void visit(ir_loop
*);
324 virtual void visit(ir_loop_jump
*);
325 virtual void visit(ir_function_signature
*);
326 virtual void visit(ir_function
*);
327 virtual void visit(ir_expression
*);
328 virtual void visit(ir_swizzle
*);
329 virtual void visit(ir_dereference_variable
*);
330 virtual void visit(ir_dereference_array
*);
331 virtual void visit(ir_dereference_record
*);
332 virtual void visit(ir_assignment
*);
333 virtual void visit(ir_constant
*);
334 virtual void visit(ir_call
*);
335 virtual void visit(ir_return
*);
336 virtual void visit(ir_discard
*);
337 virtual void visit(ir_texture
*);
338 virtual void visit(ir_if
*);
343 /** List of variable_storage */
346 /** List of immediate_storage */
347 exec_list immediates
;
350 /** List of function_entry */
351 exec_list function_signatures
;
352 int next_signature_id
;
354 /** List of glsl_to_tgsi_instruction */
355 exec_list instructions
;
357 glsl_to_tgsi_instruction
*emit(ir_instruction
*ir
, unsigned op
);
359 glsl_to_tgsi_instruction
*emit(ir_instruction
*ir
, unsigned op
,
360 st_dst_reg dst
, st_src_reg src0
);
362 glsl_to_tgsi_instruction
*emit(ir_instruction
*ir
, unsigned op
,
363 st_dst_reg dst
, st_src_reg src0
, st_src_reg src1
);
365 glsl_to_tgsi_instruction
*emit(ir_instruction
*ir
, unsigned op
,
367 st_src_reg src0
, st_src_reg src1
, st_src_reg src2
);
369 unsigned get_opcode(ir_instruction
*ir
, unsigned op
,
371 st_src_reg src0
, st_src_reg src1
);
374 * Emit the correct dot-product instruction for the type of arguments
376 glsl_to_tgsi_instruction
*emit_dp(ir_instruction
*ir
,
382 void emit_scalar(ir_instruction
*ir
, unsigned op
,
383 st_dst_reg dst
, st_src_reg src0
);
385 void emit_scalar(ir_instruction
*ir
, unsigned op
,
386 st_dst_reg dst
, st_src_reg src0
, st_src_reg src1
);
388 void emit_arl(ir_instruction
*ir
, st_dst_reg dst
, st_src_reg src0
);
390 void emit_scs(ir_instruction
*ir
, unsigned op
,
391 st_dst_reg dst
, const st_src_reg
&src
);
393 bool try_emit_mad(ir_expression
*ir
,
395 bool try_emit_mad_for_and_not(ir_expression
*ir
,
397 bool try_emit_sat(ir_expression
*ir
);
399 void emit_swz(ir_expression
*ir
);
401 bool process_move_condition(ir_rvalue
*ir
);
403 void remove_output_reads(gl_register_file type
);
404 void simplify_cmp(void);
406 void rename_temp_register(int index
, int new_index
);
407 int get_first_temp_read(int index
);
408 int get_first_temp_write(int index
);
409 int get_last_temp_read(int index
);
410 int get_last_temp_write(int index
);
412 void copy_propagate(void);
413 void eliminate_dead_code(void);
414 int eliminate_dead_code_advanced(void);
415 void merge_registers(void);
416 void renumber_registers(void);
421 static st_src_reg undef_src
= st_src_reg(PROGRAM_UNDEFINED
, 0, GLSL_TYPE_ERROR
);
423 static st_dst_reg undef_dst
= st_dst_reg(PROGRAM_UNDEFINED
, SWIZZLE_NOOP
, GLSL_TYPE_ERROR
);
425 static st_dst_reg address_reg
= st_dst_reg(PROGRAM_ADDRESS
, WRITEMASK_X
, GLSL_TYPE_FLOAT
);
428 fail_link(struct gl_shader_program
*prog
, const char *fmt
, ...) PRINTFLIKE(2, 3);
431 fail_link(struct gl_shader_program
*prog
, const char *fmt
, ...)
435 ralloc_vasprintf_append(&prog
->InfoLog
, fmt
, args
);
438 prog
->LinkStatus
= GL_FALSE
;
442 swizzle_for_size(int size
)
444 int size_swizzles
[4] = {
445 MAKE_SWIZZLE4(SWIZZLE_X
, SWIZZLE_X
, SWIZZLE_X
, SWIZZLE_X
),
446 MAKE_SWIZZLE4(SWIZZLE_X
, SWIZZLE_Y
, SWIZZLE_Y
, SWIZZLE_Y
),
447 MAKE_SWIZZLE4(SWIZZLE_X
, SWIZZLE_Y
, SWIZZLE_Z
, SWIZZLE_Z
),
448 MAKE_SWIZZLE4(SWIZZLE_X
, SWIZZLE_Y
, SWIZZLE_Z
, SWIZZLE_W
),
451 assert((size
>= 1) && (size
<= 4));
452 return size_swizzles
[size
- 1];
456 is_tex_instruction(unsigned opcode
)
458 const tgsi_opcode_info
* info
= tgsi_get_opcode_info(opcode
);
463 num_inst_dst_regs(unsigned opcode
)
465 const tgsi_opcode_info
* info
= tgsi_get_opcode_info(opcode
);
466 return info
->num_dst
;
470 num_inst_src_regs(unsigned opcode
)
472 const tgsi_opcode_info
* info
= tgsi_get_opcode_info(opcode
);
473 return info
->is_tex
? info
->num_src
- 1 : info
->num_src
;
476 glsl_to_tgsi_instruction
*
477 glsl_to_tgsi_visitor::emit(ir_instruction
*ir
, unsigned op
,
479 st_src_reg src0
, st_src_reg src1
, st_src_reg src2
)
481 glsl_to_tgsi_instruction
*inst
= new(mem_ctx
) glsl_to_tgsi_instruction();
482 int num_reladdr
= 0, i
;
484 op
= get_opcode(ir
, op
, dst
, src0
, src1
);
486 /* If we have to do relative addressing, we want to load the ARL
487 * reg directly for one of the regs, and preload the other reladdr
488 * sources into temps.
490 num_reladdr
+= dst
.reladdr
!= NULL
;
491 num_reladdr
+= src0
.reladdr
!= NULL
;
492 num_reladdr
+= src1
.reladdr
!= NULL
;
493 num_reladdr
+= src2
.reladdr
!= NULL
;
495 reladdr_to_temp(ir
, &src2
, &num_reladdr
);
496 reladdr_to_temp(ir
, &src1
, &num_reladdr
);
497 reladdr_to_temp(ir
, &src0
, &num_reladdr
);
500 emit_arl(ir
, address_reg
, *dst
.reladdr
);
503 assert(num_reladdr
== 0);
513 inst
->function
= NULL
;
515 if (op
== TGSI_OPCODE_ARL
)
516 this->num_address_regs
= 1;
518 /* Update indirect addressing status used by TGSI */
521 case PROGRAM_TEMPORARY
:
522 this->indirect_addr_temps
= true;
524 case PROGRAM_LOCAL_PARAM
:
525 case PROGRAM_ENV_PARAM
:
526 case PROGRAM_STATE_VAR
:
527 case PROGRAM_NAMED_PARAM
:
528 case PROGRAM_CONSTANT
:
529 case PROGRAM_UNIFORM
:
530 this->indirect_addr_consts
= true;
532 case PROGRAM_IMMEDIATE
:
533 assert(!"immediates should not have indirect addressing");
540 for (i
=0; i
<3; i
++) {
541 if(inst
->src
[i
].reladdr
) {
542 switch(inst
->src
[i
].file
) {
543 case PROGRAM_TEMPORARY
:
544 this->indirect_addr_temps
= true;
546 case PROGRAM_LOCAL_PARAM
:
547 case PROGRAM_ENV_PARAM
:
548 case PROGRAM_STATE_VAR
:
549 case PROGRAM_NAMED_PARAM
:
550 case PROGRAM_CONSTANT
:
551 case PROGRAM_UNIFORM
:
552 this->indirect_addr_consts
= true;
554 case PROGRAM_IMMEDIATE
:
555 assert(!"immediates should not have indirect addressing");
564 this->instructions
.push_tail(inst
);
570 glsl_to_tgsi_instruction
*
571 glsl_to_tgsi_visitor::emit(ir_instruction
*ir
, unsigned op
,
572 st_dst_reg dst
, st_src_reg src0
, st_src_reg src1
)
574 return emit(ir
, op
, dst
, src0
, src1
, undef_src
);
577 glsl_to_tgsi_instruction
*
578 glsl_to_tgsi_visitor::emit(ir_instruction
*ir
, unsigned op
,
579 st_dst_reg dst
, st_src_reg src0
)
581 assert(dst
.writemask
!= 0);
582 return emit(ir
, op
, dst
, src0
, undef_src
, undef_src
);
585 glsl_to_tgsi_instruction
*
586 glsl_to_tgsi_visitor::emit(ir_instruction
*ir
, unsigned op
)
588 return emit(ir
, op
, undef_dst
, undef_src
, undef_src
, undef_src
);
592 * Determines whether to use an integer, unsigned integer, or float opcode
593 * based on the operands and input opcode, then emits the result.
595 * TODO: type checking for remaining TGSI opcodes
598 glsl_to_tgsi_visitor::get_opcode(ir_instruction
*ir
, unsigned op
,
600 st_src_reg src0
, st_src_reg src1
)
602 int type
= GLSL_TYPE_FLOAT
;
604 if (src0
.type
== GLSL_TYPE_FLOAT
|| src1
.type
== GLSL_TYPE_FLOAT
)
605 type
= GLSL_TYPE_FLOAT
;
606 else if (native_integers
)
609 #define case4(c, f, i, u) \
610 case TGSI_OPCODE_##c: \
611 if (type == GLSL_TYPE_INT) op = TGSI_OPCODE_##i; \
612 else if (type == GLSL_TYPE_UINT) op = TGSI_OPCODE_##u; \
613 else op = TGSI_OPCODE_##f; \
615 #define case3(f, i, u) case4(f, f, i, u)
616 #define case2fi(f, i) case4(f, f, i, i)
617 #define case2iu(i, u) case4(i, LAST, i, u)
623 case3(DIV
, IDIV
, UDIV
);
624 case3(MAX
, IMAX
, UMAX
);
625 case3(MIN
, IMIN
, UMIN
);
630 case3(SGE
, ISGE
, USGE
);
631 case3(SLT
, ISLT
, USLT
);
643 assert(op
!= TGSI_OPCODE_LAST
);
647 glsl_to_tgsi_instruction
*
648 glsl_to_tgsi_visitor::emit_dp(ir_instruction
*ir
,
649 st_dst_reg dst
, st_src_reg src0
, st_src_reg src1
,
652 static const unsigned dot_opcodes
[] = {
653 TGSI_OPCODE_DP2
, TGSI_OPCODE_DP3
, TGSI_OPCODE_DP4
656 return emit(ir
, dot_opcodes
[elements
- 2], dst
, src0
, src1
);
660 * Emits TGSI scalar opcodes to produce unique answers across channels.
662 * Some TGSI opcodes are scalar-only, like ARB_fp/vp. The src X
663 * channel determines the result across all channels. So to do a vec4
664 * of this operation, we want to emit a scalar per source channel used
665 * to produce dest channels.
668 glsl_to_tgsi_visitor::emit_scalar(ir_instruction
*ir
, unsigned op
,
670 st_src_reg orig_src0
, st_src_reg orig_src1
)
673 int done_mask
= ~dst
.writemask
;
675 /* TGSI RCP is a scalar operation splatting results to all channels,
676 * like ARB_fp/vp. So emit as many RCPs as necessary to cover our
679 for (i
= 0; i
< 4; i
++) {
680 GLuint this_mask
= (1 << i
);
681 glsl_to_tgsi_instruction
*inst
;
682 st_src_reg src0
= orig_src0
;
683 st_src_reg src1
= orig_src1
;
685 if (done_mask
& this_mask
)
688 GLuint src0_swiz
= GET_SWZ(src0
.swizzle
, i
);
689 GLuint src1_swiz
= GET_SWZ(src1
.swizzle
, i
);
690 for (j
= i
+ 1; j
< 4; j
++) {
691 /* If there is another enabled component in the destination that is
692 * derived from the same inputs, generate its value on this pass as
695 if (!(done_mask
& (1 << j
)) &&
696 GET_SWZ(src0
.swizzle
, j
) == src0_swiz
&&
697 GET_SWZ(src1
.swizzle
, j
) == src1_swiz
) {
698 this_mask
|= (1 << j
);
701 src0
.swizzle
= MAKE_SWIZZLE4(src0_swiz
, src0_swiz
,
702 src0_swiz
, src0_swiz
);
703 src1
.swizzle
= MAKE_SWIZZLE4(src1_swiz
, src1_swiz
,
704 src1_swiz
, src1_swiz
);
706 inst
= emit(ir
, op
, dst
, src0
, src1
);
707 inst
->dst
.writemask
= this_mask
;
708 done_mask
|= this_mask
;
713 glsl_to_tgsi_visitor::emit_scalar(ir_instruction
*ir
, unsigned op
,
714 st_dst_reg dst
, st_src_reg src0
)
716 st_src_reg undef
= undef_src
;
718 undef
.swizzle
= SWIZZLE_XXXX
;
720 emit_scalar(ir
, op
, dst
, src0
, undef
);
724 glsl_to_tgsi_visitor::emit_arl(ir_instruction
*ir
,
725 st_dst_reg dst
, st_src_reg src0
)
727 st_src_reg tmp
= get_temp(glsl_type::float_type
);
729 if (src0
.type
== GLSL_TYPE_INT
)
730 emit(NULL
, TGSI_OPCODE_I2F
, st_dst_reg(tmp
), src0
);
731 else if (src0
.type
== GLSL_TYPE_UINT
)
732 emit(NULL
, TGSI_OPCODE_U2F
, st_dst_reg(tmp
), src0
);
736 emit(NULL
, TGSI_OPCODE_ARL
, dst
, tmp
);
740 * Emit an TGSI_OPCODE_SCS instruction
742 * The \c SCS opcode functions a bit differently than the other TGSI opcodes.
743 * Instead of splatting its result across all four components of the
744 * destination, it writes one value to the \c x component and another value to
745 * the \c y component.
747 * \param ir IR instruction being processed
748 * \param op Either \c TGSI_OPCODE_SIN or \c TGSI_OPCODE_COS depending
749 * on which value is desired.
750 * \param dst Destination register
751 * \param src Source register
754 glsl_to_tgsi_visitor::emit_scs(ir_instruction
*ir
, unsigned op
,
756 const st_src_reg
&src
)
758 /* Vertex programs cannot use the SCS opcode.
760 if (this->prog
->Target
== GL_VERTEX_PROGRAM_ARB
) {
761 emit_scalar(ir
, op
, dst
, src
);
765 const unsigned component
= (op
== TGSI_OPCODE_SIN
) ? 0 : 1;
766 const unsigned scs_mask
= (1U << component
);
767 int done_mask
= ~dst
.writemask
;
770 assert(op
== TGSI_OPCODE_SIN
|| op
== TGSI_OPCODE_COS
);
772 /* If there are compnents in the destination that differ from the component
773 * that will be written by the SCS instrution, we'll need a temporary.
775 if (scs_mask
!= unsigned(dst
.writemask
)) {
776 tmp
= get_temp(glsl_type::vec4_type
);
779 for (unsigned i
= 0; i
< 4; i
++) {
780 unsigned this_mask
= (1U << i
);
781 st_src_reg src0
= src
;
783 if ((done_mask
& this_mask
) != 0)
786 /* The source swizzle specified which component of the source generates
787 * sine / cosine for the current component in the destination. The SCS
788 * instruction requires that this value be swizzle to the X component.
789 * Replace the current swizzle with a swizzle that puts the source in
792 unsigned src0_swiz
= GET_SWZ(src
.swizzle
, i
);
794 src0
.swizzle
= MAKE_SWIZZLE4(src0_swiz
, src0_swiz
,
795 src0_swiz
, src0_swiz
);
796 for (unsigned j
= i
+ 1; j
< 4; j
++) {
797 /* If there is another enabled component in the destination that is
798 * derived from the same inputs, generate its value on this pass as
801 if (!(done_mask
& (1 << j
)) &&
802 GET_SWZ(src0
.swizzle
, j
) == src0_swiz
) {
803 this_mask
|= (1 << j
);
807 if (this_mask
!= scs_mask
) {
808 glsl_to_tgsi_instruction
*inst
;
809 st_dst_reg tmp_dst
= st_dst_reg(tmp
);
811 /* Emit the SCS instruction.
813 inst
= emit(ir
, TGSI_OPCODE_SCS
, tmp_dst
, src0
);
814 inst
->dst
.writemask
= scs_mask
;
816 /* Move the result of the SCS instruction to the desired location in
819 tmp
.swizzle
= MAKE_SWIZZLE4(component
, component
,
820 component
, component
);
821 inst
= emit(ir
, TGSI_OPCODE_SCS
, dst
, tmp
);
822 inst
->dst
.writemask
= this_mask
;
824 /* Emit the SCS instruction to write directly to the destination.
826 glsl_to_tgsi_instruction
*inst
= emit(ir
, TGSI_OPCODE_SCS
, dst
, src0
);
827 inst
->dst
.writemask
= scs_mask
;
830 done_mask
|= this_mask
;
835 glsl_to_tgsi_visitor::add_constant(gl_register_file file
,
836 gl_constant_value values
[4], int size
, int datatype
,
839 if (file
== PROGRAM_CONSTANT
) {
840 return _mesa_add_typed_unnamed_constant(this->prog
->Parameters
, values
,
841 size
, datatype
, swizzle_out
);
844 immediate_storage
*entry
;
845 assert(file
== PROGRAM_IMMEDIATE
);
847 /* Search immediate storage to see if we already have an identical
848 * immediate that we can use instead of adding a duplicate entry.
850 foreach_iter(exec_list_iterator
, iter
, this->immediates
) {
851 entry
= (immediate_storage
*)iter
.get();
853 if (entry
->size
== size
&&
854 entry
->type
== datatype
&&
855 !memcmp(entry
->values
, values
, size
* sizeof(gl_constant_value
))) {
861 /* Add this immediate to the list. */
862 entry
= new(mem_ctx
) immediate_storage(values
, size
, datatype
);
863 this->immediates
.push_tail(entry
);
864 this->num_immediates
++;
870 glsl_to_tgsi_visitor::st_src_reg_for_float(float val
)
872 st_src_reg
src(PROGRAM_IMMEDIATE
, -1, GLSL_TYPE_FLOAT
);
873 union gl_constant_value uval
;
876 src
.index
= add_constant(src
.file
, &uval
, 1, GL_FLOAT
, &src
.swizzle
);
882 glsl_to_tgsi_visitor::st_src_reg_for_int(int val
)
884 st_src_reg
src(PROGRAM_IMMEDIATE
, -1, GLSL_TYPE_INT
);
885 union gl_constant_value uval
;
887 assert(native_integers
);
890 src
.index
= add_constant(src
.file
, &uval
, 1, GL_INT
, &src
.swizzle
);
896 glsl_to_tgsi_visitor::st_src_reg_for_type(int type
, int val
)
899 return type
== GLSL_TYPE_FLOAT
? st_src_reg_for_float(val
) :
900 st_src_reg_for_int(val
);
902 return st_src_reg_for_float(val
);
906 type_size(const struct glsl_type
*type
)
911 switch (type
->base_type
) {
914 case GLSL_TYPE_FLOAT
:
916 if (type
->is_matrix()) {
917 return type
->matrix_columns
;
919 /* Regardless of size of vector, it gets a vec4. This is bad
920 * packing for things like floats, but otherwise arrays become a
921 * mess. Hopefully a later pass over the code can pack scalars
922 * down if appropriate.
