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 void 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 GLboolean
try_emit_mad(ir_expression
*ir
,
395 GLboolean
try_emit_sat(ir_expression
*ir
);
397 void emit_swz(ir_expression
*ir
);
399 bool process_move_condition(ir_rvalue
*ir
);
401 void remove_output_reads(gl_register_file type
);
402 void simplify_cmp(void);
404 void rename_temp_register(int index
, int new_index
);
405 int get_first_temp_read(int index
);
406 int get_first_temp_write(int index
);
407 int get_last_temp_read(int index
);
408 int get_last_temp_write(int index
);
410 void copy_propagate(void);
411 void eliminate_dead_code(void);
412 int eliminate_dead_code_advanced(void);
413 void merge_registers(void);
414 void renumber_registers(void);
419 static st_src_reg undef_src
= st_src_reg(PROGRAM_UNDEFINED
, 0, GLSL_TYPE_ERROR
);
421 static st_dst_reg undef_dst
= st_dst_reg(PROGRAM_UNDEFINED
, SWIZZLE_NOOP
, GLSL_TYPE_ERROR
);
423 static st_dst_reg address_reg
= st_dst_reg(PROGRAM_ADDRESS
, WRITEMASK_X
, GLSL_TYPE_FLOAT
);
426 fail_link(struct gl_shader_program
*prog
, const char *fmt
, ...) PRINTFLIKE(2, 3);
429 fail_link(struct gl_shader_program
*prog
, const char *fmt
, ...)
433 ralloc_vasprintf_append(&prog
->InfoLog
, fmt
, args
);
436 prog
->LinkStatus
= GL_FALSE
;
440 swizzle_for_size(int size
)
442 int size_swizzles
[4] = {
443 MAKE_SWIZZLE4(SWIZZLE_X
, SWIZZLE_X
, SWIZZLE_X
, SWIZZLE_X
),
444 MAKE_SWIZZLE4(SWIZZLE_X
, SWIZZLE_Y
, SWIZZLE_Y
, SWIZZLE_Y
),
445 MAKE_SWIZZLE4(SWIZZLE_X
, SWIZZLE_Y
, SWIZZLE_Z
, SWIZZLE_Z
),
446 MAKE_SWIZZLE4(SWIZZLE_X
, SWIZZLE_Y
, SWIZZLE_Z
, SWIZZLE_W
),
449 assert((size
>= 1) && (size
<= 4));
450 return size_swizzles
[size
- 1];
454 is_tex_instruction(unsigned opcode
)
456 const tgsi_opcode_info
* info
= tgsi_get_opcode_info(opcode
);
461 num_inst_dst_regs(unsigned opcode
)
463 const tgsi_opcode_info
* info
= tgsi_get_opcode_info(opcode
);
464 return info
->num_dst
;
468 num_inst_src_regs(unsigned opcode
)
470 const tgsi_opcode_info
* info
= tgsi_get_opcode_info(opcode
);
471 return info
->is_tex
? info
->num_src
- 1 : info
->num_src
;
474 glsl_to_tgsi_instruction
*
475 glsl_to_tgsi_visitor::emit(ir_instruction
*ir
, unsigned op
,
477 st_src_reg src0
, st_src_reg src1
, st_src_reg src2
)
479 glsl_to_tgsi_instruction
*inst
= new(mem_ctx
) glsl_to_tgsi_instruction();
480 int num_reladdr
= 0, i
;
482 op
= get_opcode(ir
, op
, dst
, src0
, src1
);
484 /* If we have to do relative addressing, we want to load the ARL
485 * reg directly for one of the regs, and preload the other reladdr
486 * sources into temps.
488 num_reladdr
+= dst
.reladdr
!= NULL
;
489 num_reladdr
+= src0
.reladdr
!= NULL
;
490 num_reladdr
+= src1
.reladdr
!= NULL
;
491 num_reladdr
+= src2
.reladdr
!= NULL
;
493 reladdr_to_temp(ir
, &src2
, &num_reladdr
);
494 reladdr_to_temp(ir
, &src1
, &num_reladdr
);
495 reladdr_to_temp(ir
, &src0
, &num_reladdr
);
498 emit_arl(ir
, address_reg
, *dst
.reladdr
);
501 assert(num_reladdr
== 0);
511 inst
->function
= NULL
;
513 if (op
== TGSI_OPCODE_ARL
)
514 this->num_address_regs
= 1;
516 /* Update indirect addressing status used by TGSI */
519 case PROGRAM_TEMPORARY
:
520 this->indirect_addr_temps
= true;
522 case PROGRAM_LOCAL_PARAM
:
523 case PROGRAM_ENV_PARAM
:
524 case PROGRAM_STATE_VAR
:
525 case PROGRAM_NAMED_PARAM
:
526 case PROGRAM_CONSTANT
:
527 case PROGRAM_UNIFORM
:
528 this->indirect_addr_consts
= true;
530 case PROGRAM_IMMEDIATE
:
531 assert(!"immediates should not have indirect addressing");
538 for (i
=0; i
<3; i
++) {
539 if(inst
->src
[i
].reladdr
) {
540 switch(inst
->src
[i
].file
) {
541 case PROGRAM_TEMPORARY
:
542 this->indirect_addr_temps
= true;
544 case PROGRAM_LOCAL_PARAM
:
545 case PROGRAM_ENV_PARAM
:
546 case PROGRAM_STATE_VAR
:
547 case PROGRAM_NAMED_PARAM
:
548 case PROGRAM_CONSTANT
:
549 case PROGRAM_UNIFORM
:
550 this->indirect_addr_consts
= true;
552 case PROGRAM_IMMEDIATE
:
553 assert(!"immediates should not have indirect addressing");
562 this->instructions
.push_tail(inst
);
568 glsl_to_tgsi_instruction
*
569 glsl_to_tgsi_visitor::emit(ir_instruction
*ir
, unsigned op
,
570 st_dst_reg dst
, st_src_reg src0
, st_src_reg src1
)
572 return emit(ir
, op
, dst
, src0
, src1
, undef_src
);
575 glsl_to_tgsi_instruction
*
576 glsl_to_tgsi_visitor::emit(ir_instruction
*ir
, unsigned op
,
577 st_dst_reg dst
, st_src_reg src0
)
579 assert(dst
.writemask
!= 0);
580 return emit(ir
, op
, dst
, src0
, undef_src
, undef_src
);
583 glsl_to_tgsi_instruction
*
584 glsl_to_tgsi_visitor::emit(ir_instruction
*ir
, unsigned op
)
586 return emit(ir
, op
, undef_dst
, undef_src
, undef_src
, undef_src
);
590 * Determines whether to use an integer, unsigned integer, or float opcode
591 * based on the operands and input opcode, then emits the result.
593 * TODO: type checking for remaining TGSI opcodes
596 glsl_to_tgsi_visitor::get_opcode(ir_instruction
*ir
, unsigned op
,
598 st_src_reg src0
, st_src_reg src1
)
600 int type
= GLSL_TYPE_FLOAT
;
602 if (src0
.type
== GLSL_TYPE_FLOAT
|| src1
.type
== GLSL_TYPE_FLOAT
)
603 type
= GLSL_TYPE_FLOAT
;
604 else if (native_integers
)
607 #define case4(c, f, i, u) \
608 case TGSI_OPCODE_##c: \
609 if (type == GLSL_TYPE_INT) op = TGSI_OPCODE_##i; \
610 else if (type == GLSL_TYPE_UINT) op = TGSI_OPCODE_##u; \
611 else op = TGSI_OPCODE_##f; \
613 #define case3(f, i, u) case4(f, f, i, u)
614 #define case2fi(f, i) case4(f, f, i, i)
615 #define case2iu(i, u) case4(i, LAST, i, u)
621 case3(DIV
, IDIV
, UDIV
);
622 case3(MAX
, IMAX
, UMAX
);
623 case3(MIN
, IMIN
, UMIN
);
628 case3(SGE
, ISGE
, USGE
);
629 case3(SLT
, ISLT
, USLT
);
641 assert(op
!= TGSI_OPCODE_LAST
);
646 glsl_to_tgsi_visitor::emit_dp(ir_instruction
*ir
,
647 st_dst_reg dst
, st_src_reg src0
, st_src_reg src1
,
650 static const unsigned dot_opcodes
[] = {
651 TGSI_OPCODE_DP2
, TGSI_OPCODE_DP3
, TGSI_OPCODE_DP4
654 emit(ir
, dot_opcodes
[elements
- 2], dst
, src0
, src1
);
658 * Emits TGSI scalar opcodes to produce unique answers across channels.
660 * Some TGSI opcodes are scalar-only, like ARB_fp/vp. The src X
661 * channel determines the result across all channels. So to do a vec4
662 * of this operation, we want to emit a scalar per source channel used
663 * to produce dest channels.
666 glsl_to_tgsi_visitor::emit_scalar(ir_instruction
*ir
, unsigned op
,
668 st_src_reg orig_src0
, st_src_reg orig_src1
)
671 int done_mask
= ~dst
.writemask
;
673 /* TGSI RCP is a scalar operation splatting results to all channels,
674 * like ARB_fp/vp. So emit as many RCPs as necessary to cover our
677 for (i
= 0; i
< 4; i
++) {
678 GLuint this_mask
= (1 << i
);
679 glsl_to_tgsi_instruction
*inst
;
680 st_src_reg src0
= orig_src0
;
681 st_src_reg src1
= orig_src1
;
683 if (done_mask
& this_mask
)
686 GLuint src0_swiz
= GET_SWZ(src0
.swizzle
, i
);
687 GLuint src1_swiz
= GET_SWZ(src1
.swizzle
, i
);
688 for (j
= i
+ 1; j
< 4; j
++) {
689 /* If there is another enabled component in the destination that is
690 * derived from the same inputs, generate its value on this pass as
693 if (!(done_mask
& (1 << j
)) &&
694 GET_SWZ(src0
.swizzle
, j
) == src0_swiz
&&
695 GET_SWZ(src1
.swizzle
, j
) == src1_swiz
) {
696 this_mask
|= (1 << j
);
699 src0
.swizzle
= MAKE_SWIZZLE4(src0_swiz
, src0_swiz
,
700 src0_swiz
, src0_swiz
);
701 src1
.swizzle
= MAKE_SWIZZLE4(src1_swiz
, src1_swiz
,
702 src1_swiz
, src1_swiz
);
704 inst
= emit(ir
, op
, dst
, src0
, src1
);
705 inst
->dst
.writemask
= this_mask
;
706 done_mask
|= this_mask
;
711 glsl_to_tgsi_visitor::emit_scalar(ir_instruction
*ir
, unsigned op
,
712 st_dst_reg dst
, st_src_reg src0
)
714 st_src_reg undef
= undef_src
;
716 undef
.swizzle
= SWIZZLE_XXXX
;
718 emit_scalar(ir
, op
, dst
, src0
, undef
);
722 glsl_to_tgsi_visitor::emit_arl(ir_instruction
*ir
,
723 st_dst_reg dst
, st_src_reg src0
)
725 st_src_reg tmp
= get_temp(glsl_type::float_type
);
727 if (src0
.type
== GLSL_TYPE_INT
)
728 emit(NULL
, TGSI_OPCODE_I2F
, st_dst_reg(tmp
), src0
);
729 else if (src0
.type
== GLSL_TYPE_UINT
)
730 emit(NULL
, TGSI_OPCODE_U2F
, st_dst_reg(tmp
), src0
);
734 emit(NULL
, TGSI_OPCODE_ARL
, dst
, tmp
);
738 * Emit an TGSI_OPCODE_SCS instruction
740 * The \c SCS opcode functions a bit differently than the other TGSI opcodes.
741 * Instead of splatting its result across all four components of the
742 * destination, it writes one value to the \c x component and another value to
743 * the \c y component.
745 * \param ir IR instruction being processed
746 * \param op Either \c TGSI_OPCODE_SIN or \c TGSI_OPCODE_COS depending
747 * on which value is desired.
748 * \param dst Destination register
749 * \param src Source register
752 glsl_to_tgsi_visitor::emit_scs(ir_instruction
*ir
, unsigned op
,
754 const st_src_reg
&src
)
756 /* Vertex programs cannot use the SCS opcode.
758 if (this->prog
->Target
== GL_VERTEX_PROGRAM_ARB
) {
759 emit_scalar(ir
, op
, dst
, src
);
763 const unsigned component
= (op
== TGSI_OPCODE_SIN
) ? 0 : 1;
764 const unsigned scs_mask
= (1U << component
);
765 int done_mask
= ~dst
.writemask
;
768 assert(op
== TGSI_OPCODE_SIN
|| op
== TGSI_OPCODE_COS
);
770 /* If there are compnents in the destination that differ from the component
771 * that will be written by the SCS instrution, we'll need a temporary.
773 if (scs_mask
!= unsigned(dst
.writemask
)) {
774 tmp
= get_temp(glsl_type::vec4_type
);
777 for (unsigned i
= 0; i
< 4; i
++) {
778 unsigned this_mask
= (1U << i
);
779 st_src_reg src0
= src
;
781 if ((done_mask
& this_mask
) != 0)
784 /* The source swizzle specified which component of the source generates
785 * sine / cosine for the current component in the destination. The SCS
786 * instruction requires that this value be swizzle to the X component.
787 * Replace the current swizzle with a swizzle that puts the source in
790 unsigned src0_swiz
= GET_SWZ(src
.swizzle
, i
);
792 src0
.swizzle
= MAKE_SWIZZLE4(src0_swiz
, src0_swiz
,
793 src0_swiz
, src0_swiz
);
794 for (unsigned j
= i
+ 1; j
< 4; j
++) {
795 /* If there is another enabled component in the destination that is
796 * derived from the same inputs, generate its value on this pass as
799 if (!(done_mask
& (1 << j
)) &&
800 GET_SWZ(src0
.swizzle
, j
) == src0_swiz
) {
801 this_mask
|= (1 << j
);
805 if (this_mask
!= scs_mask
) {
806 glsl_to_tgsi_instruction
*inst
;
807 st_dst_reg tmp_dst
= st_dst_reg(tmp
);
809 /* Emit the SCS instruction.
811 inst
= emit(ir
, TGSI_OPCODE_SCS
, tmp_dst
, src0
);
812 inst
->dst
.writemask
= scs_mask
;
814 /* Move the result of the SCS instruction to the desired location in
817 tmp
.swizzle
= MAKE_SWIZZLE4(component
, component
,
818 component
, component
);
819 inst
= emit(ir
, TGSI_OPCODE_SCS
, dst
, tmp
);
820 inst
->dst
.writemask
= this_mask
;
822 /* Emit the SCS instruction to write directly to the destination.
824 glsl_to_tgsi_instruction
*inst
= emit(ir
, TGSI_OPCODE_SCS
, dst
, src0
);
825 inst
->dst
.writemask
= scs_mask
;
828 done_mask
|= this_mask
;
833 glsl_to_tgsi_visitor::add_constant(gl_register_file file
,
834 gl_constant_value values
[4], int size
, int datatype
,
837 if (file
== PROGRAM_CONSTANT
) {
838 return _mesa_add_typed_unnamed_constant(this->prog
->Parameters
, values
,
839 size
, datatype
, swizzle_out
);
842 immediate_storage
*entry
;
843 assert(file
== PROGRAM_IMMEDIATE
);
845 /* Search immediate storage to see if we already have an identical
846 * immediate that we can use instead of adding a duplicate entry.
848 foreach_iter(exec_list_iterator
, iter
, this->immediates
) {
849 entry
= (immediate_storage
*)iter
.get();
851 if (entry
->size
== size
&&
852 entry
->type
== datatype
&&
853 !memcmp(entry
->values
, values
, size
* sizeof(gl_constant_value
))) {
859 /* Add this immediate to the list. */
860 entry
= new(mem_ctx
) immediate_storage(values
, size
, datatype
);
861 this->immediates
.push_tail(entry
);
862 this->num_immediates
++;
868 glsl_to_tgsi_visitor::st_src_reg_for_float(float val
)
870 st_src_reg
src(PROGRAM_IMMEDIATE
, -1, GLSL_TYPE_FLOAT
);
871 union gl_constant_value uval
;
874 src
.index
= add_constant(src
.file
, &uval
, 1, GL_FLOAT
, &src
.swizzle
);
880 glsl_to_tgsi_visitor::st_src_reg_for_int(int val
)
882 st_src_reg
src(PROGRAM_IMMEDIATE
, -1, GLSL_TYPE_INT
);
883 union gl_constant_value uval
;
885 assert(native_integers
);
888 src
.index
= add_constant(src
.file
, &uval
, 1, GL_INT
, &src
.swizzle
);
894 glsl_to_tgsi_visitor::st_src_reg_for_type(int type
, int val
)
897 return type
== GLSL_TYPE_FLOAT
? st_src_reg_for_float(val
) :
898 st_src_reg_for_int(val
);
900 return st_src_reg_for_float(val
);
904 type_size(const struct glsl_type
*type
)
909 switch (type
->base_type
) {
912 case GLSL_TYPE_FLOAT
:
914 if (type
->is_matrix()) {
915 return type
->matrix_columns
;
917 /* Regardless of size of vector, it gets a vec4. This is bad
918 * packing for things like floats, but otherwise arrays become a
919 * mess. Hopefully a later pass over the code can pack scalars
920 * down if appropriate.