926 case GLSL_TYPE_ARRAY
:
927 assert(type
->length
> 0);
928 return type_size(type
->fields
.array
) * type
->length
;
929 case GLSL_TYPE_STRUCT
:
931 for (i
= 0; i
< type
->length
; i
++) {
932 size
+= type_size(type
->fields
.structure
[i
].type
);
935 case GLSL_TYPE_SAMPLER
:
936 /* Samplers take up one slot in UNIFORMS[], but they're baked in
947 * In the initial pass of codegen, we assign temporary numbers to
948 * intermediate results. (not SSA -- variable assignments will reuse
952 glsl_to_tgsi_visitor::get_temp(const glsl_type
*type
)
956 src
.type
= native_integers
? type
->base_type
: GLSL_TYPE_FLOAT
;
957 src
.file
= PROGRAM_TEMPORARY
;
958 src
.index
= next_temp
;
960 next_temp
+= type_size(type
);
962 if (type
->is_array() || type
->is_record()) {
963 src
.swizzle
= SWIZZLE_NOOP
;
965 src
.swizzle
= swizzle_for_size(type
->vector_elements
);
973 glsl_to_tgsi_visitor::find_variable_storage(ir_variable
*var
)
976 variable_storage
*entry
;
978 foreach_iter(exec_list_iterator
, iter
, this->variables
) {
979 entry
= (variable_storage
*)iter
.get();
981 if (entry
->var
== var
)
989 glsl_to_tgsi_visitor::visit(ir_variable
*ir
)
991 if (strcmp(ir
->name
, "gl_FragCoord") == 0) {
992 struct gl_fragment_program
*fp
= (struct gl_fragment_program
*)this->prog
;
994 fp
->OriginUpperLeft
= ir
->origin_upper_left
;
995 fp
->PixelCenterInteger
= ir
->pixel_center_integer
;
997 } else if (strcmp(ir
->name
, "gl_FragDepth") == 0) {
998 struct gl_fragment_program
*fp
= (struct gl_fragment_program
*)this->prog
;
999 switch (ir
->depth_layout
) {
1000 case ir_depth_layout_none
:
1001 fp
->FragDepthLayout
= FRAG_DEPTH_LAYOUT_NONE
;
1003 case ir_depth_layout_any
:
1004 fp
->FragDepthLayout
= FRAG_DEPTH_LAYOUT_ANY
;
1006 case ir_depth_layout_greater
:
1007 fp
->FragDepthLayout
= FRAG_DEPTH_LAYOUT_GREATER
;
1009 case ir_depth_layout_less
:
1010 fp
->FragDepthLayout
= FRAG_DEPTH_LAYOUT_LESS
;
1012 case ir_depth_layout_unchanged
:
1013 fp
->FragDepthLayout
= FRAG_DEPTH_LAYOUT_UNCHANGED
;
1021 if (ir
->mode
== ir_var_uniform
&& strncmp(ir
->name
, "gl_", 3) == 0) {
1023 const ir_state_slot
*const slots
= ir
->state_slots
;
1024 assert(ir
->state_slots
!= NULL
);
1026 /* Check if this statevar's setup in the STATE file exactly
1027 * matches how we'll want to reference it as a
1028 * struct/array/whatever. If not, then we need to move it into
1029 * temporary storage and hope that it'll get copy-propagated
1032 for (i
= 0; i
< ir
->num_state_slots
; i
++) {
1033 if (slots
[i
].swizzle
!= SWIZZLE_XYZW
) {
1038 variable_storage
*storage
;
1040 if (i
== ir
->num_state_slots
) {
1041 /* We'll set the index later. */
1042 storage
= new(mem_ctx
) variable_storage(ir
, PROGRAM_STATE_VAR
, -1);
1043 this->variables
.push_tail(storage
);
1047 /* The variable_storage constructor allocates slots based on the size
1048 * of the type. However, this had better match the number of state
1049 * elements that we're going to copy into the new temporary.
1051 assert((int) ir
->num_state_slots
== type_size(ir
->type
));
1053 storage
= new(mem_ctx
) variable_storage(ir
, PROGRAM_TEMPORARY
,
1055 this->variables
.push_tail(storage
);
1056 this->next_temp
+= type_size(ir
->type
);
1058 dst
= st_dst_reg(st_src_reg(PROGRAM_TEMPORARY
, storage
->index
,
1059 native_integers
? ir
->type
->base_type
: GLSL_TYPE_FLOAT
));
1063 for (unsigned int i
= 0; i
< ir
->num_state_slots
; i
++) {
1064 int index
= _mesa_add_state_reference(this->prog
->Parameters
,
1065 (gl_state_index
*)slots
[i
].tokens
);
1067 if (storage
->file
== PROGRAM_STATE_VAR
) {
1068 if (storage
->index
== -1) {
1069 storage
->index
= index
;
1071 assert(index
== storage
->index
+ (int)i
);
1074 st_src_reg
src(PROGRAM_STATE_VAR
, index
,
1075 native_integers
? ir
->type
->base_type
: GLSL_TYPE_FLOAT
);
1076 src
.swizzle
= slots
[i
].swizzle
;
1077 emit(ir
, TGSI_OPCODE_MOV
, dst
, src
);
1078 /* even a float takes up a whole vec4 reg in a struct/array. */
1083 if (storage
->file
== PROGRAM_TEMPORARY
&&
1084 dst
.index
!= storage
->index
+ (int) ir
->num_state_slots
) {
1085 fail_link(this->shader_program
,
1086 "failed to load builtin uniform `%s' (%d/%d regs loaded)\n",
1087 ir
->name
, dst
.index
- storage
->index
,
1088 type_size(ir
->type
));
1094 glsl_to_tgsi_visitor::visit(ir_loop
*ir
)
1096 ir_dereference_variable
*counter
= NULL
;
1098 if (ir
->counter
!= NULL
)
1099 counter
= new(ir
) ir_dereference_variable(ir
->counter
);
1101 if (ir
->from
!= NULL
) {
1102 assert(ir
->counter
!= NULL
);
1104 ir_assignment
*a
= new(ir
) ir_assignment(counter
, ir
->from
, NULL
);
1110 emit(NULL
, TGSI_OPCODE_BGNLOOP
);
1114 new(ir
) ir_expression(ir
->cmp
, glsl_type::bool_type
,
1116 ir_if
*if_stmt
= new(ir
) ir_if(e
);
1118 ir_loop_jump
*brk
= new(ir
) ir_loop_jump(ir_loop_jump::jump_break
);
1120 if_stmt
->then_instructions
.push_tail(brk
);
1122 if_stmt
->accept(this);
1129 visit_exec_list(&ir
->body_instructions
, this);
1131 if (ir
->increment
) {
1133 new(ir
) ir_expression(ir_binop_add
, counter
->type
,
1134 counter
, ir
->increment
);
1136 ir_assignment
*a
= new(ir
) ir_assignment(counter
, e
, NULL
);
1143 emit(NULL
, TGSI_OPCODE_ENDLOOP
);
1147 glsl_to_tgsi_visitor::visit(ir_loop_jump
*ir
)
1150 case ir_loop_jump::jump_break
:
1151 emit(NULL
, TGSI_OPCODE_BRK
);
1153 case ir_loop_jump::jump_continue
:
1154 emit(NULL
, TGSI_OPCODE_CONT
);
1161 glsl_to_tgsi_visitor::visit(ir_function_signature
*ir
)
1168 glsl_to_tgsi_visitor::visit(ir_function
*ir
)
1170 /* Ignore function bodies other than main() -- we shouldn't see calls to
1171 * them since they should all be inlined before we get to glsl_to_tgsi.
1173 if (strcmp(ir
->name
, "main") == 0) {
1174 const ir_function_signature
*sig
;
1177 sig
= ir
->matching_signature(&empty
);
1181 foreach_iter(exec_list_iterator
, iter
, sig
->body
) {
1182 ir_instruction
*ir
= (ir_instruction
*)iter
.get();
1190 glsl_to_tgsi_visitor::try_emit_mad(ir_expression
*ir
, int mul_operand
)
1192 int nonmul_operand
= 1 - mul_operand
;
1194 st_dst_reg result_dst
;
1196 ir_expression
*expr
= ir
->operands
[mul_operand
]->as_expression();
1197 if (!expr
|| expr
->operation
!= ir_binop_mul
)
1200 expr
->operands
[0]->accept(this);
1202 expr
->operands
[1]->accept(this);
1204 ir
->operands
[nonmul_operand
]->accept(this);
1207 this->result
= get_temp(ir
->type
);
1208 result_dst
= st_dst_reg(this->result
);
1209 result_dst
.writemask
= (1 << ir
->type
->vector_elements
) - 1;
1210 emit(ir
, TGSI_OPCODE_MAD
, result_dst
, a
, b
, c
);
1216 * Emit MAD(a, -b, a) instead of AND(a, NOT(b))
1218 * The logic values are 1.0 for true and 0.0 for false. Logical-and is
1219 * implemented using multiplication, and logical-or is implemented using
1220 * addition. Logical-not can be implemented as (true - x), or (1.0 - x).
1221 * As result, the logical expression (a & !b) can be rewritten as:
1225 * - (a * 1) - (a * b)
1229 * This final expression can be implemented as a single MAD(a, -b, a)
1233 glsl_to_tgsi_visitor::try_emit_mad_for_and_not(ir_expression
*ir
, int try_operand
)
1235 const int other_operand
= 1 - try_operand
;
1238 ir_expression
*expr
= ir
->operands
[try_operand
]->as_expression();
1239 if (!expr
|| expr
->operation
!= ir_unop_logic_not
)
1242 ir
->operands
[other_operand
]->accept(this);
1244 expr
->operands
[0]->accept(this);
1247 b
.negate
= ~b
.negate
;
1249 this->result
= get_temp(ir
->type
);
1250 emit(ir
, TGSI_OPCODE_MAD
, st_dst_reg(this->result
), a
, b
, a
);
1256 glsl_to_tgsi_visitor::try_emit_sat(ir_expression
*ir
)
1258 /* Saturates were only introduced to vertex programs in
1259 * NV_vertex_program3, so don't give them to drivers in the VP.
1261 if (this->prog
->Target
== GL_VERTEX_PROGRAM_ARB
)
1264 ir_rvalue
*sat_src
= ir
->as_rvalue_to_saturate();
1268 sat_src
->accept(this);
1269 st_src_reg src
= this->result
;
1271 /* If we generated an expression instruction into a temporary in
1272 * processing the saturate's operand, apply the saturate to that
1273 * instruction. Otherwise, generate a MOV to do the saturate.
1275 * Note that we have to be careful to only do this optimization if
1276 * the instruction in question was what generated src->result. For
1277 * example, ir_dereference_array might generate a MUL instruction
1278 * to create the reladdr, and return us a src reg using that
1279 * reladdr. That MUL result is not the value we're trying to
1282 ir_expression
*sat_src_expr
= sat_src
->as_expression();
1283 if (sat_src_expr
&& (sat_src_expr
->operation
== ir_binop_mul
||
1284 sat_src_expr
->operation
== ir_binop_add
||
1285 sat_src_expr
->operation
== ir_binop_dot
)) {
1286 glsl_to_tgsi_instruction
*new_inst
;
1287 new_inst
= (glsl_to_tgsi_instruction
*)this->instructions
.get_tail();
1288 new_inst
->saturate
= true;
1290 this->result
= get_temp(ir
->type
);
1291 st_dst_reg result_dst
= st_dst_reg(this->result
);
1292 result_dst
.writemask
= (1 << ir
->type
->vector_elements
) - 1;
1293 glsl_to_tgsi_instruction
*inst
;
1294 inst
= emit(ir
, TGSI_OPCODE_MOV
, result_dst
, src
);
1295 inst
->saturate
= true;
1302 glsl_to_tgsi_visitor::reladdr_to_temp(ir_instruction
*ir
,
1303 st_src_reg
*reg
, int *num_reladdr
)
1308 emit_arl(ir
, address_reg
, *reg
->reladdr
);
1310 if (*num_reladdr
!= 1) {
1311 st_src_reg temp
= get_temp(glsl_type::vec4_type
);
1313 emit(ir
, TGSI_OPCODE_MOV
, st_dst_reg(temp
), *reg
);
1321 glsl_to_tgsi_visitor::visit(ir_expression
*ir
)
1323 unsigned int operand
;
1324 st_src_reg op
[Elements(ir
->operands
)];
1325 st_src_reg result_src
;
1326 st_dst_reg result_dst
;
1328 /* Quick peephole: Emit MAD(a, b, c) instead of ADD(MUL(a, b), c)
1330 if (ir
->operation
== ir_binop_add
) {
1331 if (try_emit_mad(ir
, 1))
1333 if (try_emit_mad(ir
, 0))
1337 /* Quick peephole: Emit OPCODE_MAD(-a, -b, a) instead of AND(a, NOT(b))
1339 if (ir
->operation
== ir_binop_logic_and
) {
1340 if (try_emit_mad_for_and_not(ir
, 1))
1342 if (try_emit_mad_for_and_not(ir
, 0))
1346 if (try_emit_sat(ir
))
1349 if (ir
->operation
== ir_quadop_vector
)
1350 assert(!"ir_quadop_vector should have been lowered");
1352 for (operand
= 0; operand
< ir
->get_num_operands(); operand
++) {
1353 this->result
.file
= PROGRAM_UNDEFINED
;
1354 ir
->operands
[operand
]->accept(this);
1355 if (this->result
.file
== PROGRAM_UNDEFINED
) {
1357 printf("Failed to get tree for expression operand:\n");
1358 ir
->operands
[operand
]->accept(&v
);
1361 op
[operand
] = this->result
;
1363 /* Matrix expression operands should have been broken down to vector
1364 * operations already.
1366 assert(!ir
->operands
[operand
]->type
->is_matrix());
1369 int vector_elements
= ir
->operands
[0]->type
->vector_elements
;
1370 if (ir
->operands
[1]) {
1371 vector_elements
= MAX2(vector_elements
,
1372 ir
->operands
[1]->type
->vector_elements
);
1375 this->result
.file
= PROGRAM_UNDEFINED
;
1377 /* Storage for our result. Ideally for an assignment we'd be using
1378 * the actual storage for the result here, instead.
1380 result_src
= get_temp(ir
->type
);
1381 /* convenience for the emit functions below. */
1382 result_dst
= st_dst_reg(result_src
);
1383 /* Limit writes to the channels that will be used by result_src later.
1384 * This does limit this temp's use as a temporary for multi-instruction
1387 result_dst
.writemask
= (1 << ir
->type
->vector_elements
) - 1;
1389 switch (ir
->operation
) {
1390 case ir_unop_logic_not
:
1391 if (result_dst
.type
!= GLSL_TYPE_FLOAT
)
1392 emit(ir
, TGSI_OPCODE_SEQ
, result_dst
, op
[0], st_src_reg_for_type(result_dst
.type
, 0));
1394 /* Previously 'SEQ dst, src, 0.0' was used for this. However, many
1395 * older GPUs implement SEQ using multiple instructions (i915 uses two
1396 * SGE instructions and a MUL instruction). Since our logic values are
1397 * 0.0 and 1.0, 1-x also implements !x.
1399 op
[0].negate
= ~op
[0].negate
;
1400 emit(ir
, TGSI_OPCODE_ADD
, result_dst
, op
[0], st_src_reg_for_float(1.0));
1404 assert(result_dst
.type
== GLSL_TYPE_FLOAT
|| result_dst
.type
== GLSL_TYPE_INT
);
1405 if (result_dst
.type
== GLSL_TYPE_INT
)
1406 emit(ir
, TGSI_OPCODE_INEG
, result_dst
, op
[0]);
1408 op
[0].negate
= ~op
[0].negate
;
1413 assert(result_dst
.type
== GLSL_TYPE_FLOAT
);
1414 emit(ir
, TGSI_OPCODE_ABS
, result_dst
, op
[0]);
1417 emit(ir
, TGSI_OPCODE_SSG
, result_dst
, op
[0]);
1420 emit_scalar(ir
, TGSI_OPCODE_RCP
, result_dst
, op
[0]);
1424 emit_scalar(ir
, TGSI_OPCODE_EX2
, result_dst
, op
[0]);
1428 assert(!"not reached: should be handled by ir_explog_to_explog2");
1431 emit_scalar(ir
, TGSI_OPCODE_LG2
, result_dst
, op
[0]);
1434 emit_scalar(ir
, TGSI_OPCODE_SIN
, result_dst
, op
[0]);
1437 emit_scalar(ir
, TGSI_OPCODE_COS
, result_dst
, op
[0]);
1439 case ir_unop_sin_reduced
:
1440 emit_scs(ir
, TGSI_OPCODE_SIN
, result_dst
, op
[0]);
1442 case ir_unop_cos_reduced
:
1443 emit_scs(ir
, TGSI_OPCODE_COS
, result_dst
, op
[0]);
1447 emit(ir
, TGSI_OPCODE_DDX
, result_dst
, op
[0]);
1450 op
[0].negate
= ~op
[0].negate
;
1451 emit(ir
, TGSI_OPCODE_DDY
, result_dst
, op
[0]);
1454 case ir_unop_noise
: {
1455 /* At some point, a motivated person could add a better
1456 * implementation of noise. Currently not even the nvidia
1457 * binary drivers do anything more than this. In any case, the
1458 * place to do this is in the GL state tracker, not the poor
1461 emit(ir
, TGSI_OPCODE_MOV
, result_dst
, st_src_reg_for_float(0.5));
1466 emit(ir
, TGSI_OPCODE_ADD
, result_dst
, op
[0], op
[1]);
1469 emit(ir
, TGSI_OPCODE_SUB
, result_dst
, op
[0], op
[1]);
1473 emit(ir
, TGSI_OPCODE_MUL
, result_dst
, op
[0], op
[1]);
1476 if (result_dst
.type
== GLSL_TYPE_FLOAT
)
1477 assert(!"not reached: should be handled by ir_div_to_mul_rcp");
1479 emit(ir
, TGSI_OPCODE_DIV
, result_dst
, op
[0], op
[1]);
1482 if (result_dst
.type
== GLSL_TYPE_FLOAT
)
1483 assert(!"ir_binop_mod should have been converted to b * fract(a/b)");
1485 emit(ir
, TGSI_OPCODE_MOD
, result_dst
, op
[0], op
[1]);
1489 emit(ir
, TGSI_OPCODE_SLT
, result_dst
, op
[0], op
[1]);
1491 case ir_binop_greater
:
1492 emit(ir
, TGSI_OPCODE_SGT
, result_dst
, op
[0], op
[1]);
1494 case ir_binop_lequal
:
1495 emit(ir
, TGSI_OPCODE_SLE
, result_dst
, op
[0], op
[1]);
1497 case ir_binop_gequal
:
1498 emit(ir
, TGSI_OPCODE_SGE
, result_dst
, op
[0], op
[1]);
1500 case ir_binop_equal
:
1501 emit(ir
, TGSI_OPCODE_SEQ
, result_dst
, op
[0], op
[1]);
1503 case ir_binop_nequal
:
1504 emit(ir
, TGSI_OPCODE_SNE
, result_dst
, op
[0], op
[1]);
1506 case ir_binop_all_equal
:
1507 /* "==" operator producing a scalar boolean. */
1508 if (ir
->operands
[0]->type
->is_vector() ||
1509 ir
->operands
[1]->type
->is_vector()) {
1510 st_src_reg temp
= get_temp(native_integers
?
1511 glsl_type::get_instance(ir
->operands
[0]->type
->base_type
, 4, 1) :
1512 glsl_type::vec4_type
);
1513 assert(ir
->operands
[0]->type
->base_type
== GLSL_TYPE_FLOAT
);
1514 emit(ir
, TGSI_OPCODE_SNE
, st_dst_reg(temp
), op
[0], op
[1]);
1516 /* After the dot-product, the value will be an integer on the
1517 * range [0,4]. Zero becomes 1.0, and positive values become zero.
1519 emit_dp(ir
, result_dst
, temp
, temp
, vector_elements
);
1521 if (result_dst
.type
== GLSL_TYPE_FLOAT
) {
1522 /* Negating the result of the dot-product gives values on the range
1523 * [-4, 0]. Zero becomes 1.0, and negative values become zero.
1524 * This is achieved using SGE.
1526 st_src_reg sge_src
= result_src
;
1527 sge_src
.negate
= ~sge_src
.negate
;
1528 emit(ir
, TGSI_OPCODE_SGE
, result_dst
, sge_src
, st_src_reg_for_float(0.0));
1530 /* The TGSI negate flag doesn't work for integers, so use SEQ 0
1533 emit(ir
, TGSI_OPCODE_SEQ
, result_dst
, result_src
, st_src_reg_for_int(0));
1536 emit(ir
, TGSI_OPCODE_SEQ
, result_dst
, op
[0], op
[1]);
1539 case ir_binop_any_nequal
:
1540 /* "!=" operator producing a scalar boolean. */
1541 if (ir
->operands
[0]->type
->is_vector() ||
1542 ir
->operands
[1]->type
->is_vector()) {
1543 st_src_reg temp
= get_temp(native_integers
?
1544 glsl_type::get_instance(ir
->operands
[0]->type
->base_type
, 4, 1) :
1545 glsl_type::vec4_type
);
1546 assert(ir
->operands
[0]->type
->base_type
== GLSL_TYPE_FLOAT
);
1547 emit(ir
, TGSI_OPCODE_SNE
, st_dst_reg(temp
), op
[0], op
[1]);
1549 /* After the dot-product, the value will be an integer on the
1550 * range [0,4]. Zero stays zero, and positive values become 1.0.
1552 glsl_to_tgsi_instruction
*const dp
=
1553 emit_dp(ir
, result_dst
, temp
, temp
, vector_elements
);
1554 if (this->prog
->Target
== GL_FRAGMENT_PROGRAM_ARB
&&
1555 result_dst
.type
== GLSL_TYPE_FLOAT
) {
1556 /* The clamping to [0,1] can be done for free in the fragment
1557 * shader with a saturate.
1559 dp
->saturate
= true;
1560 } else if (result_dst
.type
== GLSL_TYPE_FLOAT
) {
1561 /* Negating the result of the dot-product gives values on the range
1562 * [-4, 0]. Zero stays zero, and negative values become 1.0. This
1563 * achieved using SLT.
1565 st_src_reg slt_src
= result_src
;
1566 slt_src
.negate
= ~slt_src
.negate
;
1567 emit(ir
, TGSI_OPCODE_SLT
, result_dst
, slt_src
, st_src_reg_for_float(0.0));
1569 emit(ir
, TGSI_OPCODE_SNE
, result_dst
, result_src
, st_src_reg_for_int(0));
1572 emit(ir
, TGSI_OPCODE_SNE
, result_dst
, op
[0], op
[1]);
1577 assert(ir
->operands
[0]->type
->is_vector());
1579 /* After the dot-product, the value will be an integer on the
1580 * range [0,4]. Zero stays zero, and positive values become 1.0.
1582 glsl_to_tgsi_instruction
*const dp
=
1583 emit_dp(ir
, result_dst
, op
[0], op
[0],
1584 ir
->operands
[0]->type
->vector_elements
);
1585 if (this->prog
->Target
== GL_FRAGMENT_PROGRAM_ARB
&&
1586 result_dst
.type
== GLSL_TYPE_FLOAT
) {
1587 /* The clamping to [0,1] can be done for free in the fragment
1588 * shader with a saturate.
1590 dp
->saturate
= true;
1591 } else if (result_dst
.type
== GLSL_TYPE_FLOAT
) {
1592 /* Negating the result of the dot-product gives values on the range
1593 * [-4, 0]. Zero stays zero, and negative values become 1.0. This
1594 * is achieved using SLT.
1596 st_src_reg slt_src
= result_src
;
1597 slt_src
.negate
= ~slt_src
.negate
;
1598 emit(ir
, TGSI_OPCODE_SLT
, result_dst
, slt_src
, st_src_reg_for_float(0.0));
1601 /* Use SNE 0 if integers are being used as boolean values. */
1602 emit(ir
, TGSI_OPCODE_SNE
, result_dst
, result_src
, st_src_reg_for_int(0));
1607 case ir_binop_logic_xor
:
1608 emit(ir
, TGSI_OPCODE_SNE
, result_dst
, op
[0], op
[1]);
1611 case ir_binop_logic_or
: {
1612 /* After the addition, the value will be an integer on the
1613 * range [0,2]. Zero stays zero, and positive values become 1.0.
1615 glsl_to_tgsi_instruction
*add
=
1616 emit(ir
, TGSI_OPCODE_ADD
, result_dst
, op
[0], op
[1]);
1617 if (this->prog
->Target
== GL_FRAGMENT_PROGRAM_ARB
&&
1618 result_dst
.type
== GLSL_TYPE_FLOAT
) {
1619 /* The clamping to [0,1] can be done for free in the fragment
1620 * shader with a saturate if floats are being used as boolean values.
1622 add
->saturate
= true;
1623 } else if (result_dst
.type
== GLSL_TYPE_FLOAT
) {
1624 /* Negating the result of the addition gives values on the range
1625 * [-2, 0]. Zero stays zero, and negative values become 1.0. This
1626 * is achieved using SLT.
1628 st_src_reg slt_src
= result_src
;
1629 slt_src
.negate
= ~slt_src
.negate
;
1630 emit(ir
, TGSI_OPCODE_SLT
, result_dst
, slt_src
, st_src_reg_for_float(0.0));
1632 /* Use an SNE on the result of the addition. Zero stays zero,
1633 * 1 stays 1, and 2 becomes 1.
1635 emit(ir
, TGSI_OPCODE_SNE
, result_dst
, result_src
, st_src_reg_for_int(0));
1640 case ir_binop_logic_and
:
1641 /* the bool args are stored as float 0.0 or 1.0, so "mul" gives us "and". */
1642 emit(ir
, TGSI_OPCODE_MUL
, result_dst
, op
[0], op
[1]);
1646 assert(ir
->operands
[0]->type
->is_vector());
1647 assert(ir
->operands
[0]->type
== ir
->operands
[1]->type
);
1648 emit_dp(ir
, result_dst
, op
[0], op
[1],
1649 ir
->operands
[0]->type
->vector_elements
);
1653 /* sqrt(x) = x * rsq(x). */
1654 emit_scalar(ir
, TGSI_OPCODE_RSQ
, result_dst
, op
[0]);
1655 emit(ir
, TGSI_OPCODE_MUL
, result_dst
, result_src
, op
[0]);
1656 /* For incoming channels <= 0, set the result to 0. */
1657 op
[0].negate
= ~op
[0].negate
;
1658 emit(ir
, TGSI_OPCODE_CMP
, result_dst
,
1659 op
[0], result_src
, st_src_reg_for_float(0.0));
1662 emit_scalar(ir
, TGSI_OPCODE_RSQ
, result_dst
, op
[0]);
1666 if (native_integers
) {
1667 emit(ir
, TGSI_OPCODE_I2F
, result_dst
, op
[0]);
1672 /* Converting between signed and unsigned integers is a no-op. */
1674 /* Booleans are stored as integers (or floats in GLSL 1.20 and lower). */
1678 if (native_integers
)
1679 emit(ir
, TGSI_OPCODE_F2I
, result_dst
, op
[0]);
1681 emit(ir
, TGSI_OPCODE_TRUNC
, result_dst
, op
[0]);
1685 emit(ir
, TGSI_OPCODE_SNE
, result_dst
, op
[0],
1686 st_src_reg_for_type(result_dst
.type
, 0));
1689 emit(ir
, TGSI_OPCODE_TRUNC
, result_dst
, op
[0]);
1692 op
[0].negate
= ~op
[0].negate
;
1693 emit(ir
, TGSI_OPCODE_FLR
, result_dst
, op
[0]);
1694 result_src
.negate
= ~result_src
.negate
;
1697 emit(ir
, TGSI_OPCODE_FLR
, result_dst
, op
[0]);
1700 emit(ir
, TGSI_OPCODE_FRC
, result_dst
, op
[0]);
1704 emit(ir
, TGSI_OPCODE_MIN
, result_dst
, op
[0], op
[1]);
1707 emit(ir
, TGSI_OPCODE_MAX
, result_dst
, op
[0], op
[1]);
1710 emit_scalar(ir
, TGSI_OPCODE_POW
, result_dst
, op
[0], op
[1]);
1713 case ir_unop_bit_not
:
1714 if (glsl_version
>= 130) {
1715 emit(ir
, TGSI_OPCODE_NOT
, result_dst
, op
[0]);
1719 if (native_integers
) {
1720 emit(ir
, TGSI_OPCODE_U2F
, result_dst
, op
[0]);
1723 case ir_binop_lshift
:
1724 if (glsl_version
>= 130) {
1725 emit(ir
, TGSI_OPCODE_SHL
, result_dst
, op
[0]);
1728 case ir_binop_rshift
:
1729 if (glsl_version
>= 130) {
1730 emit(ir
, TGSI_OPCODE_ISHR
, result_dst
, op
[0]);
1733 case ir_binop_bit_and
:
1734 if (glsl_version
>= 130) {
1735 emit(ir
, TGSI_OPCODE_AND
, result_dst
, op
[0]);
1738 case ir_binop_bit_xor
:
1739 if (glsl_version
>= 130) {
1740 emit(ir
, TGSI_OPCODE_XOR
, result_dst
, op
[0]);
1743 case ir_binop_bit_or
:
1744 if (glsl_version
>= 130) {
1745 emit(ir
, TGSI_OPCODE_OR
, result_dst
, op
[0]);
1748 case ir_unop_round_even
:
1749 assert(!"GLSL 1.30 features unsupported");
1752 case ir_quadop_vector
:
1753 /* This operation should have already been handled.
1755 assert(!"Should not get here.");
1759 this->result
= result_src
;
1764 glsl_to_tgsi_visitor::visit(ir_swizzle
*ir
)
1770 /* Note that this is only swizzles in expressions, not those on the left
1771 * hand side of an assignment, which do write masking. See ir_assignment
1775 ir
->val
->accept(this);
1777 assert(src
.file
!= PROGRAM_UNDEFINED
);
1779 for (i
= 0; i
< 4; i
++) {
1780 if (i
< ir
->type
->vector_elements
) {
1783 swizzle
[i
] = GET_SWZ(src
.swizzle
, ir
->mask
.x
);
1786 swizzle
[i
] = GET_SWZ(src
.swizzle
, ir
->mask
.y
);
1789 swizzle
[i
] = GET_SWZ(src
.swizzle
, ir
->mask
.z
);
1792 swizzle
[i
] = GET_SWZ(src
.swizzle
, ir
->mask
.w
);
1796 /* If the type is smaller than a vec4, replicate the last
1799 swizzle
[i
] = swizzle
[ir
->type
->vector_elements
- 1];
1803 src
.swizzle
= MAKE_SWIZZLE4(swizzle
[0], swizzle
[1], swizzle
[2], swizzle
[3]);
1809 glsl_to_tgsi_visitor::visit(ir_dereference_variable
*ir
)
1811 variable_storage
*entry
= find_variable_storage(ir
->var
);
1812 ir_variable
*var
= ir
->var
;
1815 switch (var
->mode
) {
1816 case ir_var_uniform
:
1817 entry
= new(mem_ctx
) variable_storage(var
, PROGRAM_UNIFORM
,
1819 this->variables
.push_tail(entry
);
1823 /* The linker assigns locations for varyings and attributes,
1824 * including deprecated builtins (like gl_Color), user-assign
1825 * generic attributes (glBindVertexLocation), and
1826 * user-defined varyings.
1828 * FINISHME: We would hit this path for function arguments. Fix!
1830 assert(var
->location
!= -1);
1831 entry
= new(mem_ctx
) variable_storage(var
,
1834 if (this->prog
->Target
== GL_VERTEX_PROGRAM_ARB
&&
1835 var
->location
>= VERT_ATTRIB_GENERIC0
) {
1836 _mesa_add_attribute(this->prog
->Attributes
,
1838 _mesa_sizeof_glsl_type(var
->type
->gl_type
),
1840 var
->location
- VERT_ATTRIB_GENERIC0
);
1844 assert(var
->location
!= -1);
1845 entry
= new(mem_ctx
) variable_storage(var
,
1849 case ir_var_system_value
:
1850 entry
= new(mem_ctx
) variable_storage(var
,
1851 PROGRAM_SYSTEM_VALUE
,
1855 case ir_var_temporary
:
1856 entry
= new(mem_ctx
) variable_storage(var
, PROGRAM_TEMPORARY
,
1858 this->variables
.push_tail(entry
);
1860 next_temp
+= type_size(var
->type
);
1865 printf("Failed to make storage for %s\n", var
->name
);
1870 this->result
= st_src_reg(entry
->file
, entry
->index
, var
->type
);
1871 if (!native_integers
)
1872 this->result
.type
= GLSL_TYPE_FLOAT
;
1876 glsl_to_tgsi_visitor::visit(ir_dereference_array
*ir
)
1880 int element_size
= type_size(ir
->type
);
1882 index
= ir
->array_index
->constant_expression_value();
1884 ir
->array
->accept(this);
1888 src
.index
+= index
->value
.i
[0] * element_size
;
1890 /* Variable index array dereference. It eats the "vec4" of the
1891 * base of the array and an index that offsets the TGSI register
1894 ir
->array_index
->accept(this);
1896 st_src_reg index_reg
;
1898 if (element_size
== 1) {
1899 index_reg
= this->result
;
1901 index_reg
= get_temp(glsl_type::float_type
);
1903 emit(ir
, TGSI_OPCODE_MUL
, st_dst_reg(index_reg
),
1904 this->result
, st_src_reg_for_float(element_size
));
1907 /* If there was already a relative address register involved, add the
1908 * new and the old together to get the new offset.
1910 if (src
.reladdr
!= NULL
) {
1911 st_src_reg accum_reg
= get_temp(glsl_type::float_type
);
1913 emit(ir
, TGSI_OPCODE_ADD
, st_dst_reg(accum_reg
),
1914 index_reg
, *src
.reladdr
);
1916 index_reg
= accum_reg
;
1919 src
.reladdr
= ralloc(mem_ctx
, st_src_reg
);
1920 memcpy(src
.reladdr
, &index_reg
, sizeof(index_reg
));
1923 /* If the type is smaller than a vec4, replicate the last channel out. */
1924 if (ir
->type
->is_scalar() || ir
->type
->is_vector())
1925 src
.swizzle
= swizzle_for_size(ir
->type
->vector_elements
);
1927 src
.swizzle
= SWIZZLE_NOOP
;
1933 glsl_to_tgsi_visitor::visit(ir_dereference_record
*ir
)
1936 const glsl_type
*struct_type
= ir
->record
->type
;
1939 ir
->record
->accept(this);
1941 for (i
= 0; i
< struct_type
->length
; i
++) {
1942 if (strcmp(struct_type
->fields
.structure
[i
].name
, ir
->field
) == 0)
1944 offset
+= type_size(struct_type
->fields
.structure
[i
].type
);
1947 /* If the type is smaller than a vec4, replicate the last channel out. */
1948 if (ir
->type
->is_scalar() || ir
->type
->is_vector())
1949 this->result
.swizzle
= swizzle_for_size(ir
->type
->vector_elements
);
1951 this->result
.swizzle
= SWIZZLE_NOOP
;
1953 this->result
.index
+= offset
;
1957 * We want to be careful in assignment setup to hit the actual storage
1958 * instead of potentially using a temporary like we might with the
1959 * ir_dereference handler.
1962 get_assignment_lhs(ir_dereference
*ir
, glsl_to_tgsi_visitor
*v
)
1964 /* The LHS must be a dereference. If the LHS is a variable indexed array
1965 * access of a vector, it must be separated into a series conditional moves
1966 * before reaching this point (see ir_vec_index_to_cond_assign).
1968 assert(ir
->as_dereference());
1969 ir_dereference_array
*deref_array
= ir
->as_dereference_array();
1971 assert(!deref_array
->array
->type
->is_vector());
1974 /* Use the rvalue deref handler for the most part. We'll ignore
1975 * swizzles in it and write swizzles using writemask, though.
1978 return st_dst_reg(v
->result
);
1982 * Process the condition of a conditional assignment
1984 * Examines the condition of a conditional assignment to generate the optimal
1985 * first operand of a \c CMP instruction. If the condition is a relational
1986 * operator with 0 (e.g., \c ir_binop_less), the value being compared will be
1987 * used as the source for the \c CMP instruction. Otherwise the comparison
1988 * is processed to a boolean result, and the boolean result is used as the
1989 * operand to the CMP instruction.
1992 glsl_to_tgsi_visitor::process_move_condition(ir_rvalue
*ir
)
1994 ir_rvalue
*src_ir
= ir
;
1996 bool switch_order
= false;
1998 ir_expression
*const expr
= ir
->as_expression();
1999 if ((expr
!= NULL
) && (expr
->get_num_operands() == 2)) {
2000 bool zero_on_left
= false;
2002 if (expr
->operands
[0]->is_zero()) {
2003 src_ir
= expr
->operands
[1];
2004 zero_on_left
= true;
2005 } else if (expr
->operands
[1]->is_zero()) {
2006 src_ir
= expr
->operands
[0];
2007 zero_on_left
= false;
2011 * (a < 0) T F F ( a < 0) T F F
2012 * (0 < a) F F T (-a < 0) F F T
2013 * (a <= 0) T T F (-a < 0) F F T (swap order of other operands)
2014 * (0 <= a) F T T ( a < 0) T F F (swap order of other operands)
2015 * (a > 0) F F T (-a < 0) F F T
2016 * (0 > a) T F F ( a < 0) T F F
2017 * (a >= 0) F T T ( a < 0) T F F (swap order of other operands)
2018 * (0 >= a) T T F (-a < 0) F F T (swap order of other operands)
2020 * Note that exchanging the order of 0 and 'a' in the comparison simply
2021 * means that the value of 'a' should be negated.
2024 switch (expr
->operation
) {
2026 switch_order
= false;
2027 negate
= zero_on_left
;
2030 case ir_binop_greater
:
2031 switch_order
= false;
2032 negate
= !zero_on_left
;
2035 case ir_binop_lequal
:
2036 switch_order
= true;
2037 negate
= !zero_on_left
;
2040 case ir_binop_gequal
:
2041 switch_order
= true;
2042 negate
= zero_on_left
;
2046 /* This isn't the right kind of comparison afterall, so make sure
2047 * the whole condition is visited.
2055 src_ir
->accept(this);
2057 /* We use the TGSI_OPCODE_CMP (a < 0 ? b : c) for conditional moves, and the
2058 * condition we produced is 0.0 or 1.0. By flipping the sign, we can
2059 * choose which value TGSI_OPCODE_CMP produces without an extra instruction
2060 * computing the condition.
2063 this->result
.negate
= ~this->result
.negate
;
2065 return switch_order
;
2069 glsl_to_tgsi_visitor::visit(ir_assignment
*ir
)
2075 ir
->rhs
->accept(this);
2078 l
= get_assignment_lhs(ir
->lhs
, this);
2080 /* FINISHME: This should really set to the correct maximal writemask for each
2081 * FINISHME: component written (in the loops below). This case can only
2082 * FINISHME: occur for matrices, arrays, and structures.
2084 if (ir
->write_mask
== 0) {
2085 assert(!ir
->lhs
->type
->is_scalar() && !ir
->lhs
->type
->is_vector());
2086 l
.writemask
= WRITEMASK_XYZW
;
2087 } else if (ir
->lhs
->type
->is_scalar() &&
2088 ir
->lhs
->variable_referenced()->mode
== ir_var_out
) {
2089 /* FINISHME: This hack makes writing to gl_FragDepth, which lives in the
2090 * FINISHME: W component of fragment shader output zero, work correctly.
2092 l
.writemask
= WRITEMASK_XYZW
;
2095 int first_enabled_chan
= 0;
2098 l
.writemask
= ir
->write_mask
;
2100 for (int i
= 0; i
< 4; i
++) {
2101 if (l
.writemask
& (1 << i
)) {
2102 first_enabled_chan
= GET_SWZ(r
.swizzle
, i
);
2107 /* Swizzle a small RHS vector into the channels being written.
2109 * glsl ir treats write_mask as dictating how many channels are
2110 * present on the RHS while TGSI treats write_mask as just
2111 * showing which channels of the vec4 RHS get written.
2113 for (int i
= 0; i
< 4; i
++) {
2114 if (l
.writemask
& (1 << i
))
2115 swizzles
[i
] = GET_SWZ(r
.swizzle
, rhs_chan
++);
2117 swizzles
[i
] = first_enabled_chan
;
2119 r
.swizzle
= MAKE_SWIZZLE4(swizzles
[0], swizzles
[1],
2120 swizzles
[2], swizzles
[3]);
2123 assert(l
.file
!= PROGRAM_UNDEFINED
);
2124 assert(r
.file
!= PROGRAM_UNDEFINED
);
2126 if (ir
->condition
) {
2127 const bool switch_order
= this->process_move_condition(ir
->condition
);
2128 st_src_reg condition
= this->result
;
2130 for (i
= 0; i
< type_size(ir
->lhs
->type
); i
++) {
2131 st_src_reg l_src
= st_src_reg(l
);
2132 l_src
.swizzle
= swizzle_for_size(ir
->lhs
->type
->vector_elements
);
2135 emit(ir
, TGSI_OPCODE_CMP
, l
, condition
, l_src
, r
);
2137 emit(ir
, TGSI_OPCODE_CMP
, l
, condition
, r
, l_src
);
2143 } else if (ir
->rhs
->as_expression() &&
2144 this->instructions
.get_tail() &&
2145 ir
->rhs
== ((glsl_to_tgsi_instruction
*)this->instructions
.get_tail())->ir
&&
2146 type_size(ir
->lhs
->type
) == 1 &&
2147 l
.writemask
== ((glsl_to_tgsi_instruction
*)this->instructions
.get_tail())->dst
.writemask
) {
2148 /* To avoid emitting an extra MOV when assigning an expression to a
2149 * variable, emit the last instruction of the expression again, but
2150 * replace the destination register with the target of the assignment.
2151 * Dead code elimination will remove the original instruction.
2153 glsl_to_tgsi_instruction
*inst
, *new_inst
;
2154 inst
= (glsl_to_tgsi_instruction
*)this->instructions
.get_tail();
2155 new_inst
= emit(ir
, inst
->op
, l
, inst
->src
[0], inst
->src
[1], inst
->src
[2]);
2156 new_inst
->saturate
= inst
->saturate
;
2158 for (i
= 0; i
< type_size(ir
->lhs
->type
); i
++) {
2159 emit(ir
, TGSI_OPCODE_MOV
, l
, r
);
2168 glsl_to_tgsi_visitor::visit(ir_constant
*ir
)
2171 GLfloat stack_vals
[4] = { 0 };
2172 gl_constant_value
*values
= (gl_constant_value
*) stack_vals
;
2173 GLenum gl_type
= GL_NONE
;
2175 static int in_array
= 0;
2176 gl_register_file file
= in_array
? PROGRAM_CONSTANT
: PROGRAM_IMMEDIATE
;
2178 /* Unfortunately, 4 floats is all we can get into
2179 * _mesa_add_typed_unnamed_constant. So, make a temp to store an
2180 * aggregate constant and move each constant value into it. If we
2181 * get lucky, copy propagation will eliminate the extra moves.