924 case GLSL_TYPE_ARRAY
:
925 assert(type
->length
> 0);
926 return type_size(type
->fields
.array
) * type
->length
;
927 case GLSL_TYPE_STRUCT
:
929 for (i
= 0; i
< type
->length
; i
++) {
930 size
+= type_size(type
->fields
.structure
[i
].type
);
933 case GLSL_TYPE_SAMPLER
:
934 /* Samplers take up one slot in UNIFORMS[], but they're baked in
945 * In the initial pass of codegen, we assign temporary numbers to
946 * intermediate results. (not SSA -- variable assignments will reuse
950 glsl_to_tgsi_visitor::get_temp(const glsl_type
*type
)
954 src
.type
= native_integers
? type
->base_type
: GLSL_TYPE_FLOAT
;
955 src
.file
= PROGRAM_TEMPORARY
;
956 src
.index
= next_temp
;
958 next_temp
+= type_size(type
);
960 if (type
->is_array() || type
->is_record()) {
961 src
.swizzle
= SWIZZLE_NOOP
;
963 src
.swizzle
= swizzle_for_size(type
->vector_elements
);
971 glsl_to_tgsi_visitor::find_variable_storage(ir_variable
*var
)
974 variable_storage
*entry
;
976 foreach_iter(exec_list_iterator
, iter
, this->variables
) {
977 entry
= (variable_storage
*)iter
.get();
979 if (entry
->var
== var
)
987 glsl_to_tgsi_visitor::visit(ir_variable
*ir
)
989 if (strcmp(ir
->name
, "gl_FragCoord") == 0) {
990 struct gl_fragment_program
*fp
= (struct gl_fragment_program
*)this->prog
;
992 fp
->OriginUpperLeft
= ir
->origin_upper_left
;
993 fp
->PixelCenterInteger
= ir
->pixel_center_integer
;
995 } else if (strcmp(ir
->name
, "gl_FragDepth") == 0) {
996 struct gl_fragment_program
*fp
= (struct gl_fragment_program
*)this->prog
;
997 switch (ir
->depth_layout
) {
998 case ir_depth_layout_none
:
999 fp
->FragDepthLayout
= FRAG_DEPTH_LAYOUT_NONE
;
1001 case ir_depth_layout_any
:
1002 fp
->FragDepthLayout
= FRAG_DEPTH_LAYOUT_ANY
;
1004 case ir_depth_layout_greater
:
1005 fp
->FragDepthLayout
= FRAG_DEPTH_LAYOUT_GREATER
;
1007 case ir_depth_layout_less
:
1008 fp
->FragDepthLayout
= FRAG_DEPTH_LAYOUT_LESS
;
1010 case ir_depth_layout_unchanged
:
1011 fp
->FragDepthLayout
= FRAG_DEPTH_LAYOUT_UNCHANGED
;
1019 if (ir
->mode
== ir_var_uniform
&& strncmp(ir
->name
, "gl_", 3) == 0) {
1021 const ir_state_slot
*const slots
= ir
->state_slots
;
1022 assert(ir
->state_slots
!= NULL
);
1024 /* Check if this statevar's setup in the STATE file exactly
1025 * matches how we'll want to reference it as a
1026 * struct/array/whatever. If not, then we need to move it into
1027 * temporary storage and hope that it'll get copy-propagated
1030 for (i
= 0; i
< ir
->num_state_slots
; i
++) {
1031 if (slots
[i
].swizzle
!= SWIZZLE_XYZW
) {
1036 struct variable_storage
*storage
;
1038 if (i
== ir
->num_state_slots
) {
1039 /* We'll set the index later. */
1040 storage
= new(mem_ctx
) variable_storage(ir
, PROGRAM_STATE_VAR
, -1);
1041 this->variables
.push_tail(storage
);
1045 /* The variable_storage constructor allocates slots based on the size
1046 * of the type. However, this had better match the number of state
1047 * elements that we're going to copy into the new temporary.
1049 assert((int) ir
->num_state_slots
== type_size(ir
->type
));
1051 storage
= new(mem_ctx
) variable_storage(ir
, PROGRAM_TEMPORARY
,
1053 this->variables
.push_tail(storage
);
1054 this->next_temp
+= type_size(ir
->type
);
1056 dst
= st_dst_reg(st_src_reg(PROGRAM_TEMPORARY
, storage
->index
,
1057 native_integers
? ir
->type
->base_type
: GLSL_TYPE_FLOAT
));
1061 for (unsigned int i
= 0; i
< ir
->num_state_slots
; i
++) {
1062 int index
= _mesa_add_state_reference(this->prog
->Parameters
,
1063 (gl_state_index
*)slots
[i
].tokens
);
1065 if (storage
->file
== PROGRAM_STATE_VAR
) {
1066 if (storage
->index
== -1) {
1067 storage
->index
= index
;
1069 assert(index
== storage
->index
+ (int)i
);
1072 st_src_reg
src(PROGRAM_STATE_VAR
, index
,
1073 native_integers
? ir
->type
->base_type
: GLSL_TYPE_FLOAT
);
1074 src
.swizzle
= slots
[i
].swizzle
;
1075 emit(ir
, TGSI_OPCODE_MOV
, dst
, src
);
1076 /* even a float takes up a whole vec4 reg in a struct/array. */
1081 if (storage
->file
== PROGRAM_TEMPORARY
&&
1082 dst
.index
!= storage
->index
+ (int) ir
->num_state_slots
) {
1083 fail_link(this->shader_program
,
1084 "failed to load builtin uniform `%s' (%d/%d regs loaded)\n",
1085 ir
->name
, dst
.index
- storage
->index
,
1086 type_size(ir
->type
));
1092 glsl_to_tgsi_visitor::visit(ir_loop
*ir
)
1094 ir_dereference_variable
*counter
= NULL
;
1096 if (ir
->counter
!= NULL
)
1097 counter
= new(ir
) ir_dereference_variable(ir
->counter
);
1099 if (ir
->from
!= NULL
) {
1100 assert(ir
->counter
!= NULL
);
1102 ir_assignment
*a
= new(ir
) ir_assignment(counter
, ir
->from
, NULL
);
1108 emit(NULL
, TGSI_OPCODE_BGNLOOP
);
1112 new(ir
) ir_expression(ir
->cmp
, glsl_type::bool_type
,
1114 ir_if
*if_stmt
= new(ir
) ir_if(e
);
1116 ir_loop_jump
*brk
= new(ir
) ir_loop_jump(ir_loop_jump::jump_break
);
1118 if_stmt
->then_instructions
.push_tail(brk
);
1120 if_stmt
->accept(this);
1127 visit_exec_list(&ir
->body_instructions
, this);
1129 if (ir
->increment
) {
1131 new(ir
) ir_expression(ir_binop_add
, counter
->type
,
1132 counter
, ir
->increment
);
1134 ir_assignment
*a
= new(ir
) ir_assignment(counter
, e
, NULL
);
1141 emit(NULL
, TGSI_OPCODE_ENDLOOP
);
1145 glsl_to_tgsi_visitor::visit(ir_loop_jump
*ir
)
1148 case ir_loop_jump::jump_break
:
1149 emit(NULL
, TGSI_OPCODE_BRK
);
1151 case ir_loop_jump::jump_continue
:
1152 emit(NULL
, TGSI_OPCODE_CONT
);
1159 glsl_to_tgsi_visitor::visit(ir_function_signature
*ir
)
1166 glsl_to_tgsi_visitor::visit(ir_function
*ir
)
1168 /* Ignore function bodies other than main() -- we shouldn't see calls to
1169 * them since they should all be inlined before we get to glsl_to_tgsi.
1171 if (strcmp(ir
->name
, "main") == 0) {
1172 const ir_function_signature
*sig
;
1175 sig
= ir
->matching_signature(&empty
);
1179 foreach_iter(exec_list_iterator
, iter
, sig
->body
) {
1180 ir_instruction
*ir
= (ir_instruction
*)iter
.get();
1188 glsl_to_tgsi_visitor::try_emit_mad(ir_expression
*ir
, int mul_operand
)
1190 int nonmul_operand
= 1 - mul_operand
;
1192 st_dst_reg result_dst
;
1194 ir_expression
*expr
= ir
->operands
[mul_operand
]->as_expression();
1195 if (!expr
|| expr
->operation
!= ir_binop_mul
)
1198 expr
->operands
[0]->accept(this);
1200 expr
->operands
[1]->accept(this);
1202 ir
->operands
[nonmul_operand
]->accept(this);
1205 this->result
= get_temp(ir
->type
);
1206 result_dst
= st_dst_reg(this->result
);
1207 result_dst
.writemask
= (1 << ir
->type
->vector_elements
) - 1;
1208 emit(ir
, TGSI_OPCODE_MAD
, result_dst
, a
, b
, c
);
1214 glsl_to_tgsi_visitor::try_emit_sat(ir_expression
*ir
)
1216 /* Saturates were only introduced to vertex programs in
1217 * NV_vertex_program3, so don't give them to drivers in the VP.
1219 if (this->prog
->Target
== GL_VERTEX_PROGRAM_ARB
)
1222 ir_rvalue
*sat_src
= ir
->as_rvalue_to_saturate();
1226 sat_src
->accept(this);
1227 st_src_reg src
= this->result
;
1229 /* If we generated an expression instruction into a temporary in
1230 * processing the saturate's operand, apply the saturate to that
1231 * instruction. Otherwise, generate a MOV to do the saturate.
1233 * Note that we have to be careful to only do this optimization if
1234 * the instruction in question was what generated src->result. For
1235 * example, ir_dereference_array might generate a MUL instruction
1236 * to create the reladdr, and return us a src reg using that
1237 * reladdr. That MUL result is not the value we're trying to
1240 ir_expression
*sat_src_expr
= sat_src
->as_expression();
1241 if (sat_src_expr
&& (sat_src_expr
->operation
== ir_binop_mul
||
1242 sat_src_expr
->operation
== ir_binop_add
||
1243 sat_src_expr
->operation
== ir_binop_dot
)) {
1244 glsl_to_tgsi_instruction
*new_inst
;
1245 new_inst
= (glsl_to_tgsi_instruction
*)this->instructions
.get_tail();
1246 new_inst
->saturate
= true;
1248 this->result
= get_temp(ir
->type
);
1249 st_dst_reg result_dst
= st_dst_reg(this->result
);
1250 result_dst
.writemask
= (1 << ir
->type
->vector_elements
) - 1;
1251 glsl_to_tgsi_instruction
*inst
;
1252 inst
= emit(ir
, TGSI_OPCODE_MOV
, result_dst
, src
);
1253 inst
->saturate
= true;
1260 glsl_to_tgsi_visitor::reladdr_to_temp(ir_instruction
*ir
,
1261 st_src_reg
*reg
, int *num_reladdr
)
1266 emit_arl(ir
, address_reg
, *reg
->reladdr
);
1268 if (*num_reladdr
!= 1) {
1269 st_src_reg temp
= get_temp(glsl_type::vec4_type
);
1271 emit(ir
, TGSI_OPCODE_MOV
, st_dst_reg(temp
), *reg
);
1279 glsl_to_tgsi_visitor::visit(ir_expression
*ir
)
1281 unsigned int operand
;
1282 st_src_reg op
[Elements(ir
->operands
)];
1283 st_src_reg result_src
;
1284 st_dst_reg result_dst
;
1286 /* Quick peephole: Emit MAD(a, b, c) instead of ADD(MUL(a, b), c)
1288 if (ir
->operation
== ir_binop_add
) {
1289 if (try_emit_mad(ir
, 1))
1291 if (try_emit_mad(ir
, 0))
1294 if (try_emit_sat(ir
))
1297 if (ir
->operation
== ir_quadop_vector
)
1298 assert(!"ir_quadop_vector should have been lowered");
1300 for (operand
= 0; operand
< ir
->get_num_operands(); operand
++) {
1301 this->result
.file
= PROGRAM_UNDEFINED
;
1302 ir
->operands
[operand
]->accept(this);
1303 if (this->result
.file
== PROGRAM_UNDEFINED
) {
1305 printf("Failed to get tree for expression operand:\n");
1306 ir
->operands
[operand
]->accept(&v
);
1309 op
[operand
] = this->result
;
1311 /* Matrix expression operands should have been broken down to vector
1312 * operations already.
1314 assert(!ir
->operands
[operand
]->type
->is_matrix());
1317 int vector_elements
= ir
->operands
[0]->type
->vector_elements
;
1318 if (ir
->operands
[1]) {
1319 vector_elements
= MAX2(vector_elements
,
1320 ir
->operands
[1]->type
->vector_elements
);
1323 this->result
.file
= PROGRAM_UNDEFINED
;
1325 /* Storage for our result. Ideally for an assignment we'd be using
1326 * the actual storage for the result here, instead.
1328 result_src
= get_temp(ir
->type
);
1329 /* convenience for the emit functions below. */
1330 result_dst
= st_dst_reg(result_src
);
1331 /* Limit writes to the channels that will be used by result_src later.
1332 * This does limit this temp's use as a temporary for multi-instruction
1335 result_dst
.writemask
= (1 << ir
->type
->vector_elements
) - 1;
1337 switch (ir
->operation
) {
1338 case ir_unop_logic_not
:
1339 emit(ir
, TGSI_OPCODE_SEQ
, result_dst
, op
[0], st_src_reg_for_type(result_dst
.type
, 0));
1342 assert(result_dst
.type
== GLSL_TYPE_FLOAT
|| result_dst
.type
== GLSL_TYPE_INT
);
1343 if (result_dst
.type
== GLSL_TYPE_INT
)
1344 emit(ir
, TGSI_OPCODE_INEG
, result_dst
, op
[0]);
1346 op
[0].negate
= ~op
[0].negate
;
1351 assert(result_dst
.type
== GLSL_TYPE_FLOAT
);
1352 emit(ir
, TGSI_OPCODE_ABS
, result_dst
, op
[0]);
1355 emit(ir
, TGSI_OPCODE_SSG
, result_dst
, op
[0]);
1358 emit_scalar(ir
, TGSI_OPCODE_RCP
, result_dst
, op
[0]);
1362 emit_scalar(ir
, TGSI_OPCODE_EX2
, result_dst
, op
[0]);
1366 assert(!"not reached: should be handled by ir_explog_to_explog2");
1369 emit_scalar(ir
, TGSI_OPCODE_LG2
, result_dst
, op
[0]);
1372 emit_scalar(ir
, TGSI_OPCODE_SIN
, result_dst
, op
[0]);
1375 emit_scalar(ir
, TGSI_OPCODE_COS
, result_dst
, op
[0]);
1377 case ir_unop_sin_reduced
:
1378 emit_scs(ir
, TGSI_OPCODE_SIN
, result_dst
, op
[0]);
1380 case ir_unop_cos_reduced
:
1381 emit_scs(ir
, TGSI_OPCODE_COS
, result_dst
, op
[0]);
1385 emit(ir
, TGSI_OPCODE_DDX
, result_dst
, op
[0]);
1388 op
[0].negate
= ~op
[0].negate
;
1389 emit(ir
, TGSI_OPCODE_DDY
, result_dst
, op
[0]);
1392 case ir_unop_noise
: {
1393 /* At some point, a motivated person could add a better
1394 * implementation of noise. Currently not even the nvidia
1395 * binary drivers do anything more than this. In any case, the
1396 * place to do this is in the GL state tracker, not the poor
1399 emit(ir
, TGSI_OPCODE_MOV
, result_dst
, st_src_reg_for_float(0.5));
1404 emit(ir
, TGSI_OPCODE_ADD
, result_dst
, op
[0], op
[1]);
1407 emit(ir
, TGSI_OPCODE_SUB
, result_dst
, op
[0], op
[1]);
1411 emit(ir
, TGSI_OPCODE_MUL
, result_dst
, op
[0], op
[1]);
1414 if (result_dst
.type
== GLSL_TYPE_FLOAT
)
1415 assert(!"not reached: should be handled by ir_div_to_mul_rcp");
1417 emit(ir
, TGSI_OPCODE_DIV
, result_dst
, op
[0], op
[1]);
1420 if (result_dst
.type
== GLSL_TYPE_FLOAT
)
1421 assert(!"ir_binop_mod should have been converted to b * fract(a/b)");
1423 emit(ir
, TGSI_OPCODE_MOD
, result_dst
, op
[0], op
[1]);
1427 emit(ir
, TGSI_OPCODE_SLT
, result_dst
, op
[0], op
[1]);
1429 case ir_binop_greater
:
1430 emit(ir
, TGSI_OPCODE_SGT
, result_dst
, op
[0], op
[1]);
1432 case ir_binop_lequal
:
1433 emit(ir
, TGSI_OPCODE_SLE
, result_dst
, op
[0], op
[1]);
1435 case ir_binop_gequal
:
1436 emit(ir
, TGSI_OPCODE_SGE
, result_dst
, op
[0], op
[1]);
1438 case ir_binop_equal
:
1439 emit(ir
, TGSI_OPCODE_SEQ
, result_dst
, op
[0], op
[1]);
1441 case ir_binop_nequal
:
1442 emit(ir
, TGSI_OPCODE_SNE
, result_dst
, op
[0], op
[1]);
1444 case ir_binop_all_equal
:
1445 /* "==" operator producing a scalar boolean. */
1446 if (ir
->operands
[0]->type
->is_vector() ||
1447 ir
->operands
[1]->type
->is_vector()) {
1448 st_src_reg temp
= get_temp(native_integers
?
1449 glsl_type::get_instance(ir
->operands
[0]->type
->base_type
, 4, 1) :
1450 glsl_type::vec4_type
);
1451 assert(ir
->operands
[0]->type
->base_type
== GLSL_TYPE_FLOAT
);
1452 emit(ir
, TGSI_OPCODE_SNE
, st_dst_reg(temp
), op
[0], op
[1]);
1453 emit_dp(ir
, result_dst
, temp
, temp
, vector_elements
);
1454 emit(ir
, TGSI_OPCODE_SEQ
, result_dst
, result_src
, st_src_reg_for_float(0.0));
1456 emit(ir
, TGSI_OPCODE_SEQ
, result_dst
, op
[0], op
[1]);
1459 case ir_binop_any_nequal
:
1460 /* "!=" operator producing a scalar boolean. */
1461 if (ir
->operands
[0]->type
->is_vector() ||
1462 ir
->operands
[1]->type
->is_vector()) {
1463 st_src_reg temp
= get_temp(native_integers
?