2183 if (ir
->type
->base_type
== GLSL_TYPE_STRUCT
) {
2184 st_src_reg temp_base
= get_temp(ir
->type
);
2185 st_dst_reg temp
= st_dst_reg(temp_base
);
2187 foreach_iter(exec_list_iterator
, iter
, ir
->components
) {
2188 ir_constant
*field_value
= (ir_constant
*)iter
.get();
2189 int size
= type_size(field_value
->type
);
2193 field_value
->accept(this);
2196 for (i
= 0; i
< (unsigned int)size
; i
++) {
2197 emit(ir
, TGSI_OPCODE_MOV
, temp
, src
);
2203 this->result
= temp_base
;
2207 if (ir
->type
->is_array()) {
2208 st_src_reg temp_base
= get_temp(ir
->type
);
2209 st_dst_reg temp
= st_dst_reg(temp_base
);
2210 int size
= type_size(ir
->type
->fields
.array
);
2215 for (i
= 0; i
< ir
->type
->length
; i
++) {
2216 ir
->array_elements
[i
]->accept(this);
2218 for (int j
= 0; j
< size
; j
++) {
2219 emit(ir
, TGSI_OPCODE_MOV
, temp
, src
);
2225 this->result
= temp_base
;
2230 if (ir
->type
->is_matrix()) {
2231 st_src_reg mat
= get_temp(ir
->type
);
2232 st_dst_reg mat_column
= st_dst_reg(mat
);
2234 for (i
= 0; i
< ir
->type
->matrix_columns
; i
++) {
2235 assert(ir
->type
->base_type
== GLSL_TYPE_FLOAT
);
2236 values
= (gl_constant_value
*) &ir
->value
.f
[i
* ir
->type
->vector_elements
];
2238 src
= st_src_reg(file
, -1, ir
->type
->base_type
);
2239 src
.index
= add_constant(file
,
2241 ir
->type
->vector_elements
,
2244 emit(ir
, TGSI_OPCODE_MOV
, mat_column
, src
);
2253 switch (ir
->type
->base_type
) {
2254 case GLSL_TYPE_FLOAT
:
2256 for (i
= 0; i
< ir
->type
->vector_elements
; i
++) {
2257 values
[i
].f
= ir
->value
.f
[i
];
2260 case GLSL_TYPE_UINT
:
2261 gl_type
= native_integers
? GL_UNSIGNED_INT
: GL_FLOAT
;
2262 for (i
= 0; i
< ir
->type
->vector_elements
; i
++) {
2263 if (native_integers
)
2264 values
[i
].u
= ir
->value
.u
[i
];
2266 values
[i
].f
= ir
->value
.u
[i
];
2270 gl_type
= native_integers
? GL_INT
: GL_FLOAT
;
2271 for (i
= 0; i
< ir
->type
->vector_elements
; i
++) {
2272 if (native_integers
)
2273 values
[i
].i
= ir
->value
.i
[i
];
2275 values
[i
].f
= ir
->value
.i
[i
];
2278 case GLSL_TYPE_BOOL
:
2279 gl_type
= native_integers
? GL_BOOL
: GL_FLOAT
;
2280 for (i
= 0; i
< ir
->type
->vector_elements
; i
++) {
2281 if (native_integers
)
2282 values
[i
].b
= ir
->value
.b
[i
];
2284 values
[i
].f
= ir
->value
.b
[i
];
2288 assert(!"Non-float/uint/int/bool constant");
2291 this->result
= st_src_reg(file
, -1, ir
->type
);
2292 this->result
.index
= add_constant(file
,
2294 ir
->type
->vector_elements
,
2296 &this->result
.swizzle
);
2300 glsl_to_tgsi_visitor::get_function_signature(ir_function_signature
*sig
)
2302 function_entry
*entry
;
2304 foreach_iter(exec_list_iterator
, iter
, this->function_signatures
) {
2305 entry
= (function_entry
*)iter
.get();
2307 if (entry
->sig
== sig
)
2311 entry
= ralloc(mem_ctx
, function_entry
);
2313 entry
->sig_id
= this->next_signature_id
++;
2314 entry
->bgn_inst
= NULL
;
2316 /* Allocate storage for all the parameters. */
2317 foreach_iter(exec_list_iterator
, iter
, sig
->parameters
) {
2318 ir_variable
*param
= (ir_variable
*)iter
.get();
2319 variable_storage
*storage
;
2321 storage
= find_variable_storage(param
);
2324 storage
= new(mem_ctx
) variable_storage(param
, PROGRAM_TEMPORARY
,
2326 this->variables
.push_tail(storage
);
2328 this->next_temp
+= type_size(param
->type
);
2331 if (!sig
->return_type
->is_void()) {
2332 entry
->return_reg
= get_temp(sig
->return_type
);
2334 entry
->return_reg
= undef_src
;
2337 this->function_signatures
.push_tail(entry
);
2342 glsl_to_tgsi_visitor::visit(ir_call
*ir
)
2344 glsl_to_tgsi_instruction
*call_inst
;
2345 ir_function_signature
*sig
= ir
->get_callee();
2346 function_entry
*entry
= get_function_signature(sig
);
2349 /* Process in parameters. */
2350 exec_list_iterator sig_iter
= sig
->parameters
.iterator();
2351 foreach_iter(exec_list_iterator
, iter
, *ir
) {
2352 ir_rvalue
*param_rval
= (ir_rvalue
*)iter
.get();
2353 ir_variable
*param
= (ir_variable
*)sig_iter
.get();
2355 if (param
->mode
== ir_var_in
||
2356 param
->mode
== ir_var_inout
) {
2357 variable_storage
*storage
= find_variable_storage(param
);
2360 param_rval
->accept(this);
2361 st_src_reg r
= this->result
;
2364 l
.file
= storage
->file
;
2365 l
.index
= storage
->index
;
2367 l
.writemask
= WRITEMASK_XYZW
;
2368 l
.cond_mask
= COND_TR
;
2370 for (i
= 0; i
< type_size(param
->type
); i
++) {
2371 emit(ir
, TGSI_OPCODE_MOV
, l
, r
);
2379 assert(!sig_iter
.has_next());
2381 /* Emit call instruction */
2382 call_inst
= emit(ir
, TGSI_OPCODE_CAL
);
2383 call_inst
->function
= entry
;
2385 /* Process out parameters. */
2386 sig_iter
= sig
->parameters
.iterator();
2387 foreach_iter(exec_list_iterator
, iter
, *ir
) {
2388 ir_rvalue
*param_rval
= (ir_rvalue
*)iter
.get();
2389 ir_variable
*param
= (ir_variable
*)sig_iter
.get();
2391 if (param
->mode
== ir_var_out
||
2392 param
->mode
== ir_var_inout
) {
2393 variable_storage
*storage
= find_variable_storage(param
);
2397 r
.file
= storage
->file
;
2398 r
.index
= storage
->index
;
2400 r
.swizzle
= SWIZZLE_NOOP
;
2403 param_rval
->accept(this);
2404 st_dst_reg l
= st_dst_reg(this->result
);
2406 for (i
= 0; i
< type_size(param
->type
); i
++) {
2407 emit(ir
, TGSI_OPCODE_MOV
, l
, r
);
2415 assert(!sig_iter
.has_next());
2417 /* Process return value. */
2418 this->result
= entry
->return_reg
;
2422 glsl_to_tgsi_visitor::visit(ir_texture
*ir
)
2424 st_src_reg result_src
, coord
, lod_info
, projector
, dx
, dy
;
2425 st_dst_reg result_dst
, coord_dst
;
2426 glsl_to_tgsi_instruction
*inst
= NULL
;
2427 unsigned opcode
= TGSI_OPCODE_NOP
;
2429 if (ir
->coordinate
) {
2430 ir
->coordinate
->accept(this);
2432 /* Put our coords in a temp. We'll need to modify them for shadow,
2433 * projection, or LOD, so the only case we'd use it as is is if
2434 * we're doing plain old texturing. The optimization passes on
2435 * glsl_to_tgsi_visitor should handle cleaning up our mess in that case.
2437 coord
= get_temp(glsl_type::vec4_type
);
2438 coord_dst
= st_dst_reg(coord
);
2439 emit(ir
, TGSI_OPCODE_MOV
, coord_dst
, this->result
);
2442 if (ir
->projector
) {
2443 ir
->projector
->accept(this);
2444 projector
= this->result
;
2447 /* Storage for our result. Ideally for an assignment we'd be using
2448 * the actual storage for the result here, instead.
2450 result_src
= get_temp(glsl_type::vec4_type
);
2451 result_dst
= st_dst_reg(result_src
);
2455 opcode
= TGSI_OPCODE_TEX
;
2458 opcode
= TGSI_OPCODE_TXB
;
2459 ir
->lod_info
.bias
->accept(this);
2460 lod_info
= this->result
;
2463 opcode
= TGSI_OPCODE_TXL
;
2464 ir
->lod_info
.lod
->accept(this);
2465 lod_info
= this->result
;
2468 opcode
= TGSI_OPCODE_TXD
;
2469 ir
->lod_info
.grad
.dPdx
->accept(this);
2471 ir
->lod_info
.grad
.dPdy
->accept(this);
2475 opcode
= TGSI_OPCODE_TXQ
;
2476 ir
->lod_info
.lod
->accept(this);
2477 lod_info
= this->result
;
2480 opcode
= TGSI_OPCODE_TXF
;
2481 ir
->lod_info
.lod
->accept(this);
2482 lod_info
= this->result
;
2486 if (ir
->projector
) {
2487 if (opcode
== TGSI_OPCODE_TEX
) {
2488 /* Slot the projector in as the last component of the coord. */
2489 coord_dst
.writemask
= WRITEMASK_W
;
2490 emit(ir
, TGSI_OPCODE_MOV
, coord_dst
, projector
);
2491 coord_dst
.writemask
= WRITEMASK_XYZW
;
2492 opcode
= TGSI_OPCODE_TXP
;
2494 st_src_reg coord_w
= coord
;
2495 coord_w
.swizzle
= SWIZZLE_WWWW
;
2497 /* For the other TEX opcodes there's no projective version
2498 * since the last slot is taken up by LOD info. Do the
2499 * projective divide now.
2501 coord_dst
.writemask
= WRITEMASK_W
;
2502 emit(ir
, TGSI_OPCODE_RCP
, coord_dst
, projector
);
2504 /* In the case where we have to project the coordinates "by hand,"
2505 * the shadow comparator value must also be projected.
2507 st_src_reg tmp_src
= coord
;
2508 if (ir
->shadow_comparitor
) {
2509 /* Slot the shadow value in as the second to last component of the
2512 ir
->shadow_comparitor
->accept(this);
2514 tmp_src
= get_temp(glsl_type::vec4_type
);
2515 st_dst_reg tmp_dst
= st_dst_reg(tmp_src
);
2517 tmp_dst
.writemask
= WRITEMASK_Z
;
2518 emit(ir
, TGSI_OPCODE_MOV
, tmp_dst
, this->result
);
2520 tmp_dst
.writemask
= WRITEMASK_XY
;
2521 emit(ir
, TGSI_OPCODE_MOV
, tmp_dst
, coord
);
2524 coord_dst
.writemask
= WRITEMASK_XYZ
;
2525 emit(ir
, TGSI_OPCODE_MUL
, coord_dst
, tmp_src
, coord_w
);
2527 coord_dst
.writemask
= WRITEMASK_XYZW
;
2528 coord
.swizzle
= SWIZZLE_XYZW
;
2532 /* If projection is done and the opcode is not TGSI_OPCODE_TXP, then the shadow
2533 * comparator was put in the correct place (and projected) by the code,
2534 * above, that handles by-hand projection.
2536 if (ir
->shadow_comparitor
&& (!ir
->projector
|| opcode
== TGSI_OPCODE_TXP
)) {
2537 /* Slot the shadow value in as the second to last component of the
2540 ir
->shadow_comparitor
->accept(this);
2541 coord_dst
.writemask
= WRITEMASK_Z
;
2542 emit(ir
, TGSI_OPCODE_MOV
, coord_dst
, this->result
);
2543 coord_dst
.writemask
= WRITEMASK_XYZW
;
2546 if (opcode
== TGSI_OPCODE_TXL
|| opcode
== TGSI_OPCODE_TXB
||
2547 opcode
== TGSI_OPCODE_TXF
) {
2548 /* TGSI stores LOD or LOD bias in the last channel of the coords. */
2549 coord_dst
.writemask
= WRITEMASK_W
;
2550 emit(ir
, TGSI_OPCODE_MOV
, coord_dst
, lod_info
);
2551 coord_dst
.writemask
= WRITEMASK_XYZW
;
2554 if (opcode
== TGSI_OPCODE_TXD
)
2555 inst
= emit(ir
, opcode
, result_dst
, coord
, dx
, dy
);
2556 else if (opcode
== TGSI_OPCODE_TXQ
)
2557 inst
= emit(ir
, opcode
, result_dst
, lod_info
);
2559 inst
= emit(ir
, opcode
, result_dst
, coord
);
2561 if (ir
->shadow_comparitor
)
2562 inst
->tex_shadow
= GL_TRUE
;
2564 inst
->sampler
= _mesa_get_sampler_uniform_value(ir
->sampler
,
2565 this->shader_program
,
2568 const glsl_type
*sampler_type
= ir
->sampler
->type
;
2570 switch (sampler_type
->sampler_dimensionality
) {
2571 case GLSL_SAMPLER_DIM_1D
:
2572 inst
->tex_target
= (sampler_type
->sampler_array
)
2573 ? TEXTURE_1D_ARRAY_INDEX
: TEXTURE_1D_INDEX
;
2575 case GLSL_SAMPLER_DIM_2D
:
2576 inst
->tex_target
= (sampler_type
->sampler_array
)
2577 ? TEXTURE_2D_ARRAY_INDEX
: TEXTURE_2D_INDEX
;
2579 case GLSL_SAMPLER_DIM_3D
:
2580 inst
->tex_target
= TEXTURE_3D_INDEX
;
2582 case GLSL_SAMPLER_DIM_CUBE
:
2583 inst
->tex_target
= TEXTURE_CUBE_INDEX
;
2585 case GLSL_SAMPLER_DIM_RECT
:
2586 inst
->tex_target
= TEXTURE_RECT_INDEX
;
2588 case GLSL_SAMPLER_DIM_BUF
:
2589 assert(!"FINISHME: Implement ARB_texture_buffer_object");
2592 assert(!"Should not get here.");
2595 this->result
= result_src
;
2599 glsl_to_tgsi_visitor::visit(ir_return
*ir
)
2601 if (ir
->get_value()) {
2605 assert(current_function
);
2607 ir
->get_value()->accept(this);
2608 st_src_reg r
= this->result
;
2610 l
= st_dst_reg(current_function
->return_reg
);
2612 for (i
= 0; i
< type_size(current_function
->sig
->return_type
); i
++) {
2613 emit(ir
, TGSI_OPCODE_MOV
, l
, r
);
2619 emit(ir
, TGSI_OPCODE_RET
);
2623 glsl_to_tgsi_visitor::visit(ir_discard
*ir
)
2625 struct gl_fragment_program
*fp
= (struct gl_fragment_program
*)this->prog
;
2627 if (ir
->condition
) {
2628 ir
->condition
->accept(this);
2629 this->result
.negate
= ~this->result
.negate
;
2630 emit(ir
, TGSI_OPCODE_KIL
, undef_dst
, this->result
);
2632 emit(ir
, TGSI_OPCODE_KILP
);
2635 fp
->UsesKill
= GL_TRUE
;
2639 glsl_to_tgsi_visitor::visit(ir_if
*ir
)
2641 glsl_to_tgsi_instruction
*cond_inst
, *if_inst
;
2642 glsl_to_tgsi_instruction
*prev_inst
;
2644 prev_inst
= (glsl_to_tgsi_instruction
*)this->instructions
.get_tail();
2646 ir
->condition
->accept(this);
2647 assert(this->result
.file
!= PROGRAM_UNDEFINED
);
2649 if (this->options
->EmitCondCodes
) {
2650 cond_inst
= (glsl_to_tgsi_instruction
*)this->instructions
.get_tail();
2652 /* See if we actually generated any instruction for generating
2653 * the condition. If not, then cook up a move to a temp so we
2654 * have something to set cond_update on.
2656 if (cond_inst
== prev_inst
) {
2657 st_src_reg temp
= get_temp(glsl_type::bool_type
);
2658 cond_inst
= emit(ir
->condition
, TGSI_OPCODE_MOV
, st_dst_reg(temp
), result
);
2660 cond_inst
->cond_update
= GL_TRUE
;
2662 if_inst
= emit(ir
->condition
, TGSI_OPCODE_IF
);
2663 if_inst
->dst
.cond_mask
= COND_NE
;
2665 if_inst
= emit(ir
->condition
, TGSI_OPCODE_IF
, undef_dst
, this->result
);
2668 this->instructions
.push_tail(if_inst
);
2670 visit_exec_list(&ir
->then_instructions
, this);
2672 if (!ir
->else_instructions
.is_empty()) {
2673 emit(ir
->condition
, TGSI_OPCODE_ELSE
);
2674 visit_exec_list(&ir
->else_instructions
, this);
2677 if_inst
= emit(ir
->condition
, TGSI_OPCODE_ENDIF
);
2680 glsl_to_tgsi_visitor::glsl_to_tgsi_visitor()
2682 result
.file
= PROGRAM_UNDEFINED
;
2684 next_signature_id
= 1;
2686 current_function
= NULL
;
2687 num_address_regs
= 0;
2688 indirect_addr_temps
= false;
2689 indirect_addr_consts
= false;
2690 mem_ctx
= ralloc_context(NULL
);
2693 glsl_to_tgsi_visitor::~glsl_to_tgsi_visitor()
2695 ralloc_free(mem_ctx
);
2698 extern "C" void free_glsl_to_tgsi_visitor(glsl_to_tgsi_visitor
*v
)
2705 * Count resources used by the given gpu program (number of texture
2709 count_resources(glsl_to_tgsi_visitor
*v
, gl_program
*prog
)
2711 v
->samplers_used
= 0;
2713 foreach_iter(exec_list_iterator
, iter
, v
->instructions
) {
2714 glsl_to_tgsi_instruction
*inst
= (glsl_to_tgsi_instruction
*)iter
.get();
2716 if (is_tex_instruction(inst
->op
)) {
2717 v
->samplers_used
|= 1 << inst
->sampler
;
2719 prog
->SamplerTargets
[inst
->sampler
] =
2720 (gl_texture_index
)inst
->tex_target
;
2721 if (inst
->tex_shadow
) {
2722 prog
->ShadowSamplers
|= 1 << inst
->sampler
;
2727 prog
->SamplersUsed
= v
->samplers_used
;
2728 _mesa_update_shader_textures_used(prog
);
2733 * Check if the given vertex/fragment/shader program is within the
2734 * resource limits of the context (number of texture units, etc).
2735 * If any of those checks fail, record a linker error.
2737 * XXX more checks are needed...
2740 check_resources(const struct gl_context
*ctx
,
2741 struct gl_shader_program
*shader_program
,
2742 glsl_to_tgsi_visitor
*prog
,
2743 struct gl_program
*proginfo
)
2745 switch (proginfo
->Target
) {
2746 case GL_VERTEX_PROGRAM_ARB
:
2747 if (_mesa_bitcount(prog
->samplers_used
) >
2748 ctx
->Const
.MaxVertexTextureImageUnits
) {
2749 fail_link(shader_program
, "Too many vertex shader texture samplers");
2751 if (proginfo
->Parameters
->NumParameters
> MAX_UNIFORMS
) {
2752 fail_link(shader_program
, "Too many vertex shader constants");
2755 case MESA_GEOMETRY_PROGRAM
:
2756 if (_mesa_bitcount(prog
->samplers_used
) >
2757 ctx
->Const
.MaxGeometryTextureImageUnits
) {
2758 fail_link(shader_program
, "Too many geometry shader texture samplers");
2760 if (proginfo
->Parameters
->NumParameters
>
2761 MAX_GEOMETRY_UNIFORM_COMPONENTS
/ 4) {
2762 fail_link(shader_program
, "Too many geometry shader constants");
2765 case GL_FRAGMENT_PROGRAM_ARB
:
2766 if (_mesa_bitcount(prog
->samplers_used
) >
2767 ctx
->Const
.MaxTextureImageUnits
) {
2768 fail_link(shader_program
, "Too many fragment shader texture samplers");
2770 if (proginfo
->Parameters
->NumParameters
> MAX_UNIFORMS
) {
2771 fail_link(shader_program
, "Too many fragment shader constants");
2775 _mesa_problem(ctx
, "unexpected program type in check_resources()");
2781 struct uniform_sort
{
2782 struct gl_uniform
*u
;
2786 /* The shader_program->Uniforms list is almost sorted in increasing
2787 * uniform->{Frag,Vert}Pos locations, but not quite when there are
2788 * uniforms shared between targets. We need to add parameters in
2789 * increasing order for the targets.
2792 sort_uniforms(const void *a
, const void *b
)
2794 struct uniform_sort
*u1
= (struct uniform_sort
*)a
;
2795 struct uniform_sort
*u2
= (struct uniform_sort
*)b
;
2797 return u1
->pos
- u2
->pos
;
2800 /* Add the uniforms to the parameters. The linker chose locations
2801 * in our parameters lists (which weren't created yet), which the
2802 * uniforms code will use to poke values into our parameters list
2803 * when uniforms are updated.
2806 add_uniforms_to_parameters_list(struct gl_shader_program
*shader_program
,
2807 struct gl_shader
*shader
,
2808 struct gl_program
*prog
)
2811 unsigned int next_sampler
= 0, num_uniforms
= 0;
2812 struct uniform_sort
*sorted_uniforms
;
2814 sorted_uniforms
= ralloc_array(NULL
, struct uniform_sort
,
2815 shader_program
->Uniforms
->NumUniforms
);
2817 for (i
= 0; i
< shader_program
->Uniforms
->NumUniforms
; i
++) {
2818 struct gl_uniform
*uniform
= shader_program
->Uniforms
->Uniforms
+ i
;
2819 int parameter_index
= -1;
2821 switch (shader
->Type
) {
2822 case GL_VERTEX_SHADER
:
2823 parameter_index
= uniform
->VertPos
;
2825 case GL_FRAGMENT_SHADER
:
2826 parameter_index
= uniform
->FragPos
;
2828 case GL_GEOMETRY_SHADER
:
2829 parameter_index
= uniform
->GeomPos
;
2833 /* Only add uniforms used in our target. */
2834 if (parameter_index
!= -1) {
2835 sorted_uniforms
[num_uniforms
].pos
= parameter_index
;
2836 sorted_uniforms
[num_uniforms
].u
= uniform
;
2841 qsort(sorted_uniforms
, num_uniforms
, sizeof(struct uniform_sort
),
2844 for (i
= 0; i
< num_uniforms
; i
++) {
2845 struct gl_uniform
*uniform
= sorted_uniforms
[i
].u
;
2846 int parameter_index
= sorted_uniforms
[i
].pos
;
2847 const glsl_type
*type
= uniform
->Type
;
2850 if (type
->is_vector() ||
2851 type
->is_scalar()) {
2852 size
= type
->vector_elements
;
2854 size
= type_size(type
) * 4;
2857 gl_register_file file
;
2858 if (type
->is_sampler() ||
2859 (type
->is_array() && type
->fields
.array
->is_sampler())) {
2860 file
= PROGRAM_SAMPLER
;
2862 file
= PROGRAM_UNIFORM
;
2865 GLint index
= _mesa_lookup_parameter_index(prog
->Parameters
, -1,
2869 index
= _mesa_add_parameter(prog
->Parameters
, file
,
2870 uniform
->Name
, size
, type
->gl_type
,
2873 /* Sampler uniform values are stored in prog->SamplerUnits,
2874 * and the entry in that array is selected by this index we
2875 * store in ParameterValues[].
2877 if (file
== PROGRAM_SAMPLER
) {
2878 for (unsigned int j
= 0; j
< size
/ 4; j
++)
2879 prog
->Parameters
->ParameterValues
[index
+ j
][0].f
= next_sampler
++;
2882 /* The location chosen in the Parameters list here (returned
2883 * from _mesa_add_uniform) has to match what the linker chose.
2885 if (index
!= parameter_index
) {
2886 fail_link(shader_program
, "Allocation of uniform `%s' to target "
2887 "failed (%d vs %d)\n",
2888 uniform
->Name
, index
, parameter_index
);
2893 ralloc_free(sorted_uniforms
);
2897 set_uniform_initializer(struct gl_context
*ctx
, void *mem_ctx
,
2898 struct gl_shader_program
*shader_program
,
2899 const char *name
, const glsl_type
*type
,
2902 if (type
->is_record()) {
2903 ir_constant
*field_constant
;
2905 field_constant
= (ir_constant
*)val
->components
.get_head();
2907 for (unsigned int i
= 0; i
< type
->length
; i
++) {
2908 const glsl_type
*field_type
= type
->fields
.structure
[i
].type
;
2909 const char *field_name
= ralloc_asprintf(mem_ctx
, "%s.%s", name
,
2910 type
->fields
.structure
[i
].name
);
2911 set_uniform_initializer(ctx
, mem_ctx
, shader_program
, field_name
,
2912 field_type
, field_constant
);
2913 field_constant
= (ir_constant
*)field_constant
->next
;
2918 int loc
= _mesa_get_uniform_location(ctx
, shader_program
, name
);
2921 fail_link(shader_program
,
2922 "Couldn't find uniform for initializer %s\n", name
);
2926 for (unsigned int i
= 0; i
< (type
->is_array() ? type
->length
: 1); i
++) {
2927 ir_constant
*element
;
2928 const glsl_type
*element_type
;
2929 if (type
->is_array()) {
2930 element
= val
->array_elements
[i
];
2931 element_type
= type
->fields
.array
;
2934 element_type
= type
;
2939 if (element_type
->base_type
== GLSL_TYPE_BOOL
) {
2940 int *conv
= ralloc_array(mem_ctx
, int, element_type
->components());
2941 for (unsigned int j
= 0; j
< element_type
->components(); j
++) {
2942 conv
[j
] = element
->value
.b
[j
];
2944 values
= (void *)conv
;
2945 element_type
= glsl_type::get_instance(GLSL_TYPE_INT
,
2946 element_type
->vector_elements
,
2949 values
= &element
->value
;
2952 if (element_type
->is_matrix()) {
2953 _mesa_uniform_matrix(ctx
, shader_program
,
2954 element_type
->matrix_columns
,
2955 element_type
->vector_elements
,
2956 loc
, 1, GL_FALSE
, (GLfloat
*)values
);
2957 loc
+= element_type
->matrix_columns
;
2959 _mesa_uniform(ctx
, shader_program
, loc
, element_type
->matrix_columns
,
2960 values
, element_type
->gl_type
);
2961 loc
+= type_size(element_type
);
2967 set_uniform_initializers(struct gl_context
*ctx
,
2968 struct gl_shader_program
*shader_program
)
2970 void *mem_ctx
= NULL
;
2972 for (unsigned int i
= 0; i
< MESA_SHADER_TYPES
; i
++) {
2973 struct gl_shader
*shader
= shader_program
->_LinkedShaders
[i
];
2978 foreach_iter(exec_list_iterator
, iter
, *shader
->ir
) {
2979 ir_instruction
*ir
= (ir_instruction
*)iter
.get();
2980 ir_variable
*var
= ir
->as_variable();
2982 if (!var
|| var
->mode
!= ir_var_uniform
|| !var
->constant_value
)
2986 mem_ctx
= ralloc_context(NULL
);
2988 set_uniform_initializer(ctx
, mem_ctx
, shader_program
, var
->name
,
2989 var
->type
, var
->constant_value
);
2993 ralloc_free(mem_ctx
);
2997 * Scan/rewrite program to remove reads of custom (output) registers.
2998 * The passed type has to be either PROGRAM_OUTPUT or PROGRAM_VARYING
2999 * (for vertex shaders).
3000 * In GLSL shaders, varying vars can be read and written.
3001 * On some hardware, trying to read an output register causes trouble.
3002 * So, rewrite the program to use a temporary register in this case.
3004 * Based on _mesa_remove_output_reads from programopt.c.
3007 glsl_to_tgsi_visitor::remove_output_reads(gl_register_file type
)
3010 GLint outputMap
[VERT_RESULT_MAX
];
3011 GLint outputTypes
[VERT_RESULT_MAX
];
3012 GLuint numVaryingReads
= 0;
3013 GLboolean usedTemps
[MAX_TEMPS
];
3014 GLuint firstTemp
= 0;
3016 _mesa_find_used_registers(prog
, PROGRAM_TEMPORARY
,
3017 usedTemps
, MAX_TEMPS
);
3019 assert(type
== PROGRAM_VARYING
|| type
== PROGRAM_OUTPUT
);
3020 assert(prog
->Target
== GL_VERTEX_PROGRAM_ARB
|| type
!= PROGRAM_VARYING
);
3022 for (i
= 0; i
< VERT_RESULT_MAX
; i
++)
3025 /* look for instructions which read from varying vars */
3026 foreach_iter(exec_list_iterator
, iter
, this->instructions
) {
3027 glsl_to_tgsi_instruction
*inst
= (glsl_to_tgsi_instruction
*)iter
.get();
3028 const GLuint numSrc
= num_inst_src_regs(inst
->op
);
3030 for (j
= 0; j
< numSrc
; j
++) {
3031 if (inst
->src
[j
].file
== type
) {
3032 /* replace the read with a temp reg */
3033 const GLuint var
= inst
->src
[j
].index
;
3034 if (outputMap
[var
] == -1) {
3036 outputMap
[var
] = _mesa_find_free_register(usedTemps
,
3039 outputTypes
[var
] = inst
->src
[j
].type
;
3040 firstTemp
= outputMap
[var
] + 1;
3042 inst
->src
[j
].file
= PROGRAM_TEMPORARY
;
3043 inst
->src
[j
].index
= outputMap
[var
];
3048 if (numVaryingReads
== 0)
3049 return; /* nothing to be done */
3051 /* look for instructions which write to the varying vars identified above */
3052 foreach_iter(exec_list_iterator
, iter
, this->instructions
) {
3053 glsl_to_tgsi_instruction
*inst
= (glsl_to_tgsi_instruction
*)iter
.get();
3054 if (inst
->dst
.file
== type
&& outputMap
[inst
->dst
.index
] >= 0) {
3055 /* change inst to write to the temp reg, instead of the varying */
3056 inst
->dst
.file
= PROGRAM_TEMPORARY
;
3057 inst
->dst
.index
= outputMap
[inst
->dst
.index
];
3061 /* insert new MOV instructions at the end */
3062 for (i
= 0; i
< VERT_RESULT_MAX
; i
++) {
3063 if (outputMap
[i
] >= 0) {
3064 /* MOV VAR[i], TEMP[tmp]; */
3065 st_src_reg src
= st_src_reg(PROGRAM_TEMPORARY
, outputMap
[i
], outputTypes
[i
]);
3066 st_dst_reg dst
= st_dst_reg(type
, WRITEMASK_XYZW
, outputTypes
[i
]);
3068 this->emit(NULL
, TGSI_OPCODE_MOV
, dst
, src
);
3074 * Returns the mask of channels (bitmask of WRITEMASK_X,Y,Z,W) which
3075 * are read from the given src in this instruction
3078 get_src_arg_mask(st_dst_reg dst
, st_src_reg src
)
3080 int read_mask
= 0, comp
;
3082 /* Now, given the src swizzle and the written channels, find which
3083 * components are actually read
3085 for (comp
= 0; comp
< 4; ++comp
) {
3086 const unsigned coord
= GET_SWZ(src
.swizzle
, comp
);
3088 if (dst
.writemask
& (1 << comp
) && coord
<= SWIZZLE_W
)
3089 read_mask
|= 1 << coord
;
3096 * This pass replaces CMP T0, T1 T2 T0 with MOV T0, T2 when the CMP
3097 * instruction is the first instruction to write to register T0. There are
3098 * several lowering passes done in GLSL IR (e.g. branches and
3099 * relative addressing) that create a large number of conditional assignments
3100 * that ir_to_mesa converts to CMP instructions like the one mentioned above.
3102 * Here is why this conversion is safe:
3103 * CMP T0, T1 T2 T0 can be expanded to:
3109 * If (T1 < 0.0) evaluates to true then our replacement MOV T0, T2 is the same
3110 * as the original program. If (T1 < 0.0) evaluates to false, executing
3111 * MOV T0, T0 will store a garbage value in T0 since T0 is uninitialized.
3112 * Therefore, it doesn't matter that we are replacing MOV T0, T0 with MOV T0, T2
3113 * because any instruction that was going to read from T0 after this was going
3114 * to read a garbage value anyway.
3117 glsl_to_tgsi_visitor::simplify_cmp(void)
3119 unsigned tempWrites
[MAX_TEMPS
];
3120 unsigned outputWrites
[MAX_PROGRAM_OUTPUTS
];
3122 memset(tempWrites
, 0, sizeof(tempWrites
));
3123 memset(outputWrites
, 0, sizeof(outputWrites
));
3125 foreach_iter(exec_list_iterator
, iter
, this->instructions
) {
3126 glsl_to_tgsi_instruction
*inst
= (glsl_to_tgsi_instruction
*)iter
.get();
3127 unsigned prevWriteMask
= 0;
3129 /* Give up if we encounter relative addressing or flow control. */
3130 if (inst
->dst
.reladdr
||
3131 tgsi_get_opcode_info(inst
->op
)->is_branch
||
3132 inst
->op
== TGSI_OPCODE_BGNSUB
||
3133 inst
->op
== TGSI_OPCODE_CONT
||
3134 inst
->op
== TGSI_OPCODE_END
||
3135 inst
->op
== TGSI_OPCODE_ENDSUB
||
3136 inst
->op
== TGSI_OPCODE_RET
) {
3140 if (inst
->dst
.file
== PROGRAM_OUTPUT
) {
3141 assert(inst
->dst
.index
< MAX_PROGRAM_OUTPUTS
);
3142 prevWriteMask
= outputWrites
[inst
->dst
.index
];
3143 outputWrites
[inst
->dst
.index
] |= inst
->dst
.writemask
;
3144 } else if (inst
->dst
.file
== PROGRAM_TEMPORARY
) {
3145 assert(inst
->dst
.index
< MAX_TEMPS
);
3146 prevWriteMask
= tempWrites
[inst
->dst
.index
];
3147 tempWrites
[inst
->dst
.index
] |= inst
->dst
.writemask
;
3150 /* For a CMP to be considered a conditional write, the destination
3151 * register and source register two must be the same. */
3152 if (inst
->op
== TGSI_OPCODE_CMP
3153 && !(inst
->dst
.writemask
& prevWriteMask
)
3154 && inst
->src
[2].file
== inst
->dst
.file
3155 && inst
->src
[2].index
== inst
->dst
.index
3156 && inst
->dst
.writemask
== get_src_arg_mask(inst
->dst
, inst
->src
[2])) {
3158 inst
->op
= TGSI_OPCODE_MOV
;
3159 inst
->src
[0] = inst
->src
[1];
3164 /* Replaces all references to a temporary register index with another index. */
3166 glsl_to_tgsi_visitor::rename_temp_register(int index
, int new_index
)
3168 foreach_iter(exec_list_iterator
, iter
, this->instructions
) {
3169 glsl_to_tgsi_instruction
*inst
= (glsl_to_tgsi_instruction
*)iter
.get();
3172 for (j
=0; j
< num_inst_src_regs(inst
->op
); j
++) {
3173 if (inst
->src
[j
].file
== PROGRAM_TEMPORARY
&&
3174 inst
->src
[j
].index
== index
) {
3175 inst
->src
[j
].index
= new_index
;
3179 if (inst
->dst
.file
== PROGRAM_TEMPORARY
&& inst
->dst
.index
== index
) {
3180 inst
->dst
.index
= new_index
;
3186 glsl_to_tgsi_visitor::get_first_temp_read(int index
)
3188 int depth
= 0; /* loop depth */
3189 int loop_start
= -1; /* index of the first active BGNLOOP (if any) */
3192 foreach_iter(exec_list_iterator
, iter
, this->instructions
) {
3193 glsl_to_tgsi_instruction
*inst
= (glsl_to_tgsi_instruction
*)iter
.get();
3195 for (j
=0; j
< num_inst_src_regs(inst
->op
); j
++) {
3196 if (inst
->src
[j
].file
== PROGRAM_TEMPORARY
&&
3197 inst
->src
[j
].index
== index
) {
3198 return (depth
== 0) ? i
: loop_start
;
3202 if (inst
->op
== TGSI_OPCODE_BGNLOOP
) {
3205 } else if (inst
->op
== TGSI_OPCODE_ENDLOOP
) {
3218 glsl_to_tgsi_visitor::get_first_temp_write(int index
)
3220 int depth
= 0; /* loop depth */
3221 int loop_start
= -1; /* index of the first active BGNLOOP (if any) */
3224 foreach_iter(exec_list_iterator
, iter
, this->instructions
) {
3225 glsl_to_tgsi_instruction
*inst
= (glsl_to_tgsi_instruction
*)iter
.get();
3227 if (inst
->dst
.file
== PROGRAM_TEMPORARY
&& inst
->dst
.index
== index
) {
3228 return (depth
== 0) ? i
: loop_start
;
3231 if (inst
->op
== TGSI_OPCODE_BGNLOOP
) {
3234 } else if (inst
->op
== TGSI_OPCODE_ENDLOOP
) {
3247 glsl_to_tgsi_visitor::get_last_temp_read(int index
)
3249 int depth
= 0; /* loop depth */
3250 int last
= -1; /* index of last instruction that reads the temporary */
3253 foreach_iter(exec_list_iterator
, iter
, this->instructions
) {
3254 glsl_to_tgsi_instruction
*inst
= (glsl_to_tgsi_instruction
*)iter
.get();
3256 for (j
=0; j
< num_inst_src_regs(inst
->op
); j
++) {
3257 if (inst
->src
[j
].file
== PROGRAM_TEMPORARY
&&
3258 inst
->src
[j
].index
== index
) {
3259 last
= (depth
== 0) ? i
: -2;
3263 if (inst
->op
== TGSI_OPCODE_BGNLOOP
)
3265 else if (inst
->op
== TGSI_OPCODE_ENDLOOP
)
3266 if (--depth
== 0 && last
== -2)
3278 glsl_to_tgsi_visitor::get_last_temp_write(int index
)
3280 int depth
= 0; /* loop depth */
3281 int last
= -1; /* index of last instruction that writes to the temporary */
3284 foreach_iter(exec_list_iterator
, iter
, this->instructions
) {
3285 glsl_to_tgsi_instruction
*inst
= (glsl_to_tgsi_instruction
*)iter
.get();
3287 if (inst
->dst
.file
== PROGRAM_TEMPORARY
&& inst
->dst
.index
== index
)
3288 last
= (depth
== 0) ? i
: -2;
3290 if (inst
->op
== TGSI_OPCODE_BGNLOOP
)
3292 else if (inst
->op
== TGSI_OPCODE_ENDLOOP
)
3293 if (--depth
== 0 && last
== -2)
3305 * On a basic block basis, tracks available PROGRAM_TEMPORARY register
3306 * channels for copy propagation and updates following instructions to
3307 * use the original versions.
3309 * The glsl_to_tgsi_visitor lazily produces code assuming that this pass
3310 * will occur. As an example, a TXP production before this pass:
3312 * 0: MOV TEMP[1], INPUT[4].xyyy;
3313 * 1: MOV TEMP[1].w, INPUT[4].wwww;
3314 * 2: TXP TEMP[2], TEMP[1], texture[0], 2D;
3318 * 0: MOV TEMP[1], INPUT[4].xyyy;
3319 * 1: MOV TEMP[1].w, INPUT[4].wwww;
3320 * 2: TXP TEMP[2], INPUT[4].xyyw, texture[0], 2D;
3322 * which allows for dead code elimination on TEMP[1]'s writes.
3325 glsl_to_tgsi_visitor::copy_propagate(void)
3327 glsl_to_tgsi_instruction
**acp
= rzalloc_array(mem_ctx
,
3328 glsl_to_tgsi_instruction
*,
3329 this->next_temp
* 4);
3330 int *acp_level
= rzalloc_array(mem_ctx
, int, this->next_temp
* 4);
3333 foreach_iter(exec_list_iterator
, iter
, this->instructions
) {
3334 glsl_to_tgsi_instruction
*inst
= (glsl_to_tgsi_instruction
*)iter
.get();
3336 assert(inst
->dst
.file
!= PROGRAM_TEMPORARY
3337 || inst
->dst
.index
< this->next_temp
);
3339 /* First, do any copy propagation possible into the src regs. */
3340 for (int r
= 0; r
< 3; r
++) {
3341 glsl_to_tgsi_instruction
*first
= NULL
;
3343 int acp_base
= inst
->src
[r
].index
* 4;
3345 if (inst
->src
[r
].file
!= PROGRAM_TEMPORARY
||
3346 inst
->src
[r
].reladdr
)
3349 /* See if we can find entries in the ACP consisting of MOVs
3350 * from the same src register for all the swizzled channels
3351 * of this src register reference.
3353 for (int i
= 0; i
< 4; i
++) {
3354 int src_chan
= GET_SWZ(inst
->src
[r
].swizzle
, i
);
3355 glsl_to_tgsi_instruction
*copy_chan
= acp
[acp_base
+ src_chan
];
3362 assert(acp_level
[acp_base
+ src_chan
] <= level
);
3367 if (first
->src
[0].file
!= copy_chan
->src
[0].file
||
3368 first
->src
[0].index
!= copy_chan
->src
[0].index
) {
3376 /* We've now validated that we can copy-propagate to
3377 * replace this src register reference. Do it.
3379 inst
->src
[r
].file
= first
->src
[0].file
;
3380 inst
->src
[r
].index
= first
->src
[0].index
;
3383 for (int i
= 0; i
< 4; i
++) {
3384 int src_chan
= GET_SWZ(inst
->src
[r
].swizzle
, i
);
3385 glsl_to_tgsi_instruction
*copy_inst
= acp
[acp_base
+ src_chan
];
3386 swizzle
|= (GET_SWZ(copy_inst
->src
[0].swizzle
, src_chan
) <<
3389 inst
->src
[r
].swizzle
= swizzle
;
3394 case TGSI_OPCODE_BGNLOOP
:
3395 case TGSI_OPCODE_ENDLOOP
:
3396 /* End of a basic block, clear the ACP entirely. */
3397 memset(acp
, 0, sizeof(*acp
) * this->next_temp
* 4);
3400 case TGSI_OPCODE_IF
:
3404 case TGSI_OPCODE_ENDIF
:
3405 case TGSI_OPCODE_ELSE
:
3406 /* Clear all channels written inside the block from the ACP, but
3407 * leaving those that were not touched.
3409 for (int r
= 0; r
< this->next_temp
; r
++) {
3410 for (int c
= 0; c
< 4; c
++) {
3411 if (!acp
[4 * r
+ c
])
3414 if (acp_level
[4 * r
+ c
] >= level
)
3415 acp
[4 * r
+ c
] = NULL
;
3418 if (inst
->op
== TGSI_OPCODE_ENDIF
)
3423 /* Continuing the block, clear any written channels from
3426 if (inst
->dst
.file
== PROGRAM_TEMPORARY
&& inst
->dst
.reladdr
) {
3427 /* Any temporary might be written, so no copy propagation
3428 * across this instruction.
3430 memset(acp
, 0, sizeof(*acp
) * this->next_temp
* 4);
3431 } else if (inst
->dst
.file
== PROGRAM_OUTPUT
&&
3432 inst
->dst
.reladdr
) {
3433 /* Any output might be written, so no copy propagation
3434 * from outputs across this instruction.