1464 glsl_type::get_instance(ir
->operands
[0]->type
->base_type
, 4, 1) :
1465 glsl_type::vec4_type
);
1466 assert(ir
->operands
[0]->type
->base_type
== GLSL_TYPE_FLOAT
);
1467 emit(ir
, TGSI_OPCODE_SNE
, st_dst_reg(temp
), op
[0], op
[1]);
1468 emit_dp(ir
, result_dst
, temp
, temp
, vector_elements
);
1469 emit(ir
, TGSI_OPCODE_SNE
, result_dst
, result_src
, st_src_reg_for_float(0.0));
1471 emit(ir
, TGSI_OPCODE_SNE
, result_dst
, op
[0], op
[1]);
1476 assert(ir
->operands
[0]->type
->is_vector());
1477 emit_dp(ir
, result_dst
, op
[0], op
[0],
1478 ir
->operands
[0]->type
->vector_elements
);
1479 emit(ir
, TGSI_OPCODE_SNE
, result_dst
, result_src
, st_src_reg_for_float(0.0));
1482 case ir_binop_logic_xor
:
1483 emit(ir
, TGSI_OPCODE_SNE
, result_dst
, op
[0], op
[1]);
1486 case ir_binop_logic_or
:
1487 /* This could be a saturated add and skip the SNE. */
1488 emit(ir
, TGSI_OPCODE_ADD
, result_dst
, op
[0], op
[1]);
1489 emit(ir
, TGSI_OPCODE_SNE
, result_dst
, result_src
, st_src_reg_for_float(0.0));
1492 case ir_binop_logic_and
:
1493 /* the bool args are stored as float 0.0 or 1.0, so "mul" gives us "and". */
1494 emit(ir
, TGSI_OPCODE_MUL
, result_dst
, op
[0], op
[1]);
1498 assert(ir
->operands
[0]->type
->is_vector());
1499 assert(ir
->operands
[0]->type
== ir
->operands
[1]->type
);
1500 emit_dp(ir
, result_dst
, op
[0], op
[1],
1501 ir
->operands
[0]->type
->vector_elements
);
1505 /* sqrt(x) = x * rsq(x). */
1506 emit_scalar(ir
, TGSI_OPCODE_RSQ
, result_dst
, op
[0]);
1507 emit(ir
, TGSI_OPCODE_MUL
, result_dst
, result_src
, op
[0]);
1508 /* For incoming channels <= 0, set the result to 0. */
1509 op
[0].negate
= ~op
[0].negate
;
1510 emit(ir
, TGSI_OPCODE_CMP
, result_dst
,
1511 op
[0], result_src
, st_src_reg_for_float(0.0));
1514 emit_scalar(ir
, TGSI_OPCODE_RSQ
, result_dst
, op
[0]);
1518 if (native_integers
) {
1519 emit(ir
, TGSI_OPCODE_I2F
, result_dst
, op
[0]);
1524 /* Converting between signed and unsigned integers is a no-op. */
1526 /* Booleans are stored as integers (or floats in GLSL 1.20 and lower). */
1530 if (native_integers
)
1531 emit(ir
, TGSI_OPCODE_F2I
, result_dst
, op
[0]);
1533 emit(ir
, TGSI_OPCODE_TRUNC
, result_dst
, op
[0]);
1537 emit(ir
, TGSI_OPCODE_SNE
, result_dst
, op
[0],
1538 st_src_reg_for_type(result_dst
.type
, 0));
1541 emit(ir
, TGSI_OPCODE_TRUNC
, result_dst
, op
[0]);
1544 op
[0].negate
= ~op
[0].negate
;
1545 emit(ir
, TGSI_OPCODE_FLR
, result_dst
, op
[0]);
1546 result_src
.negate
= ~result_src
.negate
;
1549 emit(ir
, TGSI_OPCODE_FLR
, result_dst
, op
[0]);
1552 emit(ir
, TGSI_OPCODE_FRC
, result_dst
, op
[0]);
1556 emit(ir
, TGSI_OPCODE_MIN
, result_dst
, op
[0], op
[1]);
1559 emit(ir
, TGSI_OPCODE_MAX
, result_dst
, op
[0], op
[1]);
1562 emit_scalar(ir
, TGSI_OPCODE_POW
, result_dst
, op
[0], op
[1]);
1565 case ir_unop_bit_not
:
1566 if (glsl_version
>= 130) {
1567 emit(ir
, TGSI_OPCODE_NOT
, result_dst
, op
[0]);
1571 if (native_integers
) {
1572 emit(ir
, TGSI_OPCODE_U2F
, result_dst
, op
[0]);
1575 case ir_binop_lshift
:
1576 if (glsl_version
>= 130) {
1577 emit(ir
, TGSI_OPCODE_SHL
, result_dst
, op
[0]);
1580 case ir_binop_rshift
:
1581 if (glsl_version
>= 130) {
1582 emit(ir
, TGSI_OPCODE_ISHR
, result_dst
, op
[0]);
1585 case ir_binop_bit_and
:
1586 if (glsl_version
>= 130) {
1587 emit(ir
, TGSI_OPCODE_AND
, result_dst
, op
[0]);
1590 case ir_binop_bit_xor
:
1591 if (glsl_version
>= 130) {
1592 emit(ir
, TGSI_OPCODE_XOR
, result_dst
, op
[0]);
1595 case ir_binop_bit_or
:
1596 if (glsl_version
>= 130) {
1597 emit(ir
, TGSI_OPCODE_OR
, result_dst
, op
[0]);
1600 case ir_unop_round_even
:
1601 assert(!"GLSL 1.30 features unsupported");
1604 case ir_quadop_vector
:
1605 /* This operation should have already been handled.
1607 assert(!"Should not get here.");
1611 this->result
= result_src
;
1616 glsl_to_tgsi_visitor::visit(ir_swizzle
*ir
)
1622 /* Note that this is only swizzles in expressions, not those on the left
1623 * hand side of an assignment, which do write masking. See ir_assignment
1627 ir
->val
->accept(this);
1629 assert(src
.file
!= PROGRAM_UNDEFINED
);
1631 for (i
= 0; i
< 4; i
++) {
1632 if (i
< ir
->type
->vector_elements
) {
1635 swizzle
[i
] = GET_SWZ(src
.swizzle
, ir
->mask
.x
);
1638 swizzle
[i
] = GET_SWZ(src
.swizzle
, ir
->mask
.y
);
1641 swizzle
[i
] = GET_SWZ(src
.swizzle
, ir
->mask
.z
);
1644 swizzle
[i
] = GET_SWZ(src
.swizzle
, ir
->mask
.w
);
1648 /* If the type is smaller than a vec4, replicate the last
1651 swizzle
[i
] = swizzle
[ir
->type
->vector_elements
- 1];
1655 src
.swizzle
= MAKE_SWIZZLE4(swizzle
[0], swizzle
[1], swizzle
[2], swizzle
[3]);
1661 glsl_to_tgsi_visitor::visit(ir_dereference_variable
*ir
)
1663 variable_storage
*entry
= find_variable_storage(ir
->var
);
1664 ir_variable
*var
= ir
->var
;
1667 switch (var
->mode
) {
1668 case ir_var_uniform
:
1669 entry
= new(mem_ctx
) variable_storage(var
, PROGRAM_UNIFORM
,
1671 this->variables
.push_tail(entry
);
1675 /* The linker assigns locations for varyings and attributes,
1676 * including deprecated builtins (like gl_Color), user-assign
1677 * generic attributes (glBindVertexLocation), and
1678 * user-defined varyings.
1680 * FINISHME: We would hit this path for function arguments. Fix!
1682 assert(var
->location
!= -1);
1683 entry
= new(mem_ctx
) variable_storage(var
,
1686 if (this->prog
->Target
== GL_VERTEX_PROGRAM_ARB
&&
1687 var
->location
>= VERT_ATTRIB_GENERIC0
) {
1688 _mesa_add_attribute(this->prog
->Attributes
,
1690 _mesa_sizeof_glsl_type(var
->type
->gl_type
),
1692 var
->location
- VERT_ATTRIB_GENERIC0
);
1696 assert(var
->location
!= -1);
1697 entry
= new(mem_ctx
) variable_storage(var
,
1701 case ir_var_system_value
:
1702 entry
= new(mem_ctx
) variable_storage(var
,
1703 PROGRAM_SYSTEM_VALUE
,
1707 case ir_var_temporary
:
1708 entry
= new(mem_ctx
) variable_storage(var
, PROGRAM_TEMPORARY
,
1710 this->variables
.push_tail(entry
);
1712 next_temp
+= type_size(var
->type
);
1717 printf("Failed to make storage for %s\n", var
->name
);
1722 this->result
= st_src_reg(entry
->file
, entry
->index
, var
->type
);
1723 if (!native_integers
)
1724 this->result
.type
= GLSL_TYPE_FLOAT
;
1728 glsl_to_tgsi_visitor::visit(ir_dereference_array
*ir
)
1732 int element_size
= type_size(ir
->type
);
1734 index
= ir
->array_index
->constant_expression_value();
1736 ir
->array
->accept(this);
1740 src
.index
+= index
->value
.i
[0] * element_size
;
1742 /* Variable index array dereference. It eats the "vec4" of the
1743 * base of the array and an index that offsets the TGSI register
1746 ir
->array_index
->accept(this);
1748 st_src_reg index_reg
;
1750 if (element_size
== 1) {
1751 index_reg
= this->result
;
1753 index_reg
= get_temp(glsl_type::float_type
);
1755 emit(ir
, TGSI_OPCODE_MUL
, st_dst_reg(index_reg
),
1756 this->result
, st_src_reg_for_float(element_size
));
1759 /* If there was already a relative address register involved, add the
1760 * new and the old together to get the new offset.
1762 if (src
.reladdr
!= NULL
) {
1763 st_src_reg accum_reg
= get_temp(glsl_type::float_type
);
1765 emit(ir
, TGSI_OPCODE_ADD
, st_dst_reg(accum_reg
),
1766 index_reg
, *src
.reladdr
);
1768 index_reg
= accum_reg
;
1771 src
.reladdr
= ralloc(mem_ctx
, st_src_reg
);
1772 memcpy(src
.reladdr
, &index_reg
, sizeof(index_reg
));
1775 /* If the type is smaller than a vec4, replicate the last channel out. */
1776 if (ir
->type
->is_scalar() || ir
->type
->is_vector())
1777 src
.swizzle
= swizzle_for_size(ir
->type
->vector_elements
);
1779 src
.swizzle
= SWIZZLE_NOOP
;
1785 glsl_to_tgsi_visitor::visit(ir_dereference_record
*ir
)
1788 const glsl_type
*struct_type
= ir
->record
->type
;
1791 ir
->record
->accept(this);
1793 for (i
= 0; i
< struct_type
->length
; i
++) {
1794 if (strcmp(struct_type
->fields
.structure
[i
].name
, ir
->field
) == 0)
1796 offset
+= type_size(struct_type
->fields
.structure
[i
].type
);
1799 /* If the type is smaller than a vec4, replicate the last channel out. */
1800 if (ir
->type
->is_scalar() || ir
->type
->is_vector())
1801 this->result
.swizzle
= swizzle_for_size(ir
->type
->vector_elements
);
1803 this->result
.swizzle
= SWIZZLE_NOOP
;
1805 this->result
.index
+= offset
;
1809 * We want to be careful in assignment setup to hit the actual storage
1810 * instead of potentially using a temporary like we might with the
1811 * ir_dereference handler.
1814 get_assignment_lhs(ir_dereference
*ir
, glsl_to_tgsi_visitor
*v
)
1816 /* The LHS must be a dereference. If the LHS is a variable indexed array
1817 * access of a vector, it must be separated into a series conditional moves
1818 * before reaching this point (see ir_vec_index_to_cond_assign).
1820 assert(ir
->as_dereference());
1821 ir_dereference_array
*deref_array
= ir
->as_dereference_array();
1823 assert(!deref_array
->array
->type
->is_vector());
1826 /* Use the rvalue deref handler for the most part. We'll ignore
1827 * swizzles in it and write swizzles using writemask, though.
1830 return st_dst_reg(v
->result
);
1834 * Process the condition of a conditional assignment
1836 * Examines the condition of a conditional assignment to generate the optimal
1837 * first operand of a \c CMP instruction. If the condition is a relational
1838 * operator with 0 (e.g., \c ir_binop_less), the value being compared will be
1839 * used as the source for the \c CMP instruction. Otherwise the comparison
1840 * is processed to a boolean result, and the boolean result is used as the
1841 * operand to the CMP instruction.
1844 glsl_to_tgsi_visitor::process_move_condition(ir_rvalue
*ir
)
1846 ir_rvalue
*src_ir
= ir
;
1848 bool switch_order
= false;
1850 ir_expression
*const expr
= ir
->as_expression();
1851 if ((expr
!= NULL
) && (expr
->get_num_operands() == 2)) {
1852 bool zero_on_left
= false;
1854 if (expr
->operands
[0]->is_zero()) {
1855 src_ir
= expr
->operands
[1];
1856 zero_on_left
= true;
1857 } else if (expr
->operands
[1]->is_zero()) {
1858 src_ir
= expr
->operands
[0];
1859 zero_on_left
= false;
1863 * (a < 0) T F F ( a < 0) T F F
1864 * (0 < a) F F T (-a < 0) F F T
1865 * (a <= 0) T T F (-a < 0) F F T (swap order of other operands)
1866 * (0 <= a) F T T ( a < 0) T F F (swap order of other operands)
1867 * (a > 0) F F T (-a < 0) F F T
1868 * (0 > a) T F F ( a < 0) T F F
1869 * (a >= 0) F T T ( a < 0) T F F (swap order of other operands)
1870 * (0 >= a) T T F (-a < 0) F F T (swap order of other operands)
1872 * Note that exchanging the order of 0 and 'a' in the comparison simply
1873 * means that the value of 'a' should be negated.
1876 switch (expr
->operation
) {
1878 switch_order
= false;
1879 negate
= zero_on_left
;
1882 case ir_binop_greater
:
1883 switch_order
= false;
1884 negate
= !zero_on_left
;
1887 case ir_binop_lequal
:
1888 switch_order
= true;
1889 negate
= !zero_on_left
;
1892 case ir_binop_gequal
:
1893 switch_order
= true;
1894 negate
= zero_on_left
;
1898 /* This isn't the right kind of comparison afterall, so make sure
1899 * the whole condition is visited.
1907 src_ir
->accept(this);
1909 /* We use the TGSI_OPCODE_CMP (a < 0 ? b : c) for conditional moves, and the
1910 * condition we produced is 0.0 or 1.0. By flipping the sign, we can
1911 * choose which value TGSI_OPCODE_CMP produces without an extra instruction
1912 * computing the condition.
1915 this->result
.negate
= ~this->result
.negate
;
1917 return switch_order
;
1921 glsl_to_tgsi_visitor::visit(ir_assignment
*ir
)
1927 ir
->rhs
->accept(this);
1930 l
= get_assignment_lhs(ir
->lhs
, this);
1932 /* FINISHME: This should really set to the correct maximal writemask for each
1933 * FINISHME: component written (in the loops below). This case can only
1934 * FINISHME: occur for matrices, arrays, and structures.
1936 if (ir
->write_mask
== 0) {
1937 assert(!ir
->lhs
->type
->is_scalar() && !ir
->lhs
->type
->is_vector());
1938 l
.writemask
= WRITEMASK_XYZW
;
1939 } else if (ir
->lhs
->type
->is_scalar() &&
1940 ir
->lhs
->variable_referenced()->mode
== ir_var_out
) {
1941 /* FINISHME: This hack makes writing to gl_FragDepth, which lives in the
1942 * FINISHME: W component of fragment shader output zero, work correctly.
1944 l
.writemask
= WRITEMASK_XYZW
;
1947 int first_enabled_chan
= 0;
1950 l
.writemask
= ir
->write_mask
;
1952 for (int i
= 0; i
< 4; i
++) {
1953 if (l
.writemask
& (1 << i
)) {
1954 first_enabled_chan
= GET_SWZ(r
.swizzle
, i
);
1959 /* Swizzle a small RHS vector into the channels being written.
1961 * glsl ir treats write_mask as dictating how many channels are
1962 * present on the RHS while TGSI treats write_mask as just
1963 * showing which channels of the vec4 RHS get written.
1965 for (int i
= 0; i
< 4; i
++) {
1966 if (l
.writemask
& (1 << i
))
1967 swizzles
[i
] = GET_SWZ(r
.swizzle
, rhs_chan
++);
1969 swizzles
[i
] = first_enabled_chan
;
1971 r
.swizzle
= MAKE_SWIZZLE4(swizzles
[0], swizzles
[1],
1972 swizzles
[2], swizzles
[3]);
1975 assert(l
.file
!= PROGRAM_UNDEFINED
);
1976 assert(r
.file
!= PROGRAM_UNDEFINED
);
1978 if (ir
->condition
) {
1979 const bool switch_order
= this->process_move_condition(ir
->condition
);
1980 st_src_reg condition
= this->result
;
1982 for (i
= 0; i
< type_size(ir
->lhs
->type
); i
++) {
1983 st_src_reg l_src
= st_src_reg(l
);
1984 l_src
.swizzle
= swizzle_for_size(ir
->lhs
->type
->vector_elements
);
1987 emit(ir
, TGSI_OPCODE_CMP
, l
, condition
, l_src
, r
);
1989 emit(ir
, TGSI_OPCODE_CMP
, l
, condition
, r
, l_src
);
1995 } else if (ir
->rhs
->as_expression() &&
1996 this->instructions
.get_tail() &&
1997 ir
->rhs
== ((glsl_to_tgsi_instruction
*)this->instructions
.get_tail())->ir
&&
1998 type_size(ir
->lhs
->type
) == 1 &&
1999 l
.writemask
== ((glsl_to_tgsi_instruction
*)this->instructions
.get_tail())->dst
.writemask
) {
2000 /* To avoid emitting an extra MOV when assigning an expression to a
2001 * variable, emit the last instruction of the expression again, but
2002 * replace the destination register with the target of the assignment.
2003 * Dead code elimination will remove the original instruction.
2005 glsl_to_tgsi_instruction
*inst
, *new_inst
;
2006 inst
= (glsl_to_tgsi_instruction
*)this->instructions
.get_tail();
2007 new_inst
= emit(ir
, inst
->op
, l
, inst
->src
[0], inst
->src
[1], inst
->src
[2]);
2008 new_inst
->saturate
= inst
->saturate
;
2010 for (i
= 0; i
< type_size(ir
->lhs
->type
); i
++) {
2011 emit(ir
, TGSI_OPCODE_MOV
, l
, r
);
2020 glsl_to_tgsi_visitor::visit(ir_constant
*ir
)
2023 GLfloat stack_vals
[4] = { 0 };
2024 gl_constant_value
*values
= (gl_constant_value
*) stack_vals
;
2025 GLenum gl_type
= GL_NONE
;
2027 static int in_array
= 0;
2028 gl_register_file file
= in_array
? PROGRAM_CONSTANT
: PROGRAM_IMMEDIATE
;
2030 /* Unfortunately, 4 floats is all we can get into
2031 * _mesa_add_typed_unnamed_constant. So, make a temp to store an
2032 * aggregate constant and move each constant value into it. If we
2033 * get lucky, copy propagation will eliminate the extra moves.