3436 for (int r
= 0; r
< this->next_temp
; r
++) {
3437 for (int c
= 0; c
< 4; c
++) {
3438 if (!acp
[4 * r
+ c
])
3441 if (acp
[4 * r
+ c
]->src
[0].file
== PROGRAM_OUTPUT
)
3442 acp
[4 * r
+ c
] = NULL
;
3445 } else if (inst
->dst
.file
== PROGRAM_TEMPORARY
||
3446 inst
->dst
.file
== PROGRAM_OUTPUT
) {
3447 /* Clear where it's used as dst. */
3448 if (inst
->dst
.file
== PROGRAM_TEMPORARY
) {
3449 for (int c
= 0; c
< 4; c
++) {
3450 if (inst
->dst
.writemask
& (1 << c
)) {
3451 acp
[4 * inst
->dst
.index
+ c
] = NULL
;
3456 /* Clear where it's used as src. */
3457 for (int r
= 0; r
< this->next_temp
; r
++) {
3458 for (int c
= 0; c
< 4; c
++) {
3459 if (!acp
[4 * r
+ c
])
3462 int src_chan
= GET_SWZ(acp
[4 * r
+ c
]->src
[0].swizzle
, c
);
3464 if (acp
[4 * r
+ c
]->src
[0].file
== inst
->dst
.file
&&
3465 acp
[4 * r
+ c
]->src
[0].index
== inst
->dst
.index
&&
3466 inst
->dst
.writemask
& (1 << src_chan
))
3468 acp
[4 * r
+ c
] = NULL
;
3476 /* If this is a copy, add it to the ACP. */
3477 if (inst
->op
== TGSI_OPCODE_MOV
&&
3478 inst
->dst
.file
== PROGRAM_TEMPORARY
&&
3479 !inst
->dst
.reladdr
&&
3481 !inst
->src
[0].reladdr
&&
3482 !inst
->src
[0].negate
) {
3483 for (int i
= 0; i
< 4; i
++) {
3484 if (inst
->dst
.writemask
& (1 << i
)) {
3485 acp
[4 * inst
->dst
.index
+ i
] = inst
;
3486 acp_level
[4 * inst
->dst
.index
+ i
] = level
;
3492 ralloc_free(acp_level
);
3497 * Tracks available PROGRAM_TEMPORARY registers for dead code elimination.
3499 * The glsl_to_tgsi_visitor lazily produces code assuming that this pass
3500 * will occur. As an example, a TXP production after copy propagation but
3503 * 0: MOV TEMP[1], INPUT[4].xyyy;
3504 * 1: MOV TEMP[1].w, INPUT[4].wwww;
3505 * 2: TXP TEMP[2], INPUT[4].xyyw, texture[0], 2D;
3507 * and after this pass:
3509 * 0: TXP TEMP[2], INPUT[4].xyyw, texture[0], 2D;
3511 * FIXME: assumes that all functions are inlined (no support for BGNSUB/ENDSUB)
3512 * FIXME: doesn't eliminate all dead code inside of loops; it steps around them
3515 glsl_to_tgsi_visitor::eliminate_dead_code(void)
3519 for (i
=0; i
< this->next_temp
; i
++) {
3520 int last_read
= get_last_temp_read(i
);
3523 foreach_iter(exec_list_iterator
, iter
, this->instructions
) {
3524 glsl_to_tgsi_instruction
*inst
= (glsl_to_tgsi_instruction
*)iter
.get();
3526 if (inst
->dst
.file
== PROGRAM_TEMPORARY
&& inst
->dst
.index
== i
&&
3539 * On a basic block basis, tracks available PROGRAM_TEMPORARY registers for dead
3540 * code elimination. This is less primitive than eliminate_dead_code(), as it
3541 * is per-channel and can detect consecutive writes without a read between them
3542 * as dead code. However, there is some dead code that can be eliminated by
3543 * eliminate_dead_code() but not this function - for example, this function
3544 * cannot eliminate an instruction writing to a register that is never read and
3545 * is the only instruction writing to that register.
3547 * The glsl_to_tgsi_visitor lazily produces code assuming that this pass
3551 glsl_to_tgsi_visitor::eliminate_dead_code_advanced(void)
3553 glsl_to_tgsi_instruction
**writes
= rzalloc_array(mem_ctx
,
3554 glsl_to_tgsi_instruction
*,
3555 this->next_temp
* 4);
3556 int *write_level
= rzalloc_array(mem_ctx
, int, this->next_temp
* 4);
3560 foreach_iter(exec_list_iterator
, iter
, this->instructions
) {
3561 glsl_to_tgsi_instruction
*inst
= (glsl_to_tgsi_instruction
*)iter
.get();
3563 assert(inst
->dst
.file
!= PROGRAM_TEMPORARY
3564 || inst
->dst
.index
< this->next_temp
);
3567 case TGSI_OPCODE_BGNLOOP
:
3568 case TGSI_OPCODE_ENDLOOP
:
3569 /* End of a basic block, clear the write array entirely.
3570 * FIXME: This keeps us from killing dead code when the writes are
3571 * on either side of a loop, even when the register isn't touched
3574 memset(writes
, 0, sizeof(*writes
) * this->next_temp
* 4);
3577 case TGSI_OPCODE_ENDIF
:
3581 case TGSI_OPCODE_ELSE
:
3582 /* Clear all channels written inside the preceding if block from the
3583 * write array, but leave those that were not touched.
3585 * FIXME: This destroys opportunities to remove dead code inside of
3586 * IF blocks that are followed by an ELSE block.
3588 for (int r
= 0; r
< this->next_temp
; r
++) {
3589 for (int c
= 0; c
< 4; c
++) {
3590 if (!writes
[4 * r
+ c
])
3593 if (write_level
[4 * r
+ c
] >= level
)
3594 writes
[4 * r
+ c
] = NULL
;
3599 case TGSI_OPCODE_IF
:
3601 /* fallthrough to default case to mark the condition as read */
3604 /* Continuing the block, clear any channels from the write array that
3605 * are read by this instruction.
3607 for (unsigned i
= 0; i
< Elements(inst
->src
); i
++) {
3608 if (inst
->src
[i
].file
== PROGRAM_TEMPORARY
&& inst
->src
[i
].reladdr
){
3609 /* Any temporary might be read, so no dead code elimination
3610 * across this instruction.
3612 memset(writes
, 0, sizeof(*writes
) * this->next_temp
* 4);
3613 } else if (inst
->src
[i
].file
== PROGRAM_TEMPORARY
) {
3614 /* Clear where it's used as src. */
3615 int src_chans
= 1 << GET_SWZ(inst
->src
[i
].swizzle
, 0);
3616 src_chans
|= 1 << GET_SWZ(inst
->src
[i
].swizzle
, 1);
3617 src_chans
|= 1 << GET_SWZ(inst
->src
[i
].swizzle
, 2);
3618 src_chans
|= 1 << GET_SWZ(inst
->src
[i
].swizzle
, 3);
3620 for (int c
= 0; c
< 4; c
++) {
3621 if (src_chans
& (1 << c
)) {
3622 writes
[4 * inst
->src
[i
].index
+ c
] = NULL
;
3630 /* If this instruction writes to a temporary, add it to the write array.
3631 * If there is already an instruction in the write array for one or more
3632 * of the channels, flag that channel write as dead.
3634 if (inst
->dst
.file
== PROGRAM_TEMPORARY
&&
3635 !inst
->dst
.reladdr
&&
3637 for (int c
= 0; c
< 4; c
++) {
3638 if (inst
->dst
.writemask
& (1 << c
)) {
3639 if (writes
[4 * inst
->dst
.index
+ c
]) {
3640 if (write_level
[4 * inst
->dst
.index
+ c
] < level
)
3643 writes
[4 * inst
->dst
.index
+ c
]->dead_mask
|= (1 << c
);
3645 writes
[4 * inst
->dst
.index
+ c
] = inst
;
3646 write_level
[4 * inst
->dst
.index
+ c
] = level
;
3652 /* Anything still in the write array at this point is dead code. */
3653 for (int r
= 0; r
< this->next_temp
; r
++) {
3654 for (int c
= 0; c
< 4; c
++) {
3655 glsl_to_tgsi_instruction
*inst
= writes
[4 * r
+ c
];
3657 inst
->dead_mask
|= (1 << c
);
3661 /* Now actually remove the instructions that are completely dead and update
3662 * the writemask of other instructions with dead channels.
3664 foreach_iter(exec_list_iterator
, iter
, this->instructions
) {
3665 glsl_to_tgsi_instruction
*inst
= (glsl_to_tgsi_instruction
*)iter
.get();
3667 if (!inst
->dead_mask
|| !inst
->dst
.writemask
)
3669 else if (inst
->dead_mask
== inst
->dst
.writemask
) {
3674 inst
->dst
.writemask
&= ~(inst
->dead_mask
);
3677 ralloc_free(write_level
);
3678 ralloc_free(writes
);
3683 /* Merges temporary registers together where possible to reduce the number of
3684 * registers needed to run a program.
3686 * Produces optimal code only after copy propagation and dead code elimination
3689 glsl_to_tgsi_visitor::merge_registers(void)
3691 int *last_reads
= rzalloc_array(mem_ctx
, int, this->next_temp
);
3692 int *first_writes
= rzalloc_array(mem_ctx
, int, this->next_temp
);
3695 /* Read the indices of the last read and first write to each temp register
3696 * into an array so that we don't have to traverse the instruction list as
3698 for (i
=0; i
< this->next_temp
; i
++) {
3699 last_reads
[i
] = get_last_temp_read(i
);
3700 first_writes
[i
] = get_first_temp_write(i
);
3703 /* Start looking for registers with non-overlapping usages that can be
3704 * merged together. */
3705 for (i
=0; i
< this->next_temp
; i
++) {
3706 /* Don't touch unused registers. */
3707 if (last_reads
[i
] < 0 || first_writes
[i
] < 0) continue;
3709 for (j
=0; j
< this->next_temp
; j
++) {
3710 /* Don't touch unused registers. */
3711 if (last_reads
[j
] < 0 || first_writes
[j
] < 0) continue;
3713 /* We can merge the two registers if the first write to j is after or
3714 * in the same instruction as the last read from i. Note that the
3715 * register at index i will always be used earlier or at the same time
3716 * as the register at index j. */
3717 if (first_writes
[i
] <= first_writes
[j
] &&
3718 last_reads
[i
] <= first_writes
[j
])
3720 rename_temp_register(j
, i
); /* Replace all references to j with i.*/
3722 /* Update the first_writes and last_reads arrays with the new
3723 * values for the merged register index, and mark the newly unused
3724 * register index as such. */
3725 last_reads
[i
] = last_reads
[j
];
3726 first_writes
[j
] = -1;
3732 ralloc_free(last_reads
);
3733 ralloc_free(first_writes
);
3736 /* Reassign indices to temporary registers by reusing unused indices created
3737 * by optimization passes. */
3739 glsl_to_tgsi_visitor::renumber_registers(void)
3744 for (i
=0; i
< this->next_temp
; i
++) {
3745 if (get_first_temp_read(i
) < 0) continue;
3747 rename_temp_register(i
, new_index
);
3751 this->next_temp
= new_index
;
3755 * Returns a fragment program which implements the current pixel transfer ops.
3756 * Based on get_pixel_transfer_program in st_atom_pixeltransfer.c.
3759 get_pixel_transfer_visitor(struct st_fragment_program
*fp
,
3760 glsl_to_tgsi_visitor
*original
,
3761 int scale_and_bias
, int pixel_maps
)
3763 glsl_to_tgsi_visitor
*v
= new glsl_to_tgsi_visitor();
3764 struct st_context
*st
= st_context(original
->ctx
);
3765 struct gl_program
*prog
= &fp
->Base
.Base
;
3766 struct gl_program_parameter_list
*params
= _mesa_new_parameter_list();
3767 st_src_reg coord
, src0
;
3769 glsl_to_tgsi_instruction
*inst
;
3771 /* Copy attributes of the glsl_to_tgsi_visitor in the original shader. */
3772 v
->ctx
= original
->ctx
;
3774 v
->glsl_version
= original
->glsl_version
;
3775 v
->native_integers
= original
->native_integers
;
3776 v
->options
= original
->options
;
3777 v
->next_temp
= original
->next_temp
;
3778 v
->num_address_regs
= original
->num_address_regs
;
3779 v
->samplers_used
= prog
->SamplersUsed
= original
->samplers_used
;
3780 v
->indirect_addr_temps
= original
->indirect_addr_temps
;
3781 v
->indirect_addr_consts
= original
->indirect_addr_consts
;
3782 memcpy(&v
->immediates
, &original
->immediates
, sizeof(v
->immediates
));
3785 * Get initial pixel color from the texture.
3786 * TEX colorTemp, fragment.texcoord[0], texture[0], 2D;
3788 coord
= st_src_reg(PROGRAM_INPUT
, FRAG_ATTRIB_TEX0
, glsl_type::vec2_type
);
3789 src0
= v
->get_temp(glsl_type::vec4_type
);
3790 dst0
= st_dst_reg(src0
);
3791 inst
= v
->emit(NULL
, TGSI_OPCODE_TEX
, dst0
, coord
);
3793 inst
->tex_target
= TEXTURE_2D_INDEX
;
3795 prog
->InputsRead
|= (1 << FRAG_ATTRIB_TEX0
);
3796 prog
->SamplersUsed
|= (1 << 0); /* mark sampler 0 as used */
3797 v
->samplers_used
|= (1 << 0);
3799 if (scale_and_bias
) {
3800 static const gl_state_index scale_state
[STATE_LENGTH
] =
3801 { STATE_INTERNAL
, STATE_PT_SCALE
,
3802 (gl_state_index
) 0, (gl_state_index
) 0, (gl_state_index
) 0 };
3803 static const gl_state_index bias_state
[STATE_LENGTH
] =
3804 { STATE_INTERNAL
, STATE_PT_BIAS
,
3805 (gl_state_index
) 0, (gl_state_index
) 0, (gl_state_index
) 0 };
3806 GLint scale_p
, bias_p
;
3807 st_src_reg scale
, bias
;
3809 scale_p
= _mesa_add_state_reference(params
, scale_state
);
3810 bias_p
= _mesa_add_state_reference(params
, bias_state
);
3812 /* MAD colorTemp, colorTemp, scale, bias; */
3813 scale
= st_src_reg(PROGRAM_STATE_VAR
, scale_p
, GLSL_TYPE_FLOAT
);
3814 bias
= st_src_reg(PROGRAM_STATE_VAR
, bias_p
, GLSL_TYPE_FLOAT
);
3815 inst
= v
->emit(NULL
, TGSI_OPCODE_MAD
, dst0
, src0
, scale
, bias
);
3819 st_src_reg temp
= v
->get_temp(glsl_type::vec4_type
);
3820 st_dst_reg temp_dst
= st_dst_reg(temp
);
3822 assert(st
->pixel_xfer
.pixelmap_texture
);
3824 /* With a little effort, we can do four pixel map look-ups with
3825 * two TEX instructions:
3828 /* TEX temp.rg, colorTemp.rgba, texture[1], 2D; */
3829 temp_dst
.writemask
= WRITEMASK_XY
; /* write R,G */
3830 inst
= v
->emit(NULL
, TGSI_OPCODE_TEX
, temp_dst
, src0
);
3832 inst
->tex_target
= TEXTURE_2D_INDEX
;
3834 /* TEX temp.ba, colorTemp.baba, texture[1], 2D; */
3835 src0
.swizzle
= MAKE_SWIZZLE4(SWIZZLE_Z
, SWIZZLE_W
, SWIZZLE_Z
, SWIZZLE_W
);
3836 temp_dst
.writemask
= WRITEMASK_ZW
; /* write B,A */
3837 inst
= v
->emit(NULL
, TGSI_OPCODE_TEX
, temp_dst
, src0
);
3839 inst
->tex_target
= TEXTURE_2D_INDEX
;
3841 prog
->SamplersUsed
|= (1 << 1); /* mark sampler 1 as used */
3842 v
->samplers_used
|= (1 << 1);
3844 /* MOV colorTemp, temp; */
3845 inst
= v
->emit(NULL
, TGSI_OPCODE_MOV
, dst0
, temp
);
3848 /* Now copy the instructions from the original glsl_to_tgsi_visitor into the
3850 foreach_iter(exec_list_iterator
, iter
, original
->instructions
) {
3851 glsl_to_tgsi_instruction
*inst
= (glsl_to_tgsi_instruction
*)iter
.get();
3852 st_src_reg src_regs
[3];
3854 if (inst
->dst
.file
== PROGRAM_OUTPUT
)
3855 prog
->OutputsWritten
|= BITFIELD64_BIT(inst
->dst
.index
);
3857 for (int i
=0; i
<3; i
++) {
3858 src_regs
[i
] = inst
->src
[i
];
3859 if (src_regs
[i
].file
== PROGRAM_INPUT
&&
3860 src_regs
[i
].index
== FRAG_ATTRIB_COL0
)
3862 src_regs
[i
].file
= PROGRAM_TEMPORARY
;
3863 src_regs
[i
].index
= src0
.index
;
3865 else if (src_regs
[i
].file
== PROGRAM_INPUT
)
3866 prog
->InputsRead
|= (1 << src_regs
[i
].index
);
3869 v
->emit(NULL
, inst
->op
, inst
->dst
, src_regs
[0], src_regs
[1], src_regs
[2]);
3872 /* Make modifications to fragment program info. */
3873 prog
->Parameters
= _mesa_combine_parameter_lists(params
,
3874 original
->prog
->Parameters
);
3875 prog
->Attributes
= _mesa_clone_parameter_list(original
->prog
->Attributes
);
3876 prog
->Varying
= _mesa_clone_parameter_list(original
->prog
->Varying
);
3877 _mesa_free_parameter_list(params
);
3878 count_resources(v
, prog
);
3879 fp
->glsl_to_tgsi
= v
;
3883 * Make fragment program for glBitmap:
3884 * Sample the texture and kill the fragment if the bit is 0.
3885 * This program will be combined with the user's fragment program.
3887 * Based on make_bitmap_fragment_program in st_cb_bitmap.c.
3890 get_bitmap_visitor(struct st_fragment_program
*fp
,
3891 glsl_to_tgsi_visitor
*original
, int samplerIndex
)
3893 glsl_to_tgsi_visitor
*v
= new glsl_to_tgsi_visitor();
3894 struct st_context
*st
= st_context(original
->ctx
);
3895 struct gl_program
*prog
= &fp
->Base
.Base
;
3896 st_src_reg coord
, src0
;
3898 glsl_to_tgsi_instruction
*inst
;
3900 /* Copy attributes of the glsl_to_tgsi_visitor in the original shader. */
3901 v
->ctx
= original
->ctx
;
3903 v
->glsl_version
= original
->glsl_version
;
3904 v
->native_integers
= original
->native_integers
;
3905 v
->options
= original
->options
;
3906 v
->next_temp
= original
->next_temp
;
3907 v
->num_address_regs
= original
->num_address_regs
;
3908 v
->samplers_used
= prog
->SamplersUsed
= original
->samplers_used
;
3909 v
->indirect_addr_temps
= original
->indirect_addr_temps
;
3910 v
->indirect_addr_consts
= original
->indirect_addr_consts
;
3911 memcpy(&v
->immediates
, &original
->immediates
, sizeof(v
->immediates
));
3913 /* TEX tmp0, fragment.texcoord[0], texture[0], 2D; */
3914 coord
= st_src_reg(PROGRAM_INPUT
, FRAG_ATTRIB_TEX0
, glsl_type::vec2_type
);
3915 src0
= v
->get_temp(glsl_type::vec4_type
);
3916 dst0
= st_dst_reg(src0
);
3917 inst
= v
->emit(NULL
, TGSI_OPCODE_TEX
, dst0
, coord
);
3918 inst
->sampler
= samplerIndex
;
3919 inst
->tex_target
= TEXTURE_2D_INDEX
;
3921 prog
->InputsRead
|= (1 << FRAG_ATTRIB_TEX0
);
3922 prog
->SamplersUsed
|= (1 << samplerIndex
); /* mark sampler as used */
3923 v
->samplers_used
|= (1 << samplerIndex
);
3925 /* KIL if -tmp0 < 0 # texel=0 -> keep / texel=0 -> discard */
3926 src0
.negate
= NEGATE_XYZW
;
3927 if (st
->bitmap
.tex_format
== PIPE_FORMAT_L8_UNORM
)
3928 src0
.swizzle
= SWIZZLE_XXXX
;
3929 inst
= v
->emit(NULL
, TGSI_OPCODE_KIL
, undef_dst
, src0
);
3931 /* Now copy the instructions from the original glsl_to_tgsi_visitor into the
3933 foreach_iter(exec_list_iterator
, iter
, original
->instructions
) {
3934 glsl_to_tgsi_instruction
*inst
= (glsl_to_tgsi_instruction
*)iter
.get();
3935 st_src_reg src_regs
[3];
3937 if (inst
->dst
.file
== PROGRAM_OUTPUT
)
3938 prog
->OutputsWritten
|= BITFIELD64_BIT(inst
->dst
.index
);
3940 for (int i
=0; i
<3; i
++) {
3941 src_regs
[i
] = inst
->src
[i
];
3942 if (src_regs
[i
].file
== PROGRAM_INPUT
)
3943 prog
->InputsRead
|= (1 << src_regs
[i
].index
);
3946 v
->emit(NULL
, inst
->op
, inst
->dst
, src_regs
[0], src_regs
[1], src_regs
[2]);
3949 /* Make modifications to fragment program info. */
3950 prog
->Parameters
= _mesa_clone_parameter_list(original
->prog
->Parameters
);
3951 prog
->Attributes
= _mesa_clone_parameter_list(original
->prog
->Attributes
);
3952 prog
->Varying
= _mesa_clone_parameter_list(original
->prog
->Varying
);
3953 count_resources(v
, prog
);
3954 fp
->glsl_to_tgsi
= v
;
3957 /* ------------------------- TGSI conversion stuff -------------------------- */
3959 unsigned branch_target
;
3964 * Intermediate state used during shader translation.
3966 struct st_translate
{
3967 struct ureg_program
*ureg
;
3969 struct ureg_dst temps
[MAX_TEMPS
];
3970 struct ureg_src
*constants
;
3971 struct ureg_src
*immediates
;
3972 struct ureg_dst outputs
[PIPE_MAX_SHADER_OUTPUTS
];
3973 struct ureg_src inputs
[PIPE_MAX_SHADER_INPUTS
];
3974 struct ureg_dst address
[1];
3975 struct ureg_src samplers
[PIPE_MAX_SAMPLERS
];
3976 struct ureg_src systemValues
[SYSTEM_VALUE_MAX
];
3978 /* Extra info for handling point size clamping in vertex shader */
3979 struct ureg_dst pointSizeResult
; /**< Actual point size output register */
3980 struct ureg_src pointSizeConst
; /**< Point size range constant register */
3981 GLint pointSizeOutIndex
; /**< Temp point size output register */
3982 GLboolean prevInstWrotePointSize
;
3984 const GLuint
*inputMapping
;
3985 const GLuint
*outputMapping
;
3987 /* For every instruction that contains a label (eg CALL), keep
3988 * details so that we can go back afterwards and emit the correct
3989 * tgsi instruction number for each label.
3991 struct label
*labels
;
3992 unsigned labels_size
;
3993 unsigned labels_count
;
3995 /* Keep a record of the tgsi instruction number that each mesa
3996 * instruction starts at, will be used to fix up labels after
4001 unsigned insn_count
;
4003 unsigned procType
; /**< TGSI_PROCESSOR_VERTEX/FRAGMENT */
4008 /** Map Mesa's SYSTEM_VALUE_x to TGSI_SEMANTIC_x */
4009 static unsigned mesa_sysval_to_semantic
[SYSTEM_VALUE_MAX
] = {
4011 TGSI_SEMANTIC_INSTANCEID
4015 * Make note of a branch to a label in the TGSI code.
4016 * After we've emitted all instructions, we'll go over the list
4017 * of labels built here and patch the TGSI code with the actual
4018 * location of each label.
4020 static unsigned *get_label(struct st_translate
*t
, unsigned branch_target
)
4024 if (t
->labels_count
+ 1 >= t
->labels_size
) {
4025 t
->labels_size
= 1 << (util_logbase2(t
->labels_size
) + 1);
4026 t
->labels
= (struct label
*)realloc(t
->labels
,
4027 t
->labels_size
* sizeof(struct label
));
4028 if (t
->labels
== NULL
) {
4029 static unsigned dummy
;
4035 i
= t
->labels_count
++;
4036 t
->labels
[i
].branch_target
= branch_target
;
4037 return &t
->labels
[i
].token
;
4041 * Called prior to emitting the TGSI code for each instruction.
4042 * Allocate additional space for instructions if needed.
4043 * Update the insn[] array so the next glsl_to_tgsi_instruction points to
4044 * the next TGSI instruction.
4046 static void set_insn_start(struct st_translate
*t
, unsigned start
)
4048 if (t
->insn_count
+ 1 >= t
->insn_size
) {
4049 t
->insn_size
= 1 << (util_logbase2(t
->insn_size
) + 1);
4050 t
->insn
= (unsigned *)realloc(t
->insn
, t
->insn_size
* sizeof(t
->insn
[0]));
4051 if (t
->insn
== NULL
) {
4057 t
->insn
[t
->insn_count
++] = start
;
4061 * Map a glsl_to_tgsi constant/immediate to a TGSI immediate.
4063 static struct ureg_src
4064 emit_immediate(struct st_translate
*t
,
4065 gl_constant_value values
[4],
4068 struct ureg_program
*ureg
= t
->ureg
;
4073 return ureg_DECL_immediate(ureg
, &values
[0].f
, size
);
4075 return ureg_DECL_immediate_int(ureg
, &values
[0].i
, size
);
4076 case GL_UNSIGNED_INT
:
4078 return ureg_DECL_immediate_uint(ureg
, &values
[0].u
, size
);
4080 assert(!"should not get here - type must be float, int, uint, or bool");
4081 return ureg_src_undef();
4086 * Map a glsl_to_tgsi dst register to a TGSI ureg_dst register.
4088 static struct ureg_dst
4089 dst_register(struct st_translate
*t
,
4090 gl_register_file file
,
4094 case PROGRAM_UNDEFINED
:
4095 return ureg_dst_undef();
4097 case PROGRAM_TEMPORARY
:
4098 if (ureg_dst_is_undef(t
->temps
[index
]))
4099 t
->temps
[index
] = ureg_DECL_temporary(t
->ureg
);
4101 return t
->temps
[index
];
4103 case PROGRAM_OUTPUT
:
4104 if (t
->procType
== TGSI_PROCESSOR_VERTEX
&& index
== VERT_RESULT_PSIZ
)
4105 t
->prevInstWrotePointSize
= GL_TRUE
;
4107 if (t
->procType
== TGSI_PROCESSOR_VERTEX
)
4108 assert(index
< VERT_RESULT_MAX
);
4109 else if (t
->procType
== TGSI_PROCESSOR_FRAGMENT
)
4110 assert(index
< FRAG_RESULT_MAX
);
4112 assert(index
< GEOM_RESULT_MAX
);
4114 assert(t
->outputMapping
[index
] < Elements(t
->outputs
));
4116 return t
->outputs
[t
->outputMapping
[index
]];
4118 case PROGRAM_ADDRESS
:
4119 return t
->address
[index
];
4122 assert(!"unknown dst register file");
4123 return ureg_dst_undef();
4128 * Map a glsl_to_tgsi src register to a TGSI ureg_src register.
4130 static struct ureg_src
4131 src_register(struct st_translate
*t
,
4132 gl_register_file file
,
4136 case PROGRAM_UNDEFINED
:
4137 return ureg_src_undef();
4139 case PROGRAM_TEMPORARY
:
4141 assert(index
< Elements(t
->temps
));
4142 if (ureg_dst_is_undef(t
->temps
[index
]))
4143 t
->temps
[index
] = ureg_DECL_temporary(t
->ureg
);
4144 return ureg_src(t
->temps
[index
]);
4146 case PROGRAM_NAMED_PARAM
:
4147 case PROGRAM_ENV_PARAM
:
4148 case PROGRAM_LOCAL_PARAM
:
4149 case PROGRAM_UNIFORM
:
4151 return t
->constants
[index
];
4152 case PROGRAM_STATE_VAR
:
4153 case PROGRAM_CONSTANT
: /* ie, immediate */
4155 return ureg_DECL_constant(t
->ureg
, 0);
4157 return t
->constants
[index
];
4159 case PROGRAM_IMMEDIATE
:
4160 return t
->immediates
[index
];
4163 assert(t
->inputMapping
[index
] < Elements(t
->inputs
));
4164 return t
->inputs
[t
->inputMapping
[index
]];
4166 case PROGRAM_OUTPUT
:
4167 assert(t
->outputMapping
[index
] < Elements(t
->outputs
));
4168 return ureg_src(t
->outputs
[t
->outputMapping
[index
]]); /* not needed? */
4170 case PROGRAM_ADDRESS
:
4171 return ureg_src(t
->address
[index
]);
4173 case PROGRAM_SYSTEM_VALUE
:
4174 assert(index
< Elements(t
->systemValues
));
4175 return t
->systemValues
[index
];
4178 assert(!"unknown src register file");
4179 return ureg_src_undef();
4184 * Create a TGSI ureg_dst register from an st_dst_reg.
4186 static struct ureg_dst
4187 translate_dst(struct st_translate
*t
,
4188 const st_dst_reg
*dst_reg
,
4191 struct ureg_dst dst
= dst_register(t
,
4195 dst
= ureg_writemask(dst
, dst_reg
->writemask
);
4198 dst
= ureg_saturate(dst
);
4200 if (dst_reg
->reladdr
!= NULL
)
4201 dst
= ureg_dst_indirect(dst
, ureg_src(t
->address
[0]));
4207 * Create a TGSI ureg_src register from an st_src_reg.
4209 static struct ureg_src
4210 translate_src(struct st_translate
*t
, const st_src_reg
*src_reg
)
4212 struct ureg_src src
= src_register(t
, src_reg
->file
, src_reg
->index
);
4214 src
= ureg_swizzle(src
,
4215 GET_SWZ(src_reg
->swizzle
, 0) & 0x3,
4216 GET_SWZ(src_reg
->swizzle
, 1) & 0x3,
4217 GET_SWZ(src_reg
->swizzle
, 2) & 0x3,
4218 GET_SWZ(src_reg
->swizzle
, 3) & 0x3);
4220 if ((src_reg
->negate
& 0xf) == NEGATE_XYZW
)
4221 src
= ureg_negate(src
);
4223 if (src_reg
->reladdr
!= NULL
) {
4224 /* Normally ureg_src_indirect() would be used here, but a stupid compiler
4225 * bug in g++ makes ureg_src_indirect (an inline C function) erroneously
4226 * set the bit for src.Negate. So we have to do the operation manually
4227 * here to work around the compiler's problems. */
4228 /*src = ureg_src_indirect(src, ureg_src(t->address[0]));*/
4229 struct ureg_src addr
= ureg_src(t
->address
[0]);
4231 src
.IndirectFile
= addr
.File
;
4232 src
.IndirectIndex
= addr
.Index
;
4233 src
.IndirectSwizzle
= addr
.SwizzleX
;
4235 if (src_reg
->file
!= PROGRAM_INPUT
&&
4236 src_reg
->file
!= PROGRAM_OUTPUT
) {
4237 /* If src_reg->index was negative, it was set to zero in
4238 * src_register(). Reassign it now. But don't do this
4239 * for input/output regs since they get remapped while
4240 * const buffers don't.
4242 src
.Index
= src_reg
->index
;
4250 compile_tgsi_instruction(struct st_translate
*t
,
4251 const glsl_to_tgsi_instruction
*inst
)
4253 struct ureg_program
*ureg
= t
->ureg
;
4255 struct ureg_dst dst
[1];
4256 struct ureg_src src
[4];
4260 num_dst
= num_inst_dst_regs(inst
->op
);
4261 num_src
= num_inst_src_regs(inst
->op
);
4264 dst
[0] = translate_dst(t
,
4268 for (i
= 0; i
< num_src
; i
++)
4269 src
[i
] = translate_src(t
, &inst
->src
[i
]);
4272 case TGSI_OPCODE_BGNLOOP
:
4273 case TGSI_OPCODE_CAL
:
4274 case TGSI_OPCODE_ELSE
:
4275 case TGSI_OPCODE_ENDLOOP
:
4276 case TGSI_OPCODE_IF
:
4277 assert(num_dst
== 0);
4278 ureg_label_insn(ureg
,
4282 inst
->op
== TGSI_OPCODE_CAL
? inst
->function
->sig_id
: 0));
4285 case TGSI_OPCODE_TEX
:
4286 case TGSI_OPCODE_TXB
:
4287 case TGSI_OPCODE_TXD
:
4288 case TGSI_OPCODE_TXL
:
4289 case TGSI_OPCODE_TXP
:
4290 case TGSI_OPCODE_TXQ
:
4291 case TGSI_OPCODE_TXF
:
4292 src
[num_src
++] = t
->samplers
[inst
->sampler
];
4296 translate_texture_target(inst
->tex_target
, inst
->tex_shadow
),
4300 case TGSI_OPCODE_SCS
:
4301 dst
[0] = ureg_writemask(dst
[0], TGSI_WRITEMASK_XY
);
4302 ureg_insn(ureg
, inst
->op
, dst
, num_dst
, src
, num_src
);
4315 * Emit the TGSI instructions to adjust the WPOS pixel center convention
4316 * Basically, add (adjX, adjY) to the fragment position.
4319 emit_adjusted_wpos(struct st_translate
*t
,
4320 const struct gl_program
*program
,
4321 float adjX
, float adjY
)
4323 struct ureg_program
*ureg
= t
->ureg
;
4324 struct ureg_dst wpos_temp
= ureg_DECL_temporary(ureg
);
4325 struct ureg_src wpos_input
= t
->inputs
[t
->inputMapping
[FRAG_ATTRIB_WPOS
]];
4327 /* Note that we bias X and Y and pass Z and W through unchanged.
4328 * The shader might also use gl_FragCoord.w and .z.
4330 ureg_ADD(ureg
, wpos_temp
, wpos_input
,
4331 ureg_imm4f(ureg
, adjX
, adjY
, 0.0f
, 0.0f
));
4333 t
->inputs
[t
->inputMapping
[FRAG_ATTRIB_WPOS
]] = ureg_src(wpos_temp
);
4338 * Emit the TGSI instructions for inverting the WPOS y coordinate.
4339 * This code is unavoidable because it also depends on whether
4340 * a FBO is bound (STATE_FB_WPOS_Y_TRANSFORM).
4343 emit_wpos_inversion(struct st_translate
*t
,
4344 const struct gl_program
*program
,
4347 struct ureg_program
*ureg
= t
->ureg
;
4349 /* Fragment program uses fragment position input.
4350 * Need to replace instances of INPUT[WPOS] with temp T
4351 * where T = INPUT[WPOS] by y is inverted.
4353 static const gl_state_index wposTransformState
[STATE_LENGTH
]
4354 = { STATE_INTERNAL
, STATE_FB_WPOS_Y_TRANSFORM
,
4355 (gl_state_index
)0, (gl_state_index
)0, (gl_state_index
)0 };
4357 /* XXX: note we are modifying the incoming shader here! Need to
4358 * do this before emitting the constant decls below, or this
4361 unsigned wposTransConst
= _mesa_add_state_reference(program
->Parameters
,
4362 wposTransformState
);
4364 struct ureg_src wpostrans
= ureg_DECL_constant(ureg
, wposTransConst
);
4365 struct ureg_dst wpos_temp
;
4366 struct ureg_src wpos_input
= t
->inputs
[t
->inputMapping
[FRAG_ATTRIB_WPOS
]];
4368 /* MOV wpos_temp, input[wpos]
4370 if (wpos_input
.File
== TGSI_FILE_TEMPORARY
)
4371 wpos_temp
= ureg_dst(wpos_input
);
4373 wpos_temp
= ureg_DECL_temporary(ureg
);
4374 ureg_MOV(ureg
, wpos_temp
, wpos_input
);
4378 /* MAD wpos_temp.y, wpos_input, wpostrans.xxxx, wpostrans.yyyy
4381 ureg_writemask(wpos_temp
, TGSI_WRITEMASK_Y
),
4383 ureg_scalar(wpostrans
, 0),
4384 ureg_scalar(wpostrans
, 1));
4386 /* MAD wpos_temp.y, wpos_input, wpostrans.zzzz, wpostrans.wwww
4389 ureg_writemask(wpos_temp
, TGSI_WRITEMASK_Y
),
4391 ureg_scalar(wpostrans
, 2),
4392 ureg_scalar(wpostrans
, 3));
4395 /* Use wpos_temp as position input from here on:
4397 t
->inputs
[t
->inputMapping
[FRAG_ATTRIB_WPOS
]] = ureg_src(wpos_temp
);
4402 * Emit fragment position/ooordinate code.
4405 emit_wpos(struct st_context
*st
,
4406 struct st_translate
*t
,
4407 const struct gl_program
*program
,
4408 struct ureg_program
*ureg
)
4410 const struct gl_fragment_program
*fp
=
4411 (const struct gl_fragment_program
*) program
;
4412 struct pipe_screen
*pscreen
= st
->pipe
->screen
;
4413 boolean invert
= FALSE
;
4415 if (fp
->OriginUpperLeft
) {
4416 /* Fragment shader wants origin in upper-left */
4417 if (pscreen
->get_param(pscreen
, PIPE_CAP_TGSI_FS_COORD_ORIGIN_UPPER_LEFT
)) {
4418 /* the driver supports upper-left origin */
4420 else if (pscreen
->get_param(pscreen
, PIPE_CAP_TGSI_FS_COORD_ORIGIN_LOWER_LEFT
)) {
4421 /* the driver supports lower-left origin, need to invert Y */
4422 ureg_property_fs_coord_origin(ureg
, TGSI_FS_COORD_ORIGIN_LOWER_LEFT
);
4429 /* Fragment shader wants origin in lower-left */
4430 if (pscreen
->get_param(pscreen
, PIPE_CAP_TGSI_FS_COORD_ORIGIN_LOWER_LEFT
))
4431 /* the driver supports lower-left origin */
4432 ureg_property_fs_coord_origin(ureg
, TGSI_FS_COORD_ORIGIN_LOWER_LEFT
);
4433 else if (pscreen
->get_param(pscreen
, PIPE_CAP_TGSI_FS_COORD_ORIGIN_UPPER_LEFT
))
4434 /* the driver supports upper-left origin, need to invert Y */
4440 if (fp
->PixelCenterInteger
) {
4441 /* Fragment shader wants pixel center integer */
4442 if (pscreen
->get_param(pscreen
, PIPE_CAP_TGSI_FS_COORD_PIXEL_CENTER_INTEGER
))
4443 /* the driver supports pixel center integer */
4444 ureg_property_fs_coord_pixel_center(ureg
, TGSI_FS_COORD_PIXEL_CENTER_INTEGER
);
4445 else if (pscreen
->get_param(pscreen
, PIPE_CAP_TGSI_FS_COORD_PIXEL_CENTER_HALF_INTEGER
))
4446 /* the driver supports pixel center half integer, need to bias X,Y */
4447 emit_adjusted_wpos(t
, program
, 0.5f
, invert
? 0.5f
: -0.5f
);
4452 /* Fragment shader wants pixel center half integer */
4453 if (pscreen
->get_param(pscreen
, PIPE_CAP_TGSI_FS_COORD_PIXEL_CENTER_HALF_INTEGER
)) {
4454 /* the driver supports pixel center half integer */
4456 else if (pscreen
->get_param(pscreen
, PIPE_CAP_TGSI_FS_COORD_PIXEL_CENTER_INTEGER
)) {
4457 /* the driver supports pixel center integer, need to bias X,Y */
4458 ureg_property_fs_coord_pixel_center(ureg
, TGSI_FS_COORD_PIXEL_CENTER_INTEGER
);
4459 emit_adjusted_wpos(t
, program
, 0.5f
, invert
? -0.5f
: 0.5f
);
4465 /* we invert after adjustment so that we avoid the MOV to temporary,
4466 * and reuse the adjustment ADD instead */
4467 emit_wpos_inversion(t
, program
, invert
);
4471 * OpenGL's fragment gl_FrontFace input is 1 for front-facing, 0 for back.
4472 * TGSI uses +1 for front, -1 for back.
4473 * This function converts the TGSI value to the GL value. Simply clamping/
4474 * saturating the value to [0,1] does the job.
4477 emit_face_var(struct st_translate
*t
)
4479 struct ureg_program
*ureg
= t
->ureg
;
4480 struct ureg_dst face_temp
= ureg_DECL_temporary(ureg
);
4481 struct ureg_src face_input
= t
->inputs
[t
->inputMapping
[FRAG_ATTRIB_FACE
]];
4483 /* MOV_SAT face_temp, input[face] */
4484 face_temp
= ureg_saturate(face_temp
);
4485 ureg_MOV(ureg
, face_temp
, face_input
);
4487 /* Use face_temp as face input from here on: */
4488 t
->inputs
[t
->inputMapping
[FRAG_ATTRIB_FACE
]] = ureg_src(face_temp
);
4492 emit_edgeflags(struct st_translate
*t
)
4494 struct ureg_program
*ureg
= t
->ureg
;
4495 struct ureg_dst edge_dst
= t
->outputs
[t
->outputMapping
[VERT_RESULT_EDGE
]];
4496 struct ureg_src edge_src
= t
->inputs
[t
->inputMapping
[VERT_ATTRIB_EDGEFLAG
]];
4498 ureg_MOV(ureg
, edge_dst
, edge_src
);
4502 * Translate intermediate IR (glsl_to_tgsi_instruction) to TGSI format.
4503 * \param program the program to translate
4504 * \param numInputs number of input registers used
4505 * \param inputMapping maps Mesa fragment program inputs to TGSI generic
4507 * \param inputSemanticName the TGSI_SEMANTIC flag for each input
4508 * \param inputSemanticIndex the semantic index (ex: which texcoord) for
4510 * \param interpMode the TGSI_INTERPOLATE_LINEAR/PERSP mode for each input
4511 * \param numOutputs number of output registers used
4512 * \param outputMapping maps Mesa fragment program outputs to TGSI
4514 * \param outputSemanticName the TGSI_SEMANTIC flag for each output
4515 * \param outputSemanticIndex the semantic index (ex: which texcoord) for
4518 * \return PIPE_OK or PIPE_ERROR_OUT_OF_MEMORY
4520 extern "C" enum pipe_error
4521 st_translate_program(
4522 struct gl_context
*ctx
,
4524 struct ureg_program
*ureg
,
4525 glsl_to_tgsi_visitor
*program
,
4526 const struct gl_program
*proginfo
,
4528 const GLuint inputMapping
[],
4529 const ubyte inputSemanticName
[],
4530 const ubyte inputSemanticIndex
[],
4531 const GLuint interpMode
[],
4533 const GLuint outputMapping
[],
4534 const ubyte outputSemanticName
[],
4535 const ubyte outputSemanticIndex
[],
4536 boolean passthrough_edgeflags
)
4538 struct st_translate translate
, *t
;
4540 enum pipe_error ret
= PIPE_OK
;
4542 assert(numInputs
<= Elements(t
->inputs
));
4543 assert(numOutputs
<= Elements(t
->outputs
));
4546 memset(t
, 0, sizeof *t
);
4548 t
->procType
= procType
;
4549 t
->inputMapping
= inputMapping
;
4550 t
->outputMapping
= outputMapping
;
4552 t
->pointSizeOutIndex
= -1;
4553 t
->prevInstWrotePointSize
= GL_FALSE
;
4556 * Declare input attributes.