2035 if (ir
->type
->base_type
== GLSL_TYPE_STRUCT
) {
2036 st_src_reg temp_base
= get_temp(ir
->type
);
2037 st_dst_reg temp
= st_dst_reg(temp_base
);
2039 foreach_iter(exec_list_iterator
, iter
, ir
->components
) {
2040 ir_constant
*field_value
= (ir_constant
*)iter
.get();
2041 int size
= type_size(field_value
->type
);
2045 field_value
->accept(this);
2048 for (i
= 0; i
< (unsigned int)size
; i
++) {
2049 emit(ir
, TGSI_OPCODE_MOV
, temp
, src
);
2055 this->result
= temp_base
;
2059 if (ir
->type
->is_array()) {
2060 st_src_reg temp_base
= get_temp(ir
->type
);
2061 st_dst_reg temp
= st_dst_reg(temp_base
);
2062 int size
= type_size(ir
->type
->fields
.array
);
2067 for (i
= 0; i
< ir
->type
->length
; i
++) {
2068 ir
->array_elements
[i
]->accept(this);
2070 for (int j
= 0; j
< size
; j
++) {
2071 emit(ir
, TGSI_OPCODE_MOV
, temp
, src
);
2077 this->result
= temp_base
;
2082 if (ir
->type
->is_matrix()) {
2083 st_src_reg mat
= get_temp(ir
->type
);
2084 st_dst_reg mat_column
= st_dst_reg(mat
);
2086 for (i
= 0; i
< ir
->type
->matrix_columns
; i
++) {
2087 assert(ir
->type
->base_type
== GLSL_TYPE_FLOAT
);
2088 values
= (gl_constant_value
*) &ir
->value
.f
[i
* ir
->type
->vector_elements
];
2090 src
= st_src_reg(file
, -1, ir
->type
->base_type
);
2091 src
.index
= add_constant(file
,
2093 ir
->type
->vector_elements
,
2096 emit(ir
, TGSI_OPCODE_MOV
, mat_column
, src
);
2105 switch (ir
->type
->base_type
) {
2106 case GLSL_TYPE_FLOAT
:
2108 for (i
= 0; i
< ir
->type
->vector_elements
; i
++) {
2109 values
[i
].f
= ir
->value
.f
[i
];
2112 case GLSL_TYPE_UINT
:
2113 gl_type
= native_integers
? GL_UNSIGNED_INT
: GL_FLOAT
;
2114 for (i
= 0; i
< ir
->type
->vector_elements
; i
++) {
2115 if (native_integers
)
2116 values
[i
].u
= ir
->value
.u
[i
];
2118 values
[i
].f
= ir
->value
.u
[i
];
2122 gl_type
= native_integers
? GL_INT
: GL_FLOAT
;
2123 for (i
= 0; i
< ir
->type
->vector_elements
; i
++) {
2124 if (native_integers
)
2125 values
[i
].i
= ir
->value
.i
[i
];
2127 values
[i
].f
= ir
->value
.i
[i
];
2130 case GLSL_TYPE_BOOL
:
2131 gl_type
= native_integers
? GL_BOOL
: GL_FLOAT
;
2132 for (i
= 0; i
< ir
->type
->vector_elements
; i
++) {
2133 if (native_integers
)
2134 values
[i
].b
= ir
->value
.b
[i
];
2136 values
[i
].f
= ir
->value
.b
[i
];
2140 assert(!"Non-float/uint/int/bool constant");
2143 this->result
= st_src_reg(file
, -1, ir
->type
);
2144 this->result
.index
= add_constant(file
,
2146 ir
->type
->vector_elements
,
2148 &this->result
.swizzle
);
2152 glsl_to_tgsi_visitor::get_function_signature(ir_function_signature
*sig
)
2154 function_entry
*entry
;
2156 foreach_iter(exec_list_iterator
, iter
, this->function_signatures
) {
2157 entry
= (function_entry
*)iter
.get();
2159 if (entry
->sig
== sig
)
2163 entry
= ralloc(mem_ctx
, function_entry
);
2165 entry
->sig_id
= this->next_signature_id
++;
2166 entry
->bgn_inst
= NULL
;
2168 /* Allocate storage for all the parameters. */
2169 foreach_iter(exec_list_iterator
, iter
, sig
->parameters
) {
2170 ir_variable
*param
= (ir_variable
*)iter
.get();
2171 variable_storage
*storage
;
2173 storage
= find_variable_storage(param
);
2176 storage
= new(mem_ctx
) variable_storage(param
, PROGRAM_TEMPORARY
,
2178 this->variables
.push_tail(storage
);
2180 this->next_temp
+= type_size(param
->type
);
2183 if (!sig
->return_type
->is_void()) {
2184 entry
->return_reg
= get_temp(sig
->return_type
);
2186 entry
->return_reg
= undef_src
;
2189 this->function_signatures
.push_tail(entry
);
2194 glsl_to_tgsi_visitor::visit(ir_call
*ir
)
2196 glsl_to_tgsi_instruction
*call_inst
;
2197 ir_function_signature
*sig
= ir
->get_callee();
2198 function_entry
*entry
= get_function_signature(sig
);
2201 /* Process in parameters. */
2202 exec_list_iterator sig_iter
= sig
->parameters
.iterator();
2203 foreach_iter(exec_list_iterator
, iter
, *ir
) {
2204 ir_rvalue
*param_rval
= (ir_rvalue
*)iter
.get();
2205 ir_variable
*param
= (ir_variable
*)sig_iter
.get();
2207 if (param
->mode
== ir_var_in
||
2208 param
->mode
== ir_var_inout
) {
2209 variable_storage
*storage
= find_variable_storage(param
);
2212 param_rval
->accept(this);
2213 st_src_reg r
= this->result
;
2216 l
.file
= storage
->file
;
2217 l
.index
= storage
->index
;
2219 l
.writemask
= WRITEMASK_XYZW
;
2220 l
.cond_mask
= COND_TR
;
2222 for (i
= 0; i
< type_size(param
->type
); i
++) {
2223 emit(ir
, TGSI_OPCODE_MOV
, l
, r
);
2231 assert(!sig_iter
.has_next());
2233 /* Emit call instruction */
2234 call_inst
= emit(ir
, TGSI_OPCODE_CAL
);
2235 call_inst
->function
= entry
;
2237 /* Process out parameters. */
2238 sig_iter
= sig
->parameters
.iterator();
2239 foreach_iter(exec_list_iterator
, iter
, *ir
) {
2240 ir_rvalue
*param_rval
= (ir_rvalue
*)iter
.get();
2241 ir_variable
*param
= (ir_variable
*)sig_iter
.get();
2243 if (param
->mode
== ir_var_out
||
2244 param
->mode
== ir_var_inout
) {
2245 variable_storage
*storage
= find_variable_storage(param
);
2249 r
.file
= storage
->file
;
2250 r
.index
= storage
->index
;
2252 r
.swizzle
= SWIZZLE_NOOP
;
2255 param_rval
->accept(this);
2256 st_dst_reg l
= st_dst_reg(this->result
);
2258 for (i
= 0; i
< type_size(param
->type
); i
++) {
2259 emit(ir
, TGSI_OPCODE_MOV
, l
, r
);
2267 assert(!sig_iter
.has_next());
2269 /* Process return value. */
2270 this->result
= entry
->return_reg
;
2274 glsl_to_tgsi_visitor::visit(ir_texture
*ir
)
2276 st_src_reg result_src
, coord
, lod_info
, projector
, dx
, dy
;
2277 st_dst_reg result_dst
, coord_dst
;
2278 glsl_to_tgsi_instruction
*inst
= NULL
;
2279 unsigned opcode
= TGSI_OPCODE_NOP
;
2281 ir
->coordinate
->accept(this);
2283 /* Put our coords in a temp. We'll need to modify them for shadow,
2284 * projection, or LOD, so the only case we'd use it as is is if
2285 * we're doing plain old texturing. The optimization passes on
2286 * glsl_to_tgsi_visitor should handle cleaning up our mess in that case.
2288 coord
= get_temp(glsl_type::vec4_type
);
2289 coord_dst
= st_dst_reg(coord
);
2290 emit(ir
, TGSI_OPCODE_MOV
, coord_dst
, this->result
);
2292 if (ir
->projector
) {
2293 ir
->projector
->accept(this);
2294 projector
= this->result
;
2297 /* Storage for our result. Ideally for an assignment we'd be using
2298 * the actual storage for the result here, instead.
2300 result_src
= get_temp(glsl_type::vec4_type
);
2301 result_dst
= st_dst_reg(result_src
);
2305 opcode
= TGSI_OPCODE_TEX
;
2308 opcode
= TGSI_OPCODE_TXB
;
2309 ir
->lod_info
.bias
->accept(this);
2310 lod_info
= this->result
;
2313 opcode
= TGSI_OPCODE_TXL
;
2314 ir
->lod_info
.lod
->accept(this);
2315 lod_info
= this->result
;
2318 opcode
= TGSI_OPCODE_TXD
;
2319 ir
->lod_info
.grad
.dPdx
->accept(this);
2321 ir
->lod_info
.grad
.dPdy
->accept(this);
2324 case ir_txf
: /* TODO: use TGSI_OPCODE_TXF here */
2325 assert(!"GLSL 1.30 features unsupported");
2329 if (ir
->projector
) {
2330 if (opcode
== TGSI_OPCODE_TEX
) {
2331 /* Slot the projector in as the last component of the coord. */
2332 coord_dst
.writemask
= WRITEMASK_W
;
2333 emit(ir
, TGSI_OPCODE_MOV
, coord_dst
, projector
);
2334 coord_dst
.writemask
= WRITEMASK_XYZW
;
2335 opcode
= TGSI_OPCODE_TXP
;
2337 st_src_reg coord_w
= coord
;
2338 coord_w
.swizzle
= SWIZZLE_WWWW
;
2340 /* For the other TEX opcodes there's no projective version
2341 * since the last slot is taken up by LOD info. Do the
2342 * projective divide now.
2344 coord_dst
.writemask
= WRITEMASK_W
;
2345 emit(ir
, TGSI_OPCODE_RCP
, coord_dst
, projector
);
2347 /* In the case where we have to project the coordinates "by hand,"
2348 * the shadow comparator value must also be projected.
2350 st_src_reg tmp_src
= coord
;
2351 if (ir
->shadow_comparitor
) {
2352 /* Slot the shadow value in as the second to last component of the
2355 ir
->shadow_comparitor
->accept(this);
2357 tmp_src
= get_temp(glsl_type::vec4_type
);
2358 st_dst_reg tmp_dst
= st_dst_reg(tmp_src
);
2360 tmp_dst
.writemask
= WRITEMASK_Z
;
2361 emit(ir
, TGSI_OPCODE_MOV
, tmp_dst
, this->result
);
2363 tmp_dst
.writemask
= WRITEMASK_XY
;
2364 emit(ir
, TGSI_OPCODE_MOV
, tmp_dst
, coord
);
2367 coord_dst
.writemask
= WRITEMASK_XYZ
;
2368 emit(ir
, TGSI_OPCODE_MUL
, coord_dst
, tmp_src
, coord_w
);
2370 coord_dst
.writemask
= WRITEMASK_XYZW
;
2371 coord
.swizzle
= SWIZZLE_XYZW
;
2375 /* If projection is done and the opcode is not TGSI_OPCODE_TXP, then the shadow
2376 * comparator was put in the correct place (and projected) by the code,
2377 * above, that handles by-hand projection.
2379 if (ir
->shadow_comparitor
&& (!ir
->projector
|| opcode
== TGSI_OPCODE_TXP
)) {
2380 /* Slot the shadow value in as the second to last component of the
2383 ir
->shadow_comparitor
->accept(this);
2384 coord_dst
.writemask
= WRITEMASK_Z
;
2385 emit(ir
, TGSI_OPCODE_MOV
, coord_dst
, this->result
);
2386 coord_dst
.writemask
= WRITEMASK_XYZW
;
2389 if (opcode
== TGSI_OPCODE_TXL
|| opcode
== TGSI_OPCODE_TXB
) {
2390 /* TGSI stores LOD or LOD bias in the last channel of the coords. */
2391 coord_dst
.writemask
= WRITEMASK_W
;
2392 emit(ir
, TGSI_OPCODE_MOV
, coord_dst
, lod_info
);
2393 coord_dst
.writemask
= WRITEMASK_XYZW
;
2396 if (opcode
== TGSI_OPCODE_TXD
)
2397 inst
= emit(ir
, opcode
, result_dst
, coord
, dx
, dy
);
2399 inst
= emit(ir
, opcode
, result_dst
, coord
);
2401 if (ir
->shadow_comparitor
)
2402 inst
->tex_shadow
= GL_TRUE
;
2404 inst
->sampler
= _mesa_get_sampler_uniform_value(ir
->sampler
,
2405 this->shader_program
,
2408 const glsl_type
*sampler_type
= ir
->sampler
->type
;
2410 switch (sampler_type
->sampler_dimensionality
) {
2411 case GLSL_SAMPLER_DIM_1D
:
2412 inst
->tex_target
= (sampler_type
->sampler_array
)
2413 ? TEXTURE_1D_ARRAY_INDEX
: TEXTURE_1D_INDEX
;
2415 case GLSL_SAMPLER_DIM_2D
:
2416 inst
->tex_target
= (sampler_type
->sampler_array
)
2417 ? TEXTURE_2D_ARRAY_INDEX
: TEXTURE_2D_INDEX
;
2419 case GLSL_SAMPLER_DIM_3D
:
2420 inst
->tex_target
= TEXTURE_3D_INDEX
;
2422 case GLSL_SAMPLER_DIM_CUBE
:
2423 inst
->tex_target
= TEXTURE_CUBE_INDEX
;
2425 case GLSL_SAMPLER_DIM_RECT
:
2426 inst
->tex_target
= TEXTURE_RECT_INDEX
;
2428 case GLSL_SAMPLER_DIM_BUF
:
2429 assert(!"FINISHME: Implement ARB_texture_buffer_object");
2432 assert(!"Should not get here.");
2435 this->result
= result_src
;
2439 glsl_to_tgsi_visitor::visit(ir_return
*ir
)
2441 if (ir
->get_value()) {
2445 assert(current_function
);
2447 ir
->get_value()->accept(this);
2448 st_src_reg r
= this->result
;
2450 l
= st_dst_reg(current_function
->return_reg
);
2452 for (i
= 0; i
< type_size(current_function
->sig
->return_type
); i
++) {
2453 emit(ir
, TGSI_OPCODE_MOV
, l
, r
);
2459 emit(ir
, TGSI_OPCODE_RET
);
2463 glsl_to_tgsi_visitor::visit(ir_discard
*ir
)
2465 struct gl_fragment_program
*fp
= (struct gl_fragment_program
*)this->prog
;
2467 if (ir
->condition
) {
2468 ir
->condition
->accept(this);
2469 this->result
.negate
= ~this->result
.negate
;
2470 emit(ir
, TGSI_OPCODE_KIL
, undef_dst
, this->result
);
2472 emit(ir
, TGSI_OPCODE_KILP
);
2475 fp
->UsesKill
= GL_TRUE
;
2479 glsl_to_tgsi_visitor::visit(ir_if
*ir
)
2481 glsl_to_tgsi_instruction
*cond_inst
, *if_inst
;
2482 glsl_to_tgsi_instruction
*prev_inst
;
2484 prev_inst
= (glsl_to_tgsi_instruction
*)this->instructions
.get_tail();
2486 ir
->condition
->accept(this);
2487 assert(this->result
.file
!= PROGRAM_UNDEFINED
);
2489 if (this->options
->EmitCondCodes
) {
2490 cond_inst
= (glsl_to_tgsi_instruction
*)this->instructions
.get_tail();
2492 /* See if we actually generated any instruction for generating
2493 * the condition. If not, then cook up a move to a temp so we
2494 * have something to set cond_update on.
2496 if (cond_inst
== prev_inst
) {
2497 st_src_reg temp
= get_temp(glsl_type::bool_type
);
2498 cond_inst
= emit(ir
->condition
, TGSI_OPCODE_MOV
, st_dst_reg(temp
), result
);
2500 cond_inst
->cond_update
= GL_TRUE
;
2502 if_inst
= emit(ir
->condition
, TGSI_OPCODE_IF
);
2503 if_inst
->dst
.cond_mask
= COND_NE
;
2505 if_inst
= emit(ir
->condition
, TGSI_OPCODE_IF
, undef_dst
, this->result
);
2508 this->instructions
.push_tail(if_inst
);
2510 visit_exec_list(&ir
->then_instructions
, this);
2512 if (!ir
->else_instructions
.is_empty()) {
2513 emit(ir
->condition
, TGSI_OPCODE_ELSE
);
2514 visit_exec_list(&ir
->else_instructions
, this);
2517 if_inst
= emit(ir
->condition
, TGSI_OPCODE_ENDIF
);
2520 glsl_to_tgsi_visitor::glsl_to_tgsi_visitor()
2522 result
.file
= PROGRAM_UNDEFINED
;
2524 next_signature_id
= 1;
2526 current_function
= NULL
;
2527 num_address_regs
= 0;
2528 indirect_addr_temps
= false;
2529 indirect_addr_consts
= false;
2530 mem_ctx
= ralloc_context(NULL
);
2533 glsl_to_tgsi_visitor::~glsl_to_tgsi_visitor()
2535 ralloc_free(mem_ctx
);
2538 extern "C" void free_glsl_to_tgsi_visitor(glsl_to_tgsi_visitor
*v
)
2545 * Count resources used by the given gpu program (number of texture
2549 count_resources(glsl_to_tgsi_visitor
*v
, gl_program
*prog
)
2551 v
->samplers_used
= 0;
2553 foreach_iter(exec_list_iterator
, iter
, v
->instructions
) {
2554 glsl_to_tgsi_instruction
*inst
= (glsl_to_tgsi_instruction
*)iter
.get();
2556 if (is_tex_instruction(inst
->op
)) {
2557 v
->samplers_used
|= 1 << inst
->sampler
;
2559 prog
->SamplerTargets
[inst
->sampler
] =
2560 (gl_texture_index
)inst
->tex_target
;
2561 if (inst
->tex_shadow
) {
2562 prog
->ShadowSamplers
|= 1 << inst
->sampler
;
2567 prog
->SamplersUsed
= v
->samplers_used
;
2568 _mesa_update_shader_textures_used(prog
);
2573 * Check if the given vertex/fragment/shader program is within the
2574 * resource limits of the context (number of texture units, etc).
2575 * If any of those checks fail, record a linker error.
2577 * XXX more checks are needed...
2580 check_resources(const struct gl_context
*ctx
,
2581 struct gl_shader_program
*shader_program
,
2582 glsl_to_tgsi_visitor
*prog
,
2583 struct gl_program
*proginfo
)
2585 switch (proginfo
->Target
) {
2586 case GL_VERTEX_PROGRAM_ARB
:
2587 if (_mesa_bitcount(prog
->samplers_used
) >
2588 ctx
->Const
.MaxVertexTextureImageUnits
) {
2589 fail_link(shader_program
, "Too many vertex shader texture samplers");
2591 if (proginfo
->Parameters
->NumParameters
> MAX_UNIFORMS
) {
2592 fail_link(shader_program
, "Too many vertex shader constants");
2595 case MESA_GEOMETRY_PROGRAM
:
2596 if (_mesa_bitcount(prog
->samplers_used
) >
2597 ctx
->Const
.MaxGeometryTextureImageUnits
) {
2598 fail_link(shader_program
, "Too many geometry shader texture samplers");
2600 if (proginfo
->Parameters
->NumParameters
>
2601 MAX_GEOMETRY_UNIFORM_COMPONENTS
/ 4) {
2602 fail_link(shader_program
, "Too many geometry shader constants");
2605 case GL_FRAGMENT_PROGRAM_ARB
:
2606 if (_mesa_bitcount(prog
->samplers_used
) >
2607 ctx
->Const
.MaxTextureImageUnits
) {
2608 fail_link(shader_program
, "Too many fragment shader texture samplers");
2610 if (proginfo
->Parameters
->NumParameters
> MAX_UNIFORMS
) {
2611 fail_link(shader_program
, "Too many fragment shader constants");
2615 _mesa_problem(ctx
, "unexpected program type in check_resources()");
2621 struct uniform_sort
{
2622 struct gl_uniform
*u
;
2626 /* The shader_program->Uniforms list is almost sorted in increasing
2627 * uniform->{Frag,Vert}Pos locations, but not quite when there are
2628 * uniforms shared between targets. We need to add parameters in
2629 * increasing order for the targets.
2632 sort_uniforms(const void *a
, const void *b
)
2634 struct uniform_sort
*u1
= (struct uniform_sort
*)a
;
2635 struct uniform_sort
*u2
= (struct uniform_sort
*)b
;
2637 return u1
->pos
- u2
->pos
;
2640 /* Add the uniforms to the parameters. The linker chose locations
2641 * in our parameters lists (which weren't created yet), which the
2642 * uniforms code will use to poke values into our parameters list
2643 * when uniforms are updated.
2646 add_uniforms_to_parameters_list(struct gl_shader_program
*shader_program
,
2647 struct gl_shader
*shader
,
2648 struct gl_program
*prog
)
2651 unsigned int next_sampler
= 0, num_uniforms
= 0;
2652 struct uniform_sort
*sorted_uniforms
;
2654 sorted_uniforms
= ralloc_array(NULL
, struct uniform_sort
,
2655 shader_program
->Uniforms
->NumUniforms
);
2657 for (i
= 0; i
< shader_program
->Uniforms
->NumUniforms
; i
++) {
2658 struct gl_uniform
*uniform
= shader_program
->Uniforms
->Uniforms
+ i
;
2659 int parameter_index
= -1;
2661 switch (shader
->Type
) {
2662 case GL_VERTEX_SHADER
:
2663 parameter_index
= uniform
->VertPos
;
2665 case GL_FRAGMENT_SHADER
:
2666 parameter_index
= uniform
->FragPos
;
2668 case GL_GEOMETRY_SHADER
:
2669 parameter_index
= uniform
->GeomPos
;
2673 /* Only add uniforms used in our target. */
2674 if (parameter_index
!= -1) {
2675 sorted_uniforms
[num_uniforms
].pos
= parameter_index
;
2676 sorted_uniforms
[num_uniforms
].u
= uniform
;
2681 qsort(sorted_uniforms
, num_uniforms
, sizeof(struct uniform_sort
),
2684 for (i
= 0; i
< num_uniforms
; i
++) {
2685 struct gl_uniform
*uniform
= sorted_uniforms
[i
].u
;
2686 int parameter_index
= sorted_uniforms
[i
].pos
;
2687 const glsl_type
*type
= uniform
->Type
;
2690 if (type
->is_vector() ||
2691 type
->is_scalar()) {
2692 size
= type
->vector_elements
;
2694 size
= type_size(type
) * 4;
2697 gl_register_file file
;
2698 if (type
->is_sampler() ||
2699 (type
->is_array() && type
->fields
.array
->is_sampler())) {
2700 file
= PROGRAM_SAMPLER
;
2702 file
= PROGRAM_UNIFORM
;
2705 GLint index
= _mesa_lookup_parameter_index(prog
->Parameters
, -1,
2709 index
= _mesa_add_parameter(prog
->Parameters
, file
,
2710 uniform
->Name
, size
, type
->gl_type
,
2713 /* Sampler uniform values are stored in prog->SamplerUnits,
2714 * and the entry in that array is selected by this index we
2715 * store in ParameterValues[].
2717 if (file
== PROGRAM_SAMPLER
) {
2718 for (unsigned int j
= 0; j
< size
/ 4; j
++)
2719 prog
->Parameters
->ParameterValues
[index
+ j
][0].f
= next_sampler
++;
2722 /* The location chosen in the Parameters list here (returned
2723 * from _mesa_add_uniform) has to match what the linker chose.
2725 if (index
!= parameter_index
) {
2726 fail_link(shader_program
, "Allocation of uniform `%s' to target "
2727 "failed (%d vs %d)\n",
2728 uniform
->Name
, index
, parameter_index
);
2733 ralloc_free(sorted_uniforms
);
2737 set_uniform_initializer(struct gl_context
*ctx
, void *mem_ctx
,
2738 struct gl_shader_program
*shader_program
,
2739 const char *name
, const glsl_type
*type
,
2742 if (type
->is_record()) {
2743 ir_constant
*field_constant
;
2745 field_constant
= (ir_constant
*)val
->components
.get_head();
2747 for (unsigned int i
= 0; i
< type
->length
; i
++) {
2748 const glsl_type
*field_type
= type
->fields
.structure
[i
].type
;
2749 const char *field_name
= ralloc_asprintf(mem_ctx
, "%s.%s", name
,
2750 type
->fields
.structure
[i
].name
);
2751 set_uniform_initializer(ctx
, mem_ctx
, shader_program
, field_name
,
2752 field_type
, field_constant
);
2753 field_constant
= (ir_constant
*)field_constant
->next
;
2758 int loc
= _mesa_get_uniform_location(ctx
, shader_program
, name
);
2761 fail_link(shader_program
,
2762 "Couldn't find uniform for initializer %s\n", name
);
2766 for (unsigned int i
= 0; i
< (type
->is_array() ? type
->length
: 1); i
++) {
2767 ir_constant
*element
;
2768 const glsl_type
*element_type
;
2769 if (type
->is_array()) {
2770 element
= val
->array_elements
[i
];
2771 element_type
= type
->fields
.array
;
2774 element_type
= type
;
2779 if (element_type
->base_type
== GLSL_TYPE_BOOL
) {
2780 int *conv
= ralloc_array(mem_ctx
, int, element_type
->components());
2781 for (unsigned int j
= 0; j
< element_type
->components(); j
++) {
2782 conv
[j
] = element
->value
.b
[j
];
2784 values
= (void *)conv
;
2785 element_type
= glsl_type::get_instance(GLSL_TYPE_INT
,
2786 element_type
->vector_elements
,
2789 values
= &element
->value
;
2792 if (element_type
->is_matrix()) {
2793 _mesa_uniform_matrix(ctx
, shader_program
,
2794 element_type
->matrix_columns
,
2795 element_type
->vector_elements
,
2796 loc
, 1, GL_FALSE
, (GLfloat
*)values
);
2797 loc
+= element_type
->matrix_columns
;
2799 _mesa_uniform(ctx
, shader_program
, loc
, element_type
->matrix_columns
,
2800 values
, element_type
->gl_type
);
2801 loc
+= type_size(element_type
);
2807 set_uniform_initializers(struct gl_context
*ctx
,
2808 struct gl_shader_program
*shader_program
)
2810 void *mem_ctx
= NULL
;
2812 for (unsigned int i
= 0; i
< MESA_SHADER_TYPES
; i
++) {
2813 struct gl_shader
*shader
= shader_program
->_LinkedShaders
[i
];
2818 foreach_iter(exec_list_iterator
, iter
, *shader
->ir
) {
2819 ir_instruction
*ir
= (ir_instruction
*)iter
.get();
2820 ir_variable
*var
= ir
->as_variable();
2822 if (!var
|| var
->mode
!= ir_var_uniform
|| !var
->constant_value
)
2826 mem_ctx
= ralloc_context(NULL
);
2828 set_uniform_initializer(ctx
, mem_ctx
, shader_program
, var
->name
,
2829 var
->type
, var
->constant_value
);
2833 ralloc_free(mem_ctx
);
2837 * Scan/rewrite program to remove reads of custom (output) registers.
2838 * The passed type has to be either PROGRAM_OUTPUT or PROGRAM_VARYING
2839 * (for vertex shaders).
2840 * In GLSL shaders, varying vars can be read and written.
2841 * On some hardware, trying to read an output register causes trouble.
2842 * So, rewrite the program to use a temporary register in this case.
2844 * Based on _mesa_remove_output_reads from programopt.c.
2847 glsl_to_tgsi_visitor::remove_output_reads(gl_register_file type
)
2850 GLint outputMap
[VERT_RESULT_MAX
];
2851 GLint outputTypes
[VERT_RESULT_MAX
];
2852 GLuint numVaryingReads
= 0;
2853 GLboolean usedTemps
[MAX_TEMPS
];
2854 GLuint firstTemp
= 0;
2856 _mesa_find_used_registers(prog
, PROGRAM_TEMPORARY
,
2857 usedTemps
, MAX_TEMPS
);
2859 assert(type
== PROGRAM_VARYING
|| type
== PROGRAM_OUTPUT
);
2860 assert(prog
->Target
== GL_VERTEX_PROGRAM_ARB
|| type
!= PROGRAM_VARYING
);
2862 for (i
= 0; i
< VERT_RESULT_MAX
; i
++)
2865 /* look for instructions which read from varying vars */
2866 foreach_iter(exec_list_iterator
, iter
, this->instructions
) {
2867 glsl_to_tgsi_instruction
*inst
= (glsl_to_tgsi_instruction
*)iter
.get();
2868 const GLuint numSrc
= num_inst_src_regs(inst
->op
);
2870 for (j
= 0; j
< numSrc
; j
++) {
2871 if (inst
->src
[j
].file
== type
) {
2872 /* replace the read with a temp reg */
2873 const GLuint var
= inst
->src
[j
].index
;
2874 if (outputMap
[var
] == -1) {
2876 outputMap
[var
] = _mesa_find_free_register(usedTemps
,
2879 outputTypes
[var
] = inst
->src
[j
].type
;
2880 firstTemp
= outputMap
[var
] + 1;
2882 inst
->src
[j
].file
= PROGRAM_TEMPORARY
;
2883 inst
->src
[j
].index
= outputMap
[var
];
2888 if (numVaryingReads
== 0)
2889 return; /* nothing to be done */
2891 /* look for instructions which write to the varying vars identified above */
2892 foreach_iter(exec_list_iterator
, iter
, this->instructions
) {
2893 glsl_to_tgsi_instruction
*inst
= (glsl_to_tgsi_instruction
*)iter
.get();
2894 if (inst
->dst
.file
== type
&& outputMap
[inst
->dst
.index
] >= 0) {
2895 /* change inst to write to the temp reg, instead of the varying */
2896 inst
->dst
.file
= PROGRAM_TEMPORARY
;
2897 inst
->dst
.index
= outputMap
[inst
->dst
.index
];
2901 /* insert new MOV instructions at the end */
2902 for (i
= 0; i
< VERT_RESULT_MAX
; i
++) {
2903 if (outputMap
[i
] >= 0) {
2904 /* MOV VAR[i], TEMP[tmp]; */
2905 st_src_reg src
= st_src_reg(PROGRAM_TEMPORARY
, outputMap
[i
], outputTypes
[i
]);
2906 st_dst_reg dst
= st_dst_reg(type
, WRITEMASK_XYZW
, outputTypes
[i
]);
2908 this->emit(NULL
, TGSI_OPCODE_MOV
, dst
, src
);
2914 * Returns the mask of channels (bitmask of WRITEMASK_X,Y,Z,W) which
2915 * are read from the given src in this instruction
2918 get_src_arg_mask(st_dst_reg dst
, st_src_reg src
)
2920 int read_mask
= 0, comp
;
2922 /* Now, given the src swizzle and the written channels, find which
2923 * components are actually read
2925 for (comp
= 0; comp
< 4; ++comp
) {
2926 const unsigned coord
= GET_SWZ(src
.swizzle
, comp
);
2928 if (dst
.writemask
& (1 << comp
) && coord
<= SWIZZLE_W
)
2929 read_mask
|= 1 << coord
;
2936 * This pass replaces CMP T0, T1 T2 T0 with MOV T0, T2 when the CMP
2937 * instruction is the first instruction to write to register T0. There are
2938 * several lowering passes done in GLSL IR (e.g. branches and
2939 * relative addressing) that create a large number of conditional assignments
2940 * that ir_to_mesa converts to CMP instructions like the one mentioned above.
2942 * Here is why this conversion is safe:
2943 * CMP T0, T1 T2 T0 can be expanded to:
2949 * If (T1 < 0.0) evaluates to true then our replacement MOV T0, T2 is the same
2950 * as the original program. If (T1 < 0.0) evaluates to false, executing
2951 * MOV T0, T0 will store a garbage value in T0 since T0 is uninitialized.
2952 * Therefore, it doesn't matter that we are replacing MOV T0, T0 with MOV T0, T2
2953 * because any instruction that was going to read from T0 after this was going
2954 * to read a garbage value anyway.
2957 glsl_to_tgsi_visitor::simplify_cmp(void)
2959 unsigned tempWrites
[MAX_TEMPS
];
2960 unsigned outputWrites
[MAX_PROGRAM_OUTPUTS
];
2962 memset(tempWrites
, 0, sizeof(tempWrites
));
2963 memset(outputWrites
, 0, sizeof(outputWrites
));
2965 foreach_iter(exec_list_iterator
, iter
, this->instructions
) {
2966 glsl_to_tgsi_instruction
*inst
= (glsl_to_tgsi_instruction
*)iter
.get();
2967 unsigned prevWriteMask
= 0;
2969 /* Give up if we encounter relative addressing or flow control. */
2970 if (inst
->dst
.reladdr
||
2971 tgsi_get_opcode_info(inst
->op
)->is_branch
||
2972 inst
->op
== TGSI_OPCODE_BGNSUB
||
2973 inst
->op
== TGSI_OPCODE_CONT
||
2974 inst
->op
== TGSI_OPCODE_END
||
2975 inst
->op
== TGSI_OPCODE_ENDSUB
||
2976 inst
->op
== TGSI_OPCODE_RET
) {
2980 if (inst
->dst
.file
== PROGRAM_OUTPUT
) {
2981 assert(inst
->dst
.index
< MAX_PROGRAM_OUTPUTS
);
2982 prevWriteMask
= outputWrites
[inst
->dst
.index
];
2983 outputWrites
[inst
->dst
.index
] |= inst
->dst
.writemask
;
2984 } else if (inst
->dst
.file
== PROGRAM_TEMPORARY
) {
2985 assert(inst
->dst
.index
< MAX_TEMPS
);
2986 prevWriteMask
= tempWrites
[inst
->dst
.index
];
2987 tempWrites
[inst
->dst
.index
] |= inst
->dst
.writemask
;
2990 /* For a CMP to be considered a conditional write, the destination
2991 * register and source register two must be the same. */
2992 if (inst
->op
== TGSI_OPCODE_CMP
2993 && !(inst
->dst
.writemask
& prevWriteMask
)
2994 && inst
->src
[2].file
== inst
->dst
.file
2995 && inst
->src
[2].index
== inst
->dst
.index
2996 && inst
->dst
.writemask
== get_src_arg_mask(inst
->dst
, inst
->src
[2])) {
2998 inst
->op
= TGSI_OPCODE_MOV
;
2999 inst
->src
[0] = inst
->src
[1];
3004 /* Replaces all references to a temporary register index with another index. */
3006 glsl_to_tgsi_visitor::rename_temp_register(int index
, int new_index
)
3008 foreach_iter(exec_list_iterator
, iter
, this->instructions
) {
3009 glsl_to_tgsi_instruction
*inst
= (glsl_to_tgsi_instruction
*)iter
.get();
3012 for (j
=0; j
< num_inst_src_regs(inst
->op
); j
++) {
3013 if (inst
->src
[j
].file
== PROGRAM_TEMPORARY
&&
3014 inst
->src
[j
].index
== index
) {
3015 inst
->src
[j
].index
= new_index
;
3019 if (inst
->dst
.file
== PROGRAM_TEMPORARY
&& inst
->dst
.index
== index
) {
3020 inst
->dst
.index
= new_index
;
3026 glsl_to_tgsi_visitor::get_first_temp_read(int index
)
3028 int depth
= 0; /* loop depth */
3029 int loop_start
= -1; /* index of the first active BGNLOOP (if any) */
3032 foreach_iter(exec_list_iterator
, iter
, this->instructions
) {
3033 glsl_to_tgsi_instruction
*inst
= (glsl_to_tgsi_instruction
*)iter
.get();
3035 for (j
=0; j
< num_inst_src_regs(inst
->op
); j
++) {
3036 if (inst
->src
[j
].file
== PROGRAM_TEMPORARY
&&
3037 inst
->src
[j
].index
== index
) {
3038 return (depth
== 0) ? i
: loop_start
;
3042 if (inst
->op
== TGSI_OPCODE_BGNLOOP
) {
3045 } else if (inst
->op
== TGSI_OPCODE_ENDLOOP
) {
3058 glsl_to_tgsi_visitor::get_first_temp_write(int index
)
3060 int depth
= 0; /* loop depth */
3061 int loop_start
= -1; /* index of the first active BGNLOOP (if any) */
3064 foreach_iter(exec_list_iterator
, iter
, this->instructions
) {
3065 glsl_to_tgsi_instruction
*inst
= (glsl_to_tgsi_instruction
*)iter
.get();
3067 if (inst
->dst
.file
== PROGRAM_TEMPORARY
&& inst
->dst
.index
== index
) {
3068 return (depth
== 0) ? i
: loop_start
;
3071 if (inst
->op
== TGSI_OPCODE_BGNLOOP
) {
3074 } else if (inst
->op
== TGSI_OPCODE_ENDLOOP
) {
3087 glsl_to_tgsi_visitor::get_last_temp_read(int index
)
3089 int depth
= 0; /* loop depth */
3090 int last
= -1; /* index of last instruction that reads the temporary */
3093 foreach_iter(exec_list_iterator
, iter
, this->instructions
) {
3094 glsl_to_tgsi_instruction
*inst
= (glsl_to_tgsi_instruction
*)iter
.get();
3096 for (j
=0; j
< num_inst_src_regs(inst
->op
); j
++) {
3097 if (inst
->src
[j
].file
== PROGRAM_TEMPORARY
&&
3098 inst
->src
[j
].index
== index
) {
3099 last
= (depth
== 0) ? i
: -2;
3103 if (inst
->op
== TGSI_OPCODE_BGNLOOP
)
3105 else if (inst
->op
== TGSI_OPCODE_ENDLOOP
)
3106 if (--depth
== 0 && last
== -2)
3118 glsl_to_tgsi_visitor::get_last_temp_write(int index
)
3120 int depth
= 0; /* loop depth */
3121 int last
= -1; /* index of last instruction that writes to the temporary */
3124 foreach_iter(exec_list_iterator
, iter
, this->instructions
) {
3125 glsl_to_tgsi_instruction
*inst
= (glsl_to_tgsi_instruction
*)iter
.get();
3127 if (inst
->dst
.file
== PROGRAM_TEMPORARY
&& inst
->dst
.index
== index
)
3128 last
= (depth
== 0) ? i
: -2;
3130 if (inst
->op
== TGSI_OPCODE_BGNLOOP
)
3132 else if (inst
->op
== TGSI_OPCODE_ENDLOOP
)
3133 if (--depth
== 0 && last
== -2)
3145 * On a basic block basis, tracks available PROGRAM_TEMPORARY register
3146 * channels for copy propagation and updates following instructions to
3147 * use the original versions.