4558 if (procType
== TGSI_PROCESSOR_FRAGMENT
) {
4559 for (i
= 0; i
< numInputs
; i
++) {
4560 t
->inputs
[i
] = ureg_DECL_fs_input(ureg
,
4561 inputSemanticName
[i
],
4562 inputSemanticIndex
[i
],
4566 if (proginfo
->InputsRead
& FRAG_BIT_WPOS
) {
4567 /* Must do this after setting up t->inputs, and before
4568 * emitting constant references, below:
4570 emit_wpos(st_context(ctx
), t
, proginfo
, ureg
);
4573 if (proginfo
->InputsRead
& FRAG_BIT_FACE
)
4577 * Declare output attributes.
4579 for (i
= 0; i
< numOutputs
; i
++) {
4580 switch (outputSemanticName
[i
]) {
4581 case TGSI_SEMANTIC_POSITION
:
4582 t
->outputs
[i
] = ureg_DECL_output(ureg
,
4583 TGSI_SEMANTIC_POSITION
, /* Z/Depth */
4584 outputSemanticIndex
[i
]);
4585 t
->outputs
[i
] = ureg_writemask(t
->outputs
[i
], TGSI_WRITEMASK_Z
);
4587 case TGSI_SEMANTIC_STENCIL
:
4588 t
->outputs
[i
] = ureg_DECL_output(ureg
,
4589 TGSI_SEMANTIC_STENCIL
, /* Stencil */
4590 outputSemanticIndex
[i
]);
4591 t
->outputs
[i
] = ureg_writemask(t
->outputs
[i
], TGSI_WRITEMASK_Y
);
4593 case TGSI_SEMANTIC_COLOR
:
4594 t
->outputs
[i
] = ureg_DECL_output(ureg
,
4595 TGSI_SEMANTIC_COLOR
,
4596 outputSemanticIndex
[i
]);
4599 assert(!"fragment shader outputs must be POSITION/STENCIL/COLOR");
4600 return PIPE_ERROR_BAD_INPUT
;
4604 else if (procType
== TGSI_PROCESSOR_GEOMETRY
) {
4605 for (i
= 0; i
< numInputs
; i
++) {
4606 t
->inputs
[i
] = ureg_DECL_gs_input(ureg
,
4608 inputSemanticName
[i
],
4609 inputSemanticIndex
[i
]);
4612 for (i
= 0; i
< numOutputs
; i
++) {
4613 t
->outputs
[i
] = ureg_DECL_output(ureg
,
4614 outputSemanticName
[i
],
4615 outputSemanticIndex
[i
]);
4619 assert(procType
== TGSI_PROCESSOR_VERTEX
);
4621 for (i
= 0; i
< numInputs
; i
++) {
4622 t
->inputs
[i
] = ureg_DECL_vs_input(ureg
, i
);
4625 for (i
= 0; i
< numOutputs
; i
++) {
4626 t
->outputs
[i
] = ureg_DECL_output(ureg
,
4627 outputSemanticName
[i
],
4628 outputSemanticIndex
[i
]);
4629 if ((outputSemanticName
[i
] == TGSI_SEMANTIC_PSIZE
) && proginfo
->Id
) {
4630 /* Writing to the point size result register requires special
4631 * handling to implement clamping.
4633 static const gl_state_index pointSizeClampState
[STATE_LENGTH
]
4634 = { STATE_INTERNAL
, STATE_POINT_SIZE_IMPL_CLAMP
, (gl_state_index
)0, (gl_state_index
)0, (gl_state_index
)0 };
4635 /* XXX: note we are modifying the incoming shader here! Need to
4636 * do this before emitting the constant decls below, or this
4639 unsigned pointSizeClampConst
=
4640 _mesa_add_state_reference(proginfo
->Parameters
,
4641 pointSizeClampState
);
4642 struct ureg_dst psizregtemp
= ureg_DECL_temporary(ureg
);
4643 t
->pointSizeConst
= ureg_DECL_constant(ureg
, pointSizeClampConst
);
4644 t
->pointSizeResult
= t
->outputs
[i
];
4645 t
->pointSizeOutIndex
= i
;
4646 t
->outputs
[i
] = psizregtemp
;
4649 if (passthrough_edgeflags
)
4653 /* Declare address register.
4655 if (program
->num_address_regs
> 0) {
4656 assert(program
->num_address_regs
== 1);
4657 t
->address
[0] = ureg_DECL_address(ureg
);
4660 /* Declare misc input registers
4663 GLbitfield sysInputs
= proginfo
->SystemValuesRead
;
4664 unsigned numSys
= 0;
4665 for (i
= 0; sysInputs
; i
++) {
4666 if (sysInputs
& (1 << i
)) {
4667 unsigned semName
= mesa_sysval_to_semantic
[i
];
4668 t
->systemValues
[i
] = ureg_DECL_system_value(ureg
, numSys
, semName
, 0);
4670 sysInputs
&= ~(1 << i
);
4675 if (program
->indirect_addr_temps
) {
4676 /* If temps are accessed with indirect addressing, declare temporaries
4677 * in sequential order. Else, we declare them on demand elsewhere.
4678 * (Note: the number of temporaries is equal to program->next_temp)
4680 for (i
= 0; i
< (unsigned)program
->next_temp
; i
++) {
4681 /* XXX use TGSI_FILE_TEMPORARY_ARRAY when it's supported by ureg */
4682 t
->temps
[i
] = ureg_DECL_temporary(t
->ureg
);
4686 /* Emit constants and uniforms. TGSI uses a single index space for these,
4687 * so we put all the translated regs in t->constants.
4689 if (proginfo
->Parameters
) {
4690 t
->constants
= (struct ureg_src
*)CALLOC(proginfo
->Parameters
->NumParameters
* sizeof(t
->constants
[0]));
4691 if (t
->constants
== NULL
) {
4692 ret
= PIPE_ERROR_OUT_OF_MEMORY
;
4696 for (i
= 0; i
< proginfo
->Parameters
->NumParameters
; i
++) {
4697 switch (proginfo
->Parameters
->Parameters
[i
].Type
) {
4698 case PROGRAM_ENV_PARAM
:
4699 case PROGRAM_LOCAL_PARAM
:
4700 case PROGRAM_STATE_VAR
:
4701 case PROGRAM_NAMED_PARAM
:
4702 case PROGRAM_UNIFORM
:
4703 t
->constants
[i
] = ureg_DECL_constant(ureg
, i
);
4706 /* Emit immediates for PROGRAM_CONSTANT only when there's no indirect
4707 * addressing of the const buffer.
4708 * FIXME: Be smarter and recognize param arrays:
4709 * indirect addressing is only valid within the referenced
4712 case PROGRAM_CONSTANT
:
4713 if (program
->indirect_addr_consts
)
4714 t
->constants
[i
] = ureg_DECL_constant(ureg
, i
);
4716 t
->constants
[i
] = emit_immediate(t
,
4717 proginfo
->Parameters
->ParameterValues
[i
],
4718 proginfo
->Parameters
->Parameters
[i
].DataType
,
4727 /* Emit immediate values.
4729 t
->immediates
= (struct ureg_src
*)CALLOC(program
->num_immediates
* sizeof(struct ureg_src
));
4730 if (t
->immediates
== NULL
) {
4731 ret
= PIPE_ERROR_OUT_OF_MEMORY
;
4735 foreach_iter(exec_list_iterator
, iter
, program
->immediates
) {
4736 immediate_storage
*imm
= (immediate_storage
*)iter
.get();
4737 t
->immediates
[i
++] = emit_immediate(t
, imm
->values
, imm
->type
, imm
->size
);
4740 /* texture samplers */
4741 for (i
= 0; i
< ctx
->Const
.MaxTextureImageUnits
; i
++) {
4742 if (program
->samplers_used
& (1 << i
)) {
4743 t
->samplers
[i
] = ureg_DECL_sampler(ureg
, i
);
4747 /* Emit each instruction in turn:
4749 foreach_iter(exec_list_iterator
, iter
, program
->instructions
) {
4750 set_insn_start(t
, ureg_get_instruction_number(ureg
));
4751 compile_tgsi_instruction(t
, (glsl_to_tgsi_instruction
*)iter
.get());
4753 if (t
->prevInstWrotePointSize
&& proginfo
->Id
) {
4754 /* The previous instruction wrote to the (fake) vertex point size
4755 * result register. Now we need to clamp that value to the min/max
4756 * point size range, putting the result into the real point size
4758 * Note that we can't do this easily at the end of program due to
4759 * possible early return.
4761 set_insn_start(t
, ureg_get_instruction_number(ureg
));
4763 ureg_writemask(t
->outputs
[t
->pointSizeOutIndex
], WRITEMASK_X
),
4764 ureg_src(t
->outputs
[t
->pointSizeOutIndex
]),
4765 ureg_swizzle(t
->pointSizeConst
, 1,1,1,1));
4766 ureg_MIN(t
->ureg
, ureg_writemask(t
->pointSizeResult
, WRITEMASK_X
),
4767 ureg_src(t
->outputs
[t
->pointSizeOutIndex
]),
4768 ureg_swizzle(t
->pointSizeConst
, 2,2,2,2));
4770 t
->prevInstWrotePointSize
= GL_FALSE
;
4773 /* Fix up all emitted labels:
4775 for (i
= 0; i
< t
->labels_count
; i
++) {
4776 ureg_fixup_label(ureg
, t
->labels
[i
].token
,
4777 t
->insn
[t
->labels
[i
].branch_target
]);
4784 FREE(t
->immediates
);
4787 debug_printf("%s: translate error flag set\n", __FUNCTION__
);
4792 /* ----------------------------- End TGSI code ------------------------------ */
4795 * Convert a shader's GLSL IR into a Mesa gl_program, although without
4796 * generating Mesa IR.
4798 static struct gl_program
*
4799 get_mesa_program(struct gl_context
*ctx
,
4800 struct gl_shader_program
*shader_program
,
4801 struct gl_shader
*shader
)
4803 glsl_to_tgsi_visitor
* v
= new glsl_to_tgsi_visitor();
4804 struct gl_program
*prog
;
4806 const char *target_string
;
4808 struct gl_shader_compiler_options
*options
=
4809 &ctx
->ShaderCompilerOptions
[_mesa_shader_type_to_index(shader
->Type
)];
4811 switch (shader
->Type
) {
4812 case GL_VERTEX_SHADER
:
4813 target
= GL_VERTEX_PROGRAM_ARB
;
4814 target_string
= "vertex";
4816 case GL_FRAGMENT_SHADER
:
4817 target
= GL_FRAGMENT_PROGRAM_ARB
;
4818 target_string
= "fragment";
4820 case GL_GEOMETRY_SHADER
:
4821 target
= GL_GEOMETRY_PROGRAM_NV
;
4822 target_string
= "geometry";
4825 assert(!"should not be reached");
4829 validate_ir_tree(shader
->ir
);
4831 prog
= ctx
->Driver
.NewProgram(ctx
, target
, shader_program
->Name
);
4834 prog
->Parameters
= _mesa_new_parameter_list();
4835 prog
->Varying
= _mesa_new_parameter_list();
4836 prog
->Attributes
= _mesa_new_parameter_list();
4839 v
->shader_program
= shader_program
;
4840 v
->options
= options
;
4841 v
->glsl_version
= ctx
->Const
.GLSLVersion
;
4842 v
->native_integers
= ctx
->Const
.NativeIntegers
;
4844 add_uniforms_to_parameters_list(shader_program
, shader
, prog
);
4846 /* Emit intermediate IR for main(). */
4847 visit_exec_list(shader
->ir
, v
);
4849 /* Now emit bodies for any functions that were used. */
4851 progress
= GL_FALSE
;
4853 foreach_iter(exec_list_iterator
, iter
, v
->function_signatures
) {
4854 function_entry
*entry
= (function_entry
*)iter
.get();
4856 if (!entry
->bgn_inst
) {
4857 v
->current_function
= entry
;
4859 entry
->bgn_inst
= v
->emit(NULL
, TGSI_OPCODE_BGNSUB
);
4860 entry
->bgn_inst
->function
= entry
;
4862 visit_exec_list(&entry
->sig
->body
, v
);
4864 glsl_to_tgsi_instruction
*last
;
4865 last
= (glsl_to_tgsi_instruction
*)v
->instructions
.get_tail();
4866 if (last
->op
!= TGSI_OPCODE_RET
)
4867 v
->emit(NULL
, TGSI_OPCODE_RET
);
4869 glsl_to_tgsi_instruction
*end
;
4870 end
= v
->emit(NULL
, TGSI_OPCODE_ENDSUB
);
4871 end
->function
= entry
;
4879 /* Print out some information (for debugging purposes) used by the
4880 * optimization passes. */
4881 for (i
=0; i
< v
->next_temp
; i
++) {
4882 int fr
= v
->get_first_temp_read(i
);
4883 int fw
= v
->get_first_temp_write(i
);
4884 int lr
= v
->get_last_temp_read(i
);
4885 int lw
= v
->get_last_temp_write(i
);
4887 printf("Temp %d: FR=%3d FW=%3d LR=%3d LW=%3d\n", i
, fr
, fw
, lr
, lw
);
4892 /* Remove reads to output registers, and to varyings in vertex shaders. */
4893 v
->remove_output_reads(PROGRAM_OUTPUT
);
4894 if (target
== GL_VERTEX_PROGRAM_ARB
)
4895 v
->remove_output_reads(PROGRAM_VARYING
);
4897 /* Perform optimizations on the instructions in the glsl_to_tgsi_visitor. */
4899 v
->copy_propagate();
4900 while (v
->eliminate_dead_code_advanced());
4902 /* FIXME: These passes to optimize temporary registers don't work when there
4903 * is indirect addressing of the temporary register space. We need proper
4904 * array support so that we don't have to give up these passes in every
4905 * shader that uses arrays.
4907 if (!v
->indirect_addr_temps
) {
4908 v
->eliminate_dead_code();
4909 v
->merge_registers();
4910 v
->renumber_registers();
4913 /* Write the END instruction. */
4914 v
->emit(NULL
, TGSI_OPCODE_END
);
4916 if (ctx
->Shader
.Flags
& GLSL_DUMP
) {
4918 printf("GLSL IR for linked %s program %d:\n", target_string
,
4919 shader_program
->Name
);
4920 _mesa_print_ir(shader
->ir
, NULL
);
4925 prog
->Instructions
= NULL
;
4926 prog
->NumInstructions
= 0;
4928 do_set_program_inouts(shader
->ir
, prog
);
4929 count_resources(v
, prog
);
4931 check_resources(ctx
, shader_program
, v
, prog
);
4933 _mesa_reference_program(ctx
, &shader
->Program
, prog
);
4935 struct st_vertex_program
*stvp
;
4936 struct st_fragment_program
*stfp
;
4937 struct st_geometry_program
*stgp
;
4939 switch (shader
->Type
) {
4940 case GL_VERTEX_SHADER
:
4941 stvp
= (struct st_vertex_program
*)prog
;
4942 stvp
->glsl_to_tgsi
= v
;
4944 case GL_FRAGMENT_SHADER
:
4945 stfp
= (struct st_fragment_program
*)prog
;
4946 stfp
->glsl_to_tgsi
= v
;
4948 case GL_GEOMETRY_SHADER
:
4949 stgp
= (struct st_geometry_program
*)prog
;
4950 stgp
->glsl_to_tgsi
= v
;
4953 assert(!"should not be reached");
4963 st_new_shader(struct gl_context
*ctx
, GLuint name
, GLuint type
)
4965 struct gl_shader
*shader
;
4966 assert(type
== GL_FRAGMENT_SHADER
|| type
== GL_VERTEX_SHADER
||
4967 type
== GL_GEOMETRY_SHADER_ARB
);
4968 shader
= rzalloc(NULL
, struct gl_shader
);
4970 shader
->Type
= type
;
4971 shader
->Name
= name
;
4972 _mesa_init_shader(ctx
, shader
);
4977 struct gl_shader_program
*
4978 st_new_shader_program(struct gl_context
*ctx
, GLuint name
)
4980 struct gl_shader_program
*shProg
;
4981 shProg
= rzalloc(NULL
, struct gl_shader_program
);
4983 shProg
->Name
= name
;
4984 _mesa_init_shader_program(ctx
, shProg
);
4991 * Called via ctx->Driver.LinkShader()
4992 * This actually involves converting GLSL IR into an intermediate TGSI-like IR
4993 * with code lowering and other optimizations.
4996 st_link_shader(struct gl_context
*ctx
, struct gl_shader_program
*prog
)
4998 assert(prog
->LinkStatus
);
5000 for (unsigned i
= 0; i
< MESA_SHADER_TYPES
; i
++) {
5001 if (prog
->_LinkedShaders
[i
] == NULL
)
5005 exec_list
*ir
= prog
->_LinkedShaders
[i
]->ir
;
5006 const struct gl_shader_compiler_options
*options
=
5007 &ctx
->ShaderCompilerOptions
[_mesa_shader_type_to_index(prog
->_LinkedShaders
[i
]->Type
)];
5013 do_mat_op_to_vec(ir
);
5014 lower_instructions(ir
, (MOD_TO_FRACT
| DIV_TO_MUL_RCP
| EXP_TO_EXP2
5016 | ((options
->EmitNoPow
) ? POW_TO_EXP2
: 0)));
5018 progress
= do_lower_jumps(ir
, true, true, options
->EmitNoMainReturn
, options
->EmitNoCont
, options
->EmitNoLoops
) || progress
;
5020 progress
= do_common_optimization(ir
, true, options
->MaxUnrollIterations
) || progress
;
5022 progress
= lower_quadop_vector(ir
, false) || progress
;
5024 if (options
->EmitNoIfs
) {
5025 progress
= lower_discard(ir
) || progress
;
5026 progress
= lower_if_to_cond_assign(ir
) || progress
;
5029 if (options
->EmitNoNoise
)
5030 progress
= lower_noise(ir
) || progress
;
5032 /* If there are forms of indirect addressing that the driver
5033 * cannot handle, perform the lowering pass.
5035 if (options
->EmitNoIndirectInput
|| options
->EmitNoIndirectOutput
5036 || options
->EmitNoIndirectTemp
|| options
->EmitNoIndirectUniform
)
5038 lower_variable_index_to_cond_assign(ir
,
5039 options
->EmitNoIndirectInput
,
5040 options
->EmitNoIndirectOutput
,
5041 options
->EmitNoIndirectTemp
,
5042 options
->EmitNoIndirectUniform
)
5045 progress
= do_vec_index_to_cond_assign(ir
) || progress
;
5048 validate_ir_tree(ir
);
5051 for (unsigned i
= 0; i
< MESA_SHADER_TYPES
; i
++) {
5052 struct gl_program
*linked_prog
;
5054 if (prog
->_LinkedShaders
[i
] == NULL
)
5057 linked_prog
= get_mesa_program(ctx
, prog
, prog
->_LinkedShaders
[i
]);
5062 switch (prog
->_LinkedShaders
[i
]->Type
) {
5063 case GL_VERTEX_SHADER
:
5064 _mesa_reference_vertprog(ctx
, &prog
->VertexProgram
,
5065 (struct gl_vertex_program
*)linked_prog
);
5066 ok
= ctx
->Driver
.ProgramStringNotify(ctx
, GL_VERTEX_PROGRAM_ARB
,
5069 case GL_FRAGMENT_SHADER
:
5070 _mesa_reference_fragprog(ctx
, &prog
->FragmentProgram
,
5071 (struct gl_fragment_program
*)linked_prog
);
5072 ok
= ctx
->Driver
.ProgramStringNotify(ctx
, GL_FRAGMENT_PROGRAM_ARB
,
5075 case GL_GEOMETRY_SHADER
:
5076 _mesa_reference_geomprog(ctx
, &prog
->GeometryProgram
,
5077 (struct gl_geometry_program
*)linked_prog
);
5078 ok
= ctx
->Driver
.ProgramStringNotify(ctx
, GL_GEOMETRY_PROGRAM_NV
,
5087 _mesa_reference_program(ctx
, &linked_prog
, NULL
);
5095 * Link a GLSL shader program. Called via glLinkProgram().
5098 st_glsl_link_shader(struct gl_context
*ctx
, struct gl_shader_program
*prog
)
5102 _mesa_clear_shader_program_data(ctx
, prog
);
5104 prog
->LinkStatus
= GL_TRUE
;
5106 for (i
= 0; i
< prog
->NumShaders
; i
++) {
5107 if (!prog
->Shaders
[i
]->CompileStatus
) {
5108 fail_link(prog
, "linking with uncompiled shader");
5109 prog
->LinkStatus
= GL_FALSE
;
5113 prog
->Varying
= _mesa_new_parameter_list();
5114 _mesa_reference_vertprog(ctx
, &prog
->VertexProgram
, NULL
);
5115 _mesa_reference_fragprog(ctx
, &prog
->FragmentProgram
, NULL
);
5116 _mesa_reference_geomprog(ctx
, &prog
->GeometryProgram
, NULL
);
5118 if (prog
->LinkStatus
) {
5119 link_shaders(ctx
, prog
);
5122 if (prog
->LinkStatus
) {
5123 if (!ctx
->Driver
.LinkShader(ctx
, prog
)) {
5124 prog
->LinkStatus
= GL_FALSE
;
5128 set_uniform_initializers(ctx
, prog
);
5130 if (ctx
->Shader
.Flags
& GLSL_DUMP
) {
5131 if (!prog
->LinkStatus
) {
5132 printf("GLSL shader program %d failed to link\n", prog
->Name
);
5135 if (prog
->InfoLog
&& prog
->InfoLog
[0] != 0) {
5136 printf("GLSL shader program %d info log:\n", prog
->Name
);
5137 printf("%s\n", prog
->InfoLog
);