3149 * The glsl_to_tgsi_visitor lazily produces code assuming that this pass
3150 * will occur. As an example, a TXP production before this pass:
3152 * 0: MOV TEMP[1], INPUT[4].xyyy;
3153 * 1: MOV TEMP[1].w, INPUT[4].wwww;
3154 * 2: TXP TEMP[2], TEMP[1], texture[0], 2D;
3158 * 0: MOV TEMP[1], INPUT[4].xyyy;
3159 * 1: MOV TEMP[1].w, INPUT[4].wwww;
3160 * 2: TXP TEMP[2], INPUT[4].xyyw, texture[0], 2D;
3162 * which allows for dead code elimination on TEMP[1]'s writes.
3165 glsl_to_tgsi_visitor::copy_propagate(void)
3167 glsl_to_tgsi_instruction
**acp
= rzalloc_array(mem_ctx
,
3168 glsl_to_tgsi_instruction
*,
3169 this->next_temp
* 4);
3170 int *acp_level
= rzalloc_array(mem_ctx
, int, this->next_temp
* 4);
3173 foreach_iter(exec_list_iterator
, iter
, this->instructions
) {
3174 glsl_to_tgsi_instruction
*inst
= (glsl_to_tgsi_instruction
*)iter
.get();
3176 assert(inst
->dst
.file
!= PROGRAM_TEMPORARY
3177 || inst
->dst
.index
< this->next_temp
);
3179 /* First, do any copy propagation possible into the src regs. */
3180 for (int r
= 0; r
< 3; r
++) {
3181 glsl_to_tgsi_instruction
*first
= NULL
;
3183 int acp_base
= inst
->src
[r
].index
* 4;
3185 if (inst
->src
[r
].file
!= PROGRAM_TEMPORARY
||
3186 inst
->src
[r
].reladdr
)
3189 /* See if we can find entries in the ACP consisting of MOVs
3190 * from the same src register for all the swizzled channels
3191 * of this src register reference.
3193 for (int i
= 0; i
< 4; i
++) {
3194 int src_chan
= GET_SWZ(inst
->src
[r
].swizzle
, i
);
3195 glsl_to_tgsi_instruction
*copy_chan
= acp
[acp_base
+ src_chan
];
3202 assert(acp_level
[acp_base
+ src_chan
] <= level
);
3207 if (first
->src
[0].file
!= copy_chan
->src
[0].file
||
3208 first
->src
[0].index
!= copy_chan
->src
[0].index
) {
3216 /* We've now validated that we can copy-propagate to
3217 * replace this src register reference. Do it.
3219 inst
->src
[r
].file
= first
->src
[0].file
;
3220 inst
->src
[r
].index
= first
->src
[0].index
;
3223 for (int i
= 0; i
< 4; i
++) {
3224 int src_chan
= GET_SWZ(inst
->src
[r
].swizzle
, i
);
3225 glsl_to_tgsi_instruction
*copy_inst
= acp
[acp_base
+ src_chan
];
3226 swizzle
|= (GET_SWZ(copy_inst
->src
[0].swizzle
, src_chan
) <<
3229 inst
->src
[r
].swizzle
= swizzle
;
3234 case TGSI_OPCODE_BGNLOOP
:
3235 case TGSI_OPCODE_ENDLOOP
:
3236 /* End of a basic block, clear the ACP entirely. */
3237 memset(acp
, 0, sizeof(*acp
) * this->next_temp
* 4);
3240 case TGSI_OPCODE_IF
:
3244 case TGSI_OPCODE_ENDIF
:
3245 case TGSI_OPCODE_ELSE
:
3246 /* Clear all channels written inside the block from the ACP, but
3247 * leaving those that were not touched.
3249 for (int r
= 0; r
< this->next_temp
; r
++) {
3250 for (int c
= 0; c
< 4; c
++) {
3251 if (!acp
[4 * r
+ c
])
3254 if (acp_level
[4 * r
+ c
] >= level
)
3255 acp
[4 * r
+ c
] = NULL
;
3258 if (inst
->op
== TGSI_OPCODE_ENDIF
)
3263 /* Continuing the block, clear any written channels from
3266 if (inst
->dst
.file
== PROGRAM_TEMPORARY
&& inst
->dst
.reladdr
) {
3267 /* Any temporary might be written, so no copy propagation
3268 * across this instruction.
3270 memset(acp
, 0, sizeof(*acp
) * this->next_temp
* 4);
3271 } else if (inst
->dst
.file
== PROGRAM_OUTPUT
&&
3272 inst
->dst
.reladdr
) {
3273 /* Any output might be written, so no copy propagation
3274 * from outputs across this instruction.
3276 for (int r
= 0; r
< this->next_temp
; r
++) {
3277 for (int c
= 0; c
< 4; c
++) {
3278 if (!acp
[4 * r
+ c
])
3281 if (acp
[4 * r
+ c
]->src
[0].file
== PROGRAM_OUTPUT
)
3282 acp
[4 * r
+ c
] = NULL
;
3285 } else if (inst
->dst
.file
== PROGRAM_TEMPORARY
||
3286 inst
->dst
.file
== PROGRAM_OUTPUT
) {
3287 /* Clear where it's used as dst. */
3288 if (inst
->dst
.file
== PROGRAM_TEMPORARY
) {
3289 for (int c
= 0; c
< 4; c
++) {
3290 if (inst
->dst
.writemask
& (1 << c
)) {
3291 acp
[4 * inst
->dst
.index
+ c
] = NULL
;
3296 /* Clear where it's used as src. */
3297 for (int r
= 0; r
< this->next_temp
; r
++) {
3298 for (int c
= 0; c
< 4; c
++) {
3299 if (!acp
[4 * r
+ c
])
3302 int src_chan
= GET_SWZ(acp
[4 * r
+ c
]->src
[0].swizzle
, c
);
3304 if (acp
[4 * r
+ c
]->src
[0].file
== inst
->dst
.file
&&
3305 acp
[4 * r
+ c
]->src
[0].index
== inst
->dst
.index
&&
3306 inst
->dst
.writemask
& (1 << src_chan
))
3308 acp
[4 * r
+ c
] = NULL
;
3316 /* If this is a copy, add it to the ACP. */
3317 if (inst
->op
== TGSI_OPCODE_MOV
&&
3318 inst
->dst
.file
== PROGRAM_TEMPORARY
&&
3319 !inst
->dst
.reladdr
&&
3321 !inst
->src
[0].reladdr
&&
3322 !inst
->src
[0].negate
) {
3323 for (int i
= 0; i
< 4; i
++) {
3324 if (inst
->dst
.writemask
& (1 << i
)) {
3325 acp
[4 * inst
->dst
.index
+ i
] = inst
;
3326 acp_level
[4 * inst
->dst
.index
+ i
] = level
;
3332 ralloc_free(acp_level
);
3337 * Tracks available PROGRAM_TEMPORARY registers for dead code elimination.
3339 * The glsl_to_tgsi_visitor lazily produces code assuming that this pass
3340 * will occur. As an example, a TXP production after copy propagation but
3343 * 0: MOV TEMP[1], INPUT[4].xyyy;
3344 * 1: MOV TEMP[1].w, INPUT[4].wwww;
3345 * 2: TXP TEMP[2], INPUT[4].xyyw, texture[0], 2D;
3347 * and after this pass:
3349 * 0: TXP TEMP[2], INPUT[4].xyyw, texture[0], 2D;
3351 * FIXME: assumes that all functions are inlined (no support for BGNSUB/ENDSUB)
3352 * FIXME: doesn't eliminate all dead code inside of loops; it steps around them
3355 glsl_to_tgsi_visitor::eliminate_dead_code(void)
3359 for (i
=0; i
< this->next_temp
; i
++) {
3360 int last_read
= get_last_temp_read(i
);
3363 foreach_iter(exec_list_iterator
, iter
, this->instructions
) {
3364 glsl_to_tgsi_instruction
*inst
= (glsl_to_tgsi_instruction
*)iter
.get();
3366 if (inst
->dst
.file
== PROGRAM_TEMPORARY
&& inst
->dst
.index
== i
&&
3379 * On a basic block basis, tracks available PROGRAM_TEMPORARY registers for dead
3380 * code elimination. This is less primitive than eliminate_dead_code(), as it
3381 * is per-channel and can detect consecutive writes without a read between them
3382 * as dead code. However, there is some dead code that can be eliminated by
3383 * eliminate_dead_code() but not this function - for example, this function
3384 * cannot eliminate an instruction writing to a register that is never read and
3385 * is the only instruction writing to that register.
3387 * The glsl_to_tgsi_visitor lazily produces code assuming that this pass
3391 glsl_to_tgsi_visitor::eliminate_dead_code_advanced(void)
3393 glsl_to_tgsi_instruction
**writes
= rzalloc_array(mem_ctx
,
3394 glsl_to_tgsi_instruction
*,
3395 this->next_temp
* 4);
3396 int *write_level
= rzalloc_array(mem_ctx
, int, this->next_temp
* 4);
3400 foreach_iter(exec_list_iterator
, iter
, this->instructions
) {
3401 glsl_to_tgsi_instruction
*inst
= (glsl_to_tgsi_instruction
*)iter
.get();
3403 assert(inst
->dst
.file
!= PROGRAM_TEMPORARY
3404 || inst
->dst
.index
< this->next_temp
);
3407 case TGSI_OPCODE_BGNLOOP
:
3408 case TGSI_OPCODE_ENDLOOP
:
3409 /* End of a basic block, clear the write array entirely.
3410 * FIXME: This keeps us from killing dead code when the writes are
3411 * on either side of a loop, even when the register isn't touched
3414 memset(writes
, 0, sizeof(*writes
) * this->next_temp
* 4);
3417 case TGSI_OPCODE_ENDIF
:
3421 case TGSI_OPCODE_ELSE
:
3422 /* Clear all channels written inside the preceding if block from the
3423 * write array, but leave those that were not touched.
3425 * FIXME: This destroys opportunities to remove dead code inside of
3426 * IF blocks that are followed by an ELSE block.
3428 for (int r
= 0; r
< this->next_temp
; r
++) {
3429 for (int c
= 0; c
< 4; c
++) {
3430 if (!writes
[4 * r
+ c
])
3433 if (write_level
[4 * r
+ c
] >= level
)
3434 writes
[4 * r
+ c
] = NULL
;
3439 case TGSI_OPCODE_IF
:
3441 /* fallthrough to default case to mark the condition as read */
3444 /* Continuing the block, clear any channels from the write array that
3445 * are read by this instruction.
3447 for (unsigned i
= 0; i
< Elements(inst
->src
); i
++) {
3448 if (inst
->src
[i
].file
== PROGRAM_TEMPORARY
&& inst
->src
[i
].reladdr
){
3449 /* Any temporary might be read, so no dead code elimination
3450 * across this instruction.
3452 memset(writes
, 0, sizeof(*writes
) * this->next_temp
* 4);
3453 } else if (inst
->src
[i
].file
== PROGRAM_TEMPORARY
) {
3454 /* Clear where it's used as src. */
3455 int src_chans
= 1 << GET_SWZ(inst
->src
[i
].swizzle
, 0);
3456 src_chans
|= 1 << GET_SWZ(inst
->src
[i
].swizzle
, 1);
3457 src_chans
|= 1 << GET_SWZ(inst
->src
[i
].swizzle
, 2);
3458 src_chans
|= 1 << GET_SWZ(inst
->src
[i
].swizzle
, 3);
3460 for (int c
= 0; c
< 4; c
++) {
3461 if (src_chans
& (1 << c
)) {
3462 writes
[4 * inst
->src
[i
].index
+ c
] = NULL
;
3470 /* If this instruction writes to a temporary, add it to the write array.
3471 * If there is already an instruction in the write array for one or more
3472 * of the channels, flag that channel write as dead.
3474 if (inst
->dst
.file
== PROGRAM_TEMPORARY
&&
3475 !inst
->dst
.reladdr
&&
3477 for (int c
= 0; c
< 4; c
++) {
3478 if (inst
->dst
.writemask
& (1 << c
)) {
3479 if (writes
[4 * inst
->dst
.index
+ c
]) {
3480 if (write_level
[4 * inst
->dst
.index
+ c
] < level
)
3483 writes
[4 * inst
->dst
.index
+ c
]->dead_mask
|= (1 << c
);
3485 writes
[4 * inst
->dst
.index
+ c
] = inst
;
3486 write_level
[4 * inst
->dst
.index
+ c
] = level
;
3492 /* Anything still in the write array at this point is dead code. */
3493 for (int r
= 0; r
< this->next_temp
; r
++) {
3494 for (int c
= 0; c
< 4; c
++) {
3495 glsl_to_tgsi_instruction
*inst
= writes
[4 * r
+ c
];
3497 inst
->dead_mask
|= (1 << c
);
3501 /* Now actually remove the instructions that are completely dead and update
3502 * the writemask of other instructions with dead channels.
3504 foreach_iter(exec_list_iterator
, iter
, this->instructions
) {
3505 glsl_to_tgsi_instruction
*inst
= (glsl_to_tgsi_instruction
*)iter
.get();
3507 if (!inst
->dead_mask
|| !inst
->dst
.writemask
)
3509 else if (inst
->dead_mask
== inst
->dst
.writemask
) {
3514 inst
->dst
.writemask
&= ~(inst
->dead_mask
);
3517 ralloc_free(write_level
);
3518 ralloc_free(writes
);
3523 /* Merges temporary registers together where possible to reduce the number of
3524 * registers needed to run a program.
3526 * Produces optimal code only after copy propagation and dead code elimination
3529 glsl_to_tgsi_visitor::merge_registers(void)
3531 int *last_reads
= rzalloc_array(mem_ctx
, int, this->next_temp
);
3532 int *first_writes
= rzalloc_array(mem_ctx
, int, this->next_temp
);
3535 /* Read the indices of the last read and first write to each temp register
3536 * into an array so that we don't have to traverse the instruction list as
3538 for (i
=0; i
< this->next_temp
; i
++) {
3539 last_reads
[i
] = get_last_temp_read(i
);
3540 first_writes
[i
] = get_first_temp_write(i
);
3543 /* Start looking for registers with non-overlapping usages that can be
3544 * merged together. */
3545 for (i
=0; i
< this->next_temp
; i
++) {
3546 /* Don't touch unused registers. */
3547 if (last_reads
[i
] < 0 || first_writes
[i
] < 0) continue;
3549 for (j
=0; j
< this->next_temp
; j
++) {
3550 /* Don't touch unused registers. */
3551 if (last_reads
[j
] < 0 || first_writes
[j
] < 0) continue;
3553 /* We can merge the two registers if the first write to j is after or
3554 * in the same instruction as the last read from i. Note that the
3555 * register at index i will always be used earlier or at the same time
3556 * as the register at index j. */
3557 if (first_writes
[i
] <= first_writes
[j
] &&
3558 last_reads
[i
] <= first_writes
[j
])
3560 rename_temp_register(j
, i
); /* Replace all references to j with i.*/
3562 /* Update the first_writes and last_reads arrays with the new
3563 * values for the merged register index, and mark the newly unused
3564 * register index as such. */
3565 last_reads
[i
] = last_reads
[j
];
3566 first_writes
[j
] = -1;
3572 ralloc_free(last_reads
);
3573 ralloc_free(first_writes
);
3576 /* Reassign indices to temporary registers by reusing unused indices created
3577 * by optimization passes. */
3579 glsl_to_tgsi_visitor::renumber_registers(void)
3584 for (i
=0; i
< this->next_temp
; i
++) {
3585 if (get_first_temp_read(i
) < 0) continue;
3587 rename_temp_register(i
, new_index
);
3591 this->next_temp
= new_index
;
3595 * Returns a fragment program which implements the current pixel transfer ops.
3596 * Based on get_pixel_transfer_program in st_atom_pixeltransfer.c.
3599 get_pixel_transfer_visitor(struct st_fragment_program
*fp
,
3600 glsl_to_tgsi_visitor
*original
,
3601 int scale_and_bias
, int pixel_maps
)
3603 glsl_to_tgsi_visitor
*v
= new glsl_to_tgsi_visitor();
3604 struct st_context
*st
= st_context(original
->ctx
);
3605 struct gl_program
*prog
= &fp
->Base
.Base
;
3606 struct gl_program_parameter_list
*params
= _mesa_new_parameter_list();
3607 st_src_reg coord
, src0
;
3609 glsl_to_tgsi_instruction
*inst
;
3611 /* Copy attributes of the glsl_to_tgsi_visitor in the original shader. */
3612 v
->ctx
= original
->ctx
;
3614 v
->glsl_version
= original
->glsl_version
;
3615 v
->native_integers
= original
->native_integers
;
3616 v
->options
= original
->options
;
3617 v
->next_temp
= original
->next_temp
;
3618 v
->num_address_regs
= original
->num_address_regs
;
3619 v
->samplers_used
= prog
->SamplersUsed
= original
->samplers_used
;
3620 v
->indirect_addr_temps
= original
->indirect_addr_temps
;
3621 v
->indirect_addr_consts
= original
->indirect_addr_consts
;
3622 memcpy(&v
->immediates
, &original
->immediates
, sizeof(v
->immediates
));
3625 * Get initial pixel color from the texture.
3626 * TEX colorTemp, fragment.texcoord[0], texture[0], 2D;
3628 coord
= st_src_reg(PROGRAM_INPUT
, FRAG_ATTRIB_TEX0
, glsl_type::vec2_type
);
3629 src0
= v
->get_temp(glsl_type::vec4_type
);
3630 dst0
= st_dst_reg(src0
);
3631 inst
= v
->emit(NULL
, TGSI_OPCODE_TEX
, dst0
, coord
);
3633 inst
->tex_target
= TEXTURE_2D_INDEX
;
3635 prog
->InputsRead
|= (1 << FRAG_ATTRIB_TEX0
);
3636 prog
->SamplersUsed
|= (1 << 0); /* mark sampler 0 as used */
3637 v
->samplers_used
|= (1 << 0);
3639 if (scale_and_bias
) {
3640 static const gl_state_index scale_state
[STATE_LENGTH
] =
3641 { STATE_INTERNAL
, STATE_PT_SCALE
,
3642 (gl_state_index
) 0, (gl_state_index
) 0, (gl_state_index
) 0 };
3643 static const gl_state_index bias_state
[STATE_LENGTH
] =
3644 { STATE_INTERNAL
, STATE_PT_BIAS
,
3645 (gl_state_index
) 0, (gl_state_index
) 0, (gl_state_index
) 0 };
3646 GLint scale_p
, bias_p
;
3647 st_src_reg scale
, bias
;
3649 scale_p
= _mesa_add_state_reference(params
, scale_state
);
3650 bias_p
= _mesa_add_state_reference(params
, bias_state
);
3652 /* MAD colorTemp, colorTemp, scale, bias; */
3653 scale
= st_src_reg(PROGRAM_STATE_VAR
, scale_p
, GLSL_TYPE_FLOAT
);
3654 bias
= st_src_reg(PROGRAM_STATE_VAR
, bias_p
, GLSL_TYPE_FLOAT
);
3655 inst
= v
->emit(NULL
, TGSI_OPCODE_MAD
, dst0
, src0
, scale
, bias
);
3659 st_src_reg temp
= v
->get_temp(glsl_type::vec4_type
);
3660 st_dst_reg temp_dst
= st_dst_reg(temp
);
3662 assert(st
->pixel_xfer
.pixelmap_texture
);
3664 /* With a little effort, we can do four pixel map look-ups with
3665 * two TEX instructions:
3668 /* TEX temp.rg, colorTemp.rgba, texture[1], 2D; */
3669 temp_dst
.writemask
= WRITEMASK_XY
; /* write R,G */
3670 inst
= v
->emit(NULL
, TGSI_OPCODE_TEX
, temp_dst
, src0
);
3672 inst
->tex_target
= TEXTURE_2D_INDEX
;
3674 /* TEX temp.ba, colorTemp.baba, texture[1], 2D; */
3675 src0
.swizzle
= MAKE_SWIZZLE4(SWIZZLE_Z
, SWIZZLE_W
, SWIZZLE_Z
, SWIZZLE_W
);
3676 temp_dst
.writemask
= WRITEMASK_ZW
; /* write B,A */
3677 inst
= v
->emit(NULL
, TGSI_OPCODE_TEX
, temp_dst
, src0
);
3679 inst
->tex_target
= TEXTURE_2D_INDEX
;
3681 prog
->SamplersUsed
|= (1 << 1); /* mark sampler 1 as used */
3682 v
->samplers_used
|= (1 << 1);
3684 /* MOV colorTemp, temp; */
3685 inst
= v
->emit(NULL
, TGSI_OPCODE_MOV
, dst0
, temp
);
3688 /* Now copy the instructions from the original glsl_to_tgsi_visitor into the
3690 foreach_iter(exec_list_iterator
, iter
, original
->instructions
) {
3691 glsl_to_tgsi_instruction
*inst
= (glsl_to_tgsi_instruction
*)iter
.get();
3692 st_src_reg src_regs
[3];
3694 if (inst
->dst
.file
== PROGRAM_OUTPUT
)
3695 prog
->OutputsWritten
|= BITFIELD64_BIT(inst
->dst
.index
);
3697 for (int i
=0; i
<3; i
++) {
3698 src_regs
[i
] = inst
->src
[i
];
3699 if (src_regs
[i
].file
== PROGRAM_INPUT
&&
3700 src_regs
[i
].index
== FRAG_ATTRIB_COL0
)
3702 src_regs
[i
].file
= PROGRAM_TEMPORARY
;
3703 src_regs
[i
].index
= src0
.index
;
3705 else if (src_regs
[i
].file
== PROGRAM_INPUT
)
3706 prog
->InputsRead
|= (1 << src_regs
[i
].index
);
3709 v
->emit(NULL
, inst
->op
, inst
->dst
, src_regs
[0], src_regs
[1], src_regs
[2]);
3712 /* Make modifications to fragment program info. */
3713 prog
->Parameters
= _mesa_combine_parameter_lists(params
,
3714 original
->prog
->Parameters
);
3715 prog
->Attributes
= _mesa_clone_parameter_list(original
->prog
->Attributes
);
3716 prog
->Varying
= _mesa_clone_parameter_list(original
->prog
->Varying
);
3717 _mesa_free_parameter_list(params
);
3718 count_resources(v
, prog
);
3719 fp
->glsl_to_tgsi
= v
;
3723 * Make fragment program for glBitmap:
3724 * Sample the texture and kill the fragment if the bit is 0.
3725 * This program will be combined with the user's fragment program.
3727 * Based on make_bitmap_fragment_program in st_cb_bitmap.c.
3730 get_bitmap_visitor(struct st_fragment_program
*fp
,
3731 glsl_to_tgsi_visitor
*original
, int samplerIndex
)
3733 glsl_to_tgsi_visitor
*v
= new glsl_to_tgsi_visitor();
3734 struct st_context
*st
= st_context(original
->ctx
);
3735 struct gl_program
*prog
= &fp
->Base
.Base
;
3736 st_src_reg coord
, src0
;
3738 glsl_to_tgsi_instruction
*inst
;
3740 /* Copy attributes of the glsl_to_tgsi_visitor in the original shader. */
3741 v
->ctx
= original
->ctx
;
3743 v
->glsl_version
= original
->glsl_version
;
3744 v
->native_integers
= original
->native_integers
;
3745 v
->options
= original
->options
;
3746 v
->next_temp
= original
->next_temp
;
3747 v
->num_address_regs
= original
->num_address_regs
;
3748 v
->samplers_used
= prog
->SamplersUsed
= original
->samplers_used
;
3749 v
->indirect_addr_temps
= original
->indirect_addr_temps
;
3750 v
->indirect_addr_consts
= original
->indirect_addr_consts
;
3751 memcpy(&v
->immediates
, &original
->immediates
, sizeof(v
->immediates
));
3753 /* TEX tmp0, fragment.texcoord[0], texture[0], 2D; */
3754 coord
= st_src_reg(PROGRAM_INPUT
, FRAG_ATTRIB_TEX0
, glsl_type::vec2_type
);
3755 src0
= v
->get_temp(glsl_type::vec4_type
);
3756 dst0
= st_dst_reg(src0
);
3757 inst
= v
->emit(NULL
, TGSI_OPCODE_TEX
, dst0
, coord
);
3758 inst
->sampler
= samplerIndex
;
3759 inst
->tex_target
= TEXTURE_2D_INDEX
;
3761 prog
->InputsRead
|= (1 << FRAG_ATTRIB_TEX0
);
3762 prog
->SamplersUsed
|= (1 << samplerIndex
); /* mark sampler as used */
3763 v
->samplers_used
|= (1 << samplerIndex
);
3765 /* KIL if -tmp0 < 0 # texel=0 -> keep / texel=0 -> discard */
3766 src0
.negate
= NEGATE_XYZW
;
3767 if (st
->bitmap
.tex_format
== PIPE_FORMAT_L8_UNORM
)
3768 src0
.swizzle
= SWIZZLE_XXXX
;
3769 inst
= v
->emit(NULL
, TGSI_OPCODE_KIL
, undef_dst
, src0
);
3771 /* Now copy the instructions from the original glsl_to_tgsi_visitor into the
3773 foreach_iter(exec_list_iterator
, iter
, original
->instructions
) {
3774 glsl_to_tgsi_instruction
*inst
= (glsl_to_tgsi_instruction
*)iter
.get();
3775 st_src_reg src_regs
[3];
3777 if (inst
->dst
.file
== PROGRAM_OUTPUT
)
3778 prog
->OutputsWritten
|= BITFIELD64_BIT(inst
->dst
.index
);
3780 for (int i
=0; i
<3; i
++) {
3781 src_regs
[i
] = inst
->src
[i
];
3782 if (src_regs
[i
].file
== PROGRAM_INPUT
)
3783 prog
->InputsRead
|= (1 << src_regs
[i
].index
);
3786 v
->emit(NULL
, inst
->op
, inst
->dst
, src_regs
[0], src_regs
[1], src_regs
[2]);
3789 /* Make modifications to fragment program info. */
3790 prog
->Parameters
= _mesa_clone_parameter_list(original
->prog
->Parameters
);
3791 prog
->Attributes
= _mesa_clone_parameter_list(original
->prog
->Attributes
);
3792 prog
->Varying
= _mesa_clone_parameter_list(original
->prog
->Varying
);
3793 count_resources(v
, prog
);
3794 fp
->glsl_to_tgsi
= v
;
3797 /* ------------------------- TGSI conversion stuff -------------------------- */
3799 unsigned branch_target
;
3804 * Intermediate state used during shader translation.
3806 struct st_translate
{
3807 struct ureg_program
*ureg
;
3809 struct ureg_dst temps
[MAX_TEMPS
];
3810 struct ureg_src
*constants
;
3811 struct ureg_src
*immediates
;
3812 struct ureg_dst outputs
[PIPE_MAX_SHADER_OUTPUTS
];
3813 struct ureg_src inputs
[PIPE_MAX_SHADER_INPUTS
];
3814 struct ureg_dst address
[1];
3815 struct ureg_src samplers
[PIPE_MAX_SAMPLERS
];
3816 struct ureg_src systemValues
[SYSTEM_VALUE_MAX
];
3818 /* Extra info for handling point size clamping in vertex shader */
3819 struct ureg_dst pointSizeResult
; /**< Actual point size output register */
3820 struct ureg_src pointSizeConst
; /**< Point size range constant register */
3821 GLint pointSizeOutIndex
; /**< Temp point size output register */
3822 GLboolean prevInstWrotePointSize
;
3824 const GLuint
*inputMapping
;
3825 const GLuint
*outputMapping
;
3827 /* For every instruction that contains a label (eg CALL), keep
3828 * details so that we can go back afterwards and emit the correct
3829 * tgsi instruction number for each label.
3831 struct label
*labels
;
3832 unsigned labels_size
;
3833 unsigned labels_count
;
3835 /* Keep a record of the tgsi instruction number that each mesa
3836 * instruction starts at, will be used to fix up labels after
3841 unsigned insn_count
;
3843 unsigned procType
; /**< TGSI_PROCESSOR_VERTEX/FRAGMENT */
3848 /** Map Mesa's SYSTEM_VALUE_x to TGSI_SEMANTIC_x */
3849 static unsigned mesa_sysval_to_semantic
[SYSTEM_VALUE_MAX
] = {
3851 TGSI_SEMANTIC_INSTANCEID
3855 * Make note of a branch to a label in the TGSI code.
3856 * After we've emitted all instructions, we'll go over the list
3857 * of labels built here and patch the TGSI code with the actual
3858 * location of each label.
3860 static unsigned *get_label(struct st_translate
*t
, unsigned branch_target
)
3864 if (t
->labels_count
+ 1 >= t
->labels_size
) {
3865 t
->labels_size
= 1 << (util_logbase2(t
->labels_size
) + 1);
3866 t
->labels
= (struct label
*)realloc(t
->labels
,
3867 t
->labels_size
* sizeof(struct label
));
3868 if (t
->labels
== NULL
) {
3869 static unsigned dummy
;
3875 i
= t
->labels_count
++;
3876 t
->labels
[i
].branch_target
= branch_target
;
3877 return &t
->labels
[i
].token
;
3881 * Called prior to emitting the TGSI code for each instruction.
3882 * Allocate additional space for instructions if needed.
3883 * Update the insn[] array so the next glsl_to_tgsi_instruction points to
3884 * the next TGSI instruction.
3886 static void set_insn_start(struct st_translate
*t
, unsigned start
)
3888 if (t
->insn_count
+ 1 >= t
->insn_size
) {
3889 t
->insn_size
= 1 << (util_logbase2(t
->insn_size
) + 1);
3890 t
->insn
= (unsigned *)realloc(t
->insn
, t
->insn_size
* sizeof(t
->insn
[0]));
3891 if (t
->insn
== NULL
) {
3897 t
->insn
[t
->insn_count
++] = start
;
3901 * Map a glsl_to_tgsi constant/immediate to a TGSI immediate.
3903 static struct ureg_src
3904 emit_immediate(struct st_translate
*t
,
3905 gl_constant_value values
[4],
3908 struct ureg_program
*ureg
= t
->ureg
;
3913 return ureg_DECL_immediate(ureg
, &values
[0].f
, size
);
3915 return ureg_DECL_immediate_int(ureg
, &values
[0].i
, size
);
3916 case GL_UNSIGNED_INT
:
3918 return ureg_DECL_immediate_uint(ureg
, &values
[0].u
, size
);
3920 assert(!"should not get here - type must be float, int, uint, or bool");
3921 return ureg_src_undef();
3926 * Map a glsl_to_tgsi dst register to a TGSI ureg_dst register.
3928 static struct ureg_dst
3929 dst_register(struct st_translate
*t
,
3930 gl_register_file file
,
3934 case PROGRAM_UNDEFINED
:
3935 return ureg_dst_undef();
3937 case PROGRAM_TEMPORARY
:
3938 if (ureg_dst_is_undef(t
->temps
[index
]))
3939 t
->temps
[index
] = ureg_DECL_temporary(t
->ureg
);
3941 return t
->temps
[index
];
3943 case PROGRAM_OUTPUT
:
3944 if (t
->procType
== TGSI_PROCESSOR_VERTEX
&& index
== VERT_RESULT_PSIZ
)
3945 t
->prevInstWrotePointSize
= GL_TRUE
;
3947 if (t
->procType
== TGSI_PROCESSOR_VERTEX
)
3948 assert(index
< VERT_RESULT_MAX
);
3949 else if (t
->procType
== TGSI_PROCESSOR_FRAGMENT
)
3950 assert(index
< FRAG_RESULT_MAX
);
3952 assert(index
< GEOM_RESULT_MAX
);
3954 assert(t
->outputMapping
[index
] < Elements(t
->outputs
));
3956 return t
->outputs
[t
->outputMapping
[index
]];
3958 case PROGRAM_ADDRESS
:
3959 return t
->address
[index
];
3962 assert(!"unknown dst register file");
3963 return ureg_dst_undef();
3968 * Map a glsl_to_tgsi src register to a TGSI ureg_src register.
3970 static struct ureg_src
3971 src_register(struct st_translate
*t
,
3972 gl_register_file file
,
3976 case PROGRAM_UNDEFINED
:
3977 return ureg_src_undef();
3979 case PROGRAM_TEMPORARY
:
3981 assert(index
< Elements(t
->temps
));
3982 if (ureg_dst_is_undef(t
->temps
[index
]))
3983 t
->temps
[index
] = ureg_DECL_temporary(t
->ureg
);
3984 return ureg_src(t
->temps
[index
]);
3986 case PROGRAM_NAMED_PARAM
:
3987 case PROGRAM_ENV_PARAM
:
3988 case PROGRAM_LOCAL_PARAM
:
3989 case PROGRAM_UNIFORM
:
3991 return t
->constants
[index
];
3992 case PROGRAM_STATE_VAR
:
3993 case PROGRAM_CONSTANT
: /* ie, immediate */
3995 return ureg_DECL_constant(t
->ureg
, 0);
3997 return t
->constants
[index
];
3999 case PROGRAM_IMMEDIATE
:
4000 return t
->immediates
[index
];
4003 assert(t
->inputMapping
[index
] < Elements(t
->inputs
));
4004 return t
->inputs
[t
->inputMapping
[index
]];
4006 case PROGRAM_OUTPUT
:
4007 assert(t
->outputMapping
[index
] < Elements(t
->outputs
));
4008 return ureg_src(t
->outputs
[t
->outputMapping
[index
]]); /* not needed? */
4010 case PROGRAM_ADDRESS
:
4011 return ureg_src(t
->address
[index
]);
4013 case PROGRAM_SYSTEM_VALUE
:
4014 assert(index
< Elements(t
->systemValues
));
4015 return t
->systemValues
[index
];
4018 assert(!"unknown src register file");
4019 return ureg_src_undef();
4024 * Create a TGSI ureg_dst register from an st_dst_reg.
4026 static struct ureg_dst
4027 translate_dst(struct st_translate
*t
,
4028 const st_dst_reg
*dst_reg
,
4031 struct ureg_dst dst
= dst_register(t
,
4035 dst
= ureg_writemask(dst
, dst_reg
->writemask
);
4038 dst
= ureg_saturate(dst
);
4040 if (dst_reg
->reladdr
!= NULL
)
4041 dst
= ureg_dst_indirect(dst
, ureg_src(t
->address
[0]));
4047 * Create a TGSI ureg_src register from an st_src_reg.
4049 static struct ureg_src
4050 translate_src(struct st_translate
*t
, const st_src_reg
*src_reg
)
4052 struct ureg_src src
= src_register(t
, src_reg
->file
, src_reg
->index
);
4054 src
= ureg_swizzle(src
,
4055 GET_SWZ(src_reg
->swizzle
, 0) & 0x3,
4056 GET_SWZ(src_reg
->swizzle
, 1) & 0x3,
4057 GET_SWZ(src_reg
->swizzle
, 2) & 0x3,
4058 GET_SWZ(src_reg
->swizzle
, 3) & 0x3);
4060 if ((src_reg
->negate
& 0xf) == NEGATE_XYZW
)
4061 src
= ureg_negate(src
);
4063 if (src_reg
->reladdr
!= NULL
) {
4064 /* Normally ureg_src_indirect() would be used here, but a stupid compiler
4065 * bug in g++ makes ureg_src_indirect (an inline C function) erroneously
4066 * set the bit for src.Negate. So we have to do the operation manually
4067 * here to work around the compiler's problems. */
4068 /*src = ureg_src_indirect(src, ureg_src(t->address[0]));*/
4069 struct ureg_src addr
= ureg_src(t
->address
[0]);
4071 src
.IndirectFile
= addr
.File
;
4072 src
.IndirectIndex
= addr
.Index
;
4073 src
.IndirectSwizzle
= addr
.SwizzleX
;
4075 if (src_reg
->file
!= PROGRAM_INPUT
&&
4076 src_reg
->file
!= PROGRAM_OUTPUT
) {
4077 /* If src_reg->index was negative, it was set to zero in
4078 * src_register(). Reassign it now. But don't do this
4079 * for input/output regs since they get remapped while
4080 * const buffers don't.
4082 src
.Index
= src_reg
->index
;
4090 compile_tgsi_instruction(struct st_translate
*t
,
4091 const struct glsl_to_tgsi_instruction
*inst
)
4093 struct ureg_program
*ureg
= t
->ureg
;
4095 struct ureg_dst dst
[1];
4096 struct ureg_src src
[4];
4100 num_dst
= num_inst_dst_regs(inst
->op
);
4101 num_src
= num_inst_src_regs(inst
->op
);
4104 dst
[0] = translate_dst(t
,
4108 for (i
= 0; i
< num_src
; i
++)
4109 src
[i
] = translate_src(t
, &inst
->src
[i
]);
4112 case TGSI_OPCODE_BGNLOOP
:
4113 case TGSI_OPCODE_CAL
:
4114 case TGSI_OPCODE_ELSE
:
4115 case TGSI_OPCODE_ENDLOOP
:
4116 case TGSI_OPCODE_IF
:
4117 assert(num_dst
== 0);
4118 ureg_label_insn(ureg
,
4122 inst
->op
== TGSI_OPCODE_CAL
? inst
->function
->sig_id
: 0));
4125 case TGSI_OPCODE_TEX
:
4126 case TGSI_OPCODE_TXB
:
4127 case TGSI_OPCODE_TXD
:
4128 case TGSI_OPCODE_TXL
:
4129 case TGSI_OPCODE_TXP
:
4130 src
[num_src
++] = t
->samplers
[inst
->sampler
];
4134 translate_texture_target(inst
->tex_target
, inst
->tex_shadow
),
4138 case TGSI_OPCODE_SCS
:
4139 dst
[0] = ureg_writemask(dst
[0], TGSI_WRITEMASK_XY
);
4140 ureg_insn(ureg
, inst
->op
, dst
, num_dst
, src
, num_src
);
4153 * Emit the TGSI instructions to adjust the WPOS pixel center convention
4154 * Basically, add (adjX, adjY) to the fragment position.
4157 emit_adjusted_wpos(struct st_translate
*t
,
4158 const struct gl_program
*program
,
4159 float adjX
, float adjY
)
4161 struct ureg_program
*ureg
= t
->ureg
;
4162 struct ureg_dst wpos_temp
= ureg_DECL_temporary(ureg
);
4163 struct ureg_src wpos_input
= t
->inputs
[t
->inputMapping
[FRAG_ATTRIB_WPOS
]];
4165 /* Note that we bias X and Y and pass Z and W through unchanged.
4166 * The shader might also use gl_FragCoord.w and .z.
4168 ureg_ADD(ureg
, wpos_temp
, wpos_input
,
4169 ureg_imm4f(ureg
, adjX
, adjY
, 0.0f
, 0.0f
));
4171 t
->inputs
[t
->inputMapping
[FRAG_ATTRIB_WPOS
]] = ureg_src(wpos_temp
);
4176 * Emit the TGSI instructions for inverting the WPOS y coordinate.
4177 * This code is unavoidable because it also depends on whether
4178 * a FBO is bound (STATE_FB_WPOS_Y_TRANSFORM).
4181 emit_wpos_inversion(struct st_translate
*t
,
4182 const struct gl_program
*program
,
4185 struct ureg_program
*ureg
= t
->ureg
;
4187 /* Fragment program uses fragment position input.
4188 * Need to replace instances of INPUT[WPOS] with temp T
4189 * where T = INPUT[WPOS] by y is inverted.
4191 static const gl_state_index wposTransformState
[STATE_LENGTH
]
4192 = { STATE_INTERNAL
, STATE_FB_WPOS_Y_TRANSFORM
,
4193 (gl_state_index
)0, (gl_state_index
)0, (gl_state_index
)0 };
4195 /* XXX: note we are modifying the incoming shader here! Need to
4196 * do this before emitting the constant decls below, or this
4199 unsigned wposTransConst
= _mesa_add_state_reference(program
->Parameters
,
4200 wposTransformState
);
4202 struct ureg_src wpostrans
= ureg_DECL_constant(ureg
, wposTransConst
);
4203 struct ureg_dst wpos_temp
;
4204 struct ureg_src wpos_input
= t
->inputs
[t
->inputMapping
[FRAG_ATTRIB_WPOS
]];
4206 /* MOV wpos_temp, input[wpos]
4208 if (wpos_input
.File
== TGSI_FILE_TEMPORARY
)
4209 wpos_temp
= ureg_dst(wpos_input
);
4211 wpos_temp
= ureg_DECL_temporary(ureg
);
4212 ureg_MOV(ureg
, wpos_temp
, wpos_input
);
4216 /* MAD wpos_temp.y, wpos_input, wpostrans.xxxx, wpostrans.yyyy
4219 ureg_writemask(wpos_temp
, TGSI_WRITEMASK_Y
),
4221 ureg_scalar(wpostrans
, 0),
4222 ureg_scalar(wpostrans
, 1));
4224 /* MAD wpos_temp.y, wpos_input, wpostrans.zzzz, wpostrans.wwww
4227 ureg_writemask(wpos_temp
, TGSI_WRITEMASK_Y
),
4229 ureg_scalar(wpostrans
, 2),
4230 ureg_scalar(wpostrans
, 3));
4233 /* Use wpos_temp as position input from here on:
4235 t
->inputs
[t
->inputMapping
[FRAG_ATTRIB_WPOS
]] = ureg_src(wpos_temp
);
4240 * Emit fragment position/ooordinate code.
4243 emit_wpos(struct st_context
*st
,
4244 struct st_translate
*t
,
4245 const struct gl_program
*program
,
4246 struct ureg_program
*ureg
)
4248 const struct gl_fragment_program
*fp
=
4249 (const struct gl_fragment_program
*) program
;
4250 struct pipe_screen
*pscreen
= st
->pipe
->screen
;
4251 boolean invert
= FALSE
;
4253 if (fp
->OriginUpperLeft
) {
4254 /* Fragment shader wants origin in upper-left */
4255 if (pscreen
->get_param(pscreen
, PIPE_CAP_TGSI_FS_COORD_ORIGIN_UPPER_LEFT
)) {
4256 /* the driver supports upper-left origin */
4258 else if (pscreen
->get_param(pscreen
, PIPE_CAP_TGSI_FS_COORD_ORIGIN_LOWER_LEFT
)) {
4259 /* the driver supports lower-left origin, need to invert Y */
4260 ureg_property_fs_coord_origin(ureg
, TGSI_FS_COORD_ORIGIN_LOWER_LEFT
);
4267 /* Fragment shader wants origin in lower-left */
4268 if (pscreen
->get_param(pscreen
, PIPE_CAP_TGSI_FS_COORD_ORIGIN_LOWER_LEFT
))
4269 /* the driver supports lower-left origin */
4270 ureg_property_fs_coord_origin(ureg
, TGSI_FS_COORD_ORIGIN_LOWER_LEFT
);
4271 else if (pscreen
->get_param(pscreen
, PIPE_CAP_TGSI_FS_COORD_ORIGIN_UPPER_LEFT
))
4272 /* the driver supports upper-left origin, need to invert Y */
4278 if (fp
->PixelCenterInteger
) {
4279 /* Fragment shader wants pixel center integer */
4280 if (pscreen
->get_param(pscreen
, PIPE_CAP_TGSI_FS_COORD_PIXEL_CENTER_INTEGER
))
4281 /* the driver supports pixel center integer */
4282 ureg_property_fs_coord_pixel_center(ureg
, TGSI_FS_COORD_PIXEL_CENTER_INTEGER
);
4283 else if (pscreen
->get_param(pscreen
, PIPE_CAP_TGSI_FS_COORD_PIXEL_CENTER_HALF_INTEGER
))
4284 /* the driver supports pixel center half integer, need to bias X,Y */
4285 emit_adjusted_wpos(t
, program
, 0.5f
, invert
? 0.5f
: -0.5f
);
4290 /* Fragment shader wants pixel center half integer */
4291 if (pscreen
->get_param(pscreen
, PIPE_CAP_TGSI_FS_COORD_PIXEL_CENTER_HALF_INTEGER
)) {
4292 /* the driver supports pixel center half integer */
4294 else if (pscreen
->get_param(pscreen
, PIPE_CAP_TGSI_FS_COORD_PIXEL_CENTER_INTEGER
)) {
4295 /* the driver supports pixel center integer, need to bias X,Y */
4296 ureg_property_fs_coord_pixel_center(ureg
, TGSI_FS_COORD_PIXEL_CENTER_INTEGER
);
4297 emit_adjusted_wpos(t
, program
, 0.5f
, invert
? -0.5f
: 0.5f
);
4303 /* we invert after adjustment so that we avoid the MOV to temporary,
4304 * and reuse the adjustment ADD instead */
4305 emit_wpos_inversion(t
, program
, invert
);
4309 * OpenGL's fragment gl_FrontFace input is 1 for front-facing, 0 for back.
4310 * TGSI uses +1 for front, -1 for back.
4311 * This function converts the TGSI value to the GL value. Simply clamping/
4312 * saturating the value to [0,1] does the job.
4315 emit_face_var(struct st_translate
*t
)
4317 struct ureg_program
*ureg
= t
->ureg
;
4318 struct ureg_dst face_temp
= ureg_DECL_temporary(ureg
);
4319 struct ureg_src face_input
= t
->inputs
[t
->inputMapping
[FRAG_ATTRIB_FACE
]];
4321 /* MOV_SAT face_temp, input[face] */
4322 face_temp
= ureg_saturate(face_temp
);
4323 ureg_MOV(ureg
, face_temp
, face_input
);
4325 /* Use face_temp as face input from here on: */
4326 t
->inputs
[t
->inputMapping
[FRAG_ATTRIB_FACE
]] = ureg_src(face_temp
);
4330 emit_edgeflags(struct st_translate
*t
)
4332 struct ureg_program
*ureg
= t
->ureg
;
4333 struct ureg_dst edge_dst
= t
->outputs
[t
->outputMapping
[VERT_RESULT_EDGE
]];
4334 struct ureg_src edge_src
= t
->inputs
[t
->inputMapping
[VERT_ATTRIB_EDGEFLAG
]];
4336 ureg_MOV(ureg
, edge_dst
, edge_src
);
4340 * Translate intermediate IR (glsl_to_tgsi_instruction) to TGSI format.
4341 * \param program the program to translate
4342 * \param numInputs number of input registers used
4343 * \param inputMapping maps Mesa fragment program inputs to TGSI generic
4345 * \param inputSemanticName the TGSI_SEMANTIC flag for each input
4346 * \param inputSemanticIndex the semantic index (ex: which texcoord) for
4348 * \param interpMode the TGSI_INTERPOLATE_LINEAR/PERSP mode for each input
4349 * \param numOutputs number of output registers used
4350 * \param outputMapping maps Mesa fragment program outputs to TGSI
4352 * \param outputSemanticName the TGSI_SEMANTIC flag for each output
4353 * \param outputSemanticIndex the semantic index (ex: which texcoord) for
4356 * \return PIPE_OK or PIPE_ERROR_OUT_OF_MEMORY
4358 extern "C" enum pipe_error
4359 st_translate_program(
4360 struct gl_context
*ctx
,
4362 struct ureg_program
*ureg
,
4363 glsl_to_tgsi_visitor
*program
,
4364 const struct gl_program
*proginfo
,
4366 const GLuint inputMapping
[],
4367 const ubyte inputSemanticName
[],
4368 const ubyte inputSemanticIndex
[],
4369 const GLuint interpMode
[],
4371 const GLuint outputMapping
[],
4372 const ubyte outputSemanticName
[],
4373 const ubyte outputSemanticIndex
[],
4374 boolean passthrough_edgeflags
)
4376 struct st_translate translate
, *t
;
4378 enum pipe_error ret
= PIPE_OK
;
4380 assert(numInputs
<= Elements(t
->inputs
));
4381 assert(numOutputs
<= Elements(t
->outputs
));
4384 memset(t
, 0, sizeof *t
);
4386 t
->procType
= procType
;
4387 t
->inputMapping
= inputMapping
;
4388 t
->outputMapping
= outputMapping
;
4390 t
->pointSizeOutIndex
= -1;
4391 t
->prevInstWrotePointSize
= GL_FALSE
;
4394 * Declare input attributes.
4396 if (procType
== TGSI_PROCESSOR_FRAGMENT
) {
4397 for (i
= 0; i
< numInputs
; i
++) {
4398 t
->inputs
[i
] = ureg_DECL_fs_input(ureg
,
4399 inputSemanticName
[i
],
4400 inputSemanticIndex
[i
],
4404 if (proginfo
->InputsRead
& FRAG_BIT_WPOS
) {
4405 /* Must do this after setting up t->inputs, and before
4406 * emitting constant references, below:
4408 emit_wpos(st_context(ctx
), t
, proginfo
, ureg
);
4411 if (proginfo
->InputsRead
& FRAG_BIT_FACE
)
4415 * Declare output attributes.
4417 for (i
= 0; i
< numOutputs
; i
++) {
4418 switch (outputSemanticName
[i
]) {
4419 case TGSI_SEMANTIC_POSITION
:
4420 t
->outputs
[i
] = ureg_DECL_output(ureg
,
4421 TGSI_SEMANTIC_POSITION
, /* Z/Depth */
4422 outputSemanticIndex
[i
]);
4423 t
->outputs
[i
] = ureg_writemask(t
->outputs
[i
], TGSI_WRITEMASK_Z
);
4425 case TGSI_SEMANTIC_STENCIL
:
4426 t
->outputs
[i
] = ureg_DECL_output(ureg
,
4427 TGSI_SEMANTIC_STENCIL
, /* Stencil */
4428 outputSemanticIndex
[i
]);
4429 t
->outputs
[i
] = ureg_writemask(t
->outputs
[i
], TGSI_WRITEMASK_Y
);
4431 case TGSI_SEMANTIC_COLOR
:
4432 t
->outputs
[i
] = ureg_DECL_output(ureg
,
4433 TGSI_SEMANTIC_COLOR
,
4434 outputSemanticIndex
[i
]);
4437 assert(!"fragment shader outputs must be POSITION/STENCIL/COLOR");
4438 return PIPE_ERROR_BAD_INPUT
;
4442 else if (procType
== TGSI_PROCESSOR_GEOMETRY
) {
4443 for (i
= 0; i
< numInputs
; i
++) {
4444 t
->inputs
[i
] = ureg_DECL_gs_input(ureg
,
4446 inputSemanticName
[i
],
4447 inputSemanticIndex
[i
]);
4450 for (i
= 0; i
< numOutputs
; i
++) {
4451 t
->outputs
[i
] = ureg_DECL_output(ureg
,
4452 outputSemanticName
[i
],
4453 outputSemanticIndex
[i
]);
4457 assert(procType
== TGSI_PROCESSOR_VERTEX
);
4459 for (i
= 0; i
< numInputs
; i
++) {
4460 t
->inputs
[i
] = ureg_DECL_vs_input(ureg
, i
);
4463 for (i
= 0; i
< numOutputs
; i
++) {
4464 t
->outputs
[i
] = ureg_DECL_output(ureg
,
4465 outputSemanticName
[i
],
4466 outputSemanticIndex
[i
]);
4467 if ((outputSemanticName
[i
] == TGSI_SEMANTIC_PSIZE
) && proginfo
->Id
) {
4468 /* Writing to the point size result register requires special
4469 * handling to implement clamping.
4471 static const gl_state_index pointSizeClampState
[STATE_LENGTH
]
4472 = { STATE_INTERNAL
, STATE_POINT_SIZE_IMPL_CLAMP
, (gl_state_index
)0, (gl_state_index
)0, (gl_state_index
)0 };
4473 /* XXX: note we are modifying the incoming shader here! Need to
4474 * do this before emitting the constant decls below, or this
4477 unsigned pointSizeClampConst
=
4478 _mesa_add_state_reference(proginfo
->Parameters
,
4479 pointSizeClampState
);
4480 struct ureg_dst psizregtemp
= ureg_DECL_temporary(ureg
);
4481 t
->pointSizeConst
= ureg_DECL_constant(ureg
, pointSizeClampConst
);
4482 t
->pointSizeResult
= t
->outputs
[i
];
4483 t
->pointSizeOutIndex
= i
;
4484 t
->outputs
[i
] = psizregtemp
;
4487 if (passthrough_edgeflags
)
4491 /* Declare address register.
4493 if (program
->num_address_regs
> 0) {
4494 assert(program
->num_address_regs
== 1);
4495 t
->address
[0] = ureg_DECL_address(ureg
);
4498 /* Declare misc input registers
4501 GLbitfield sysInputs
= proginfo
->SystemValuesRead
;
4502 unsigned numSys
= 0;
4503 for (i
= 0; sysInputs
; i
++) {
4504 if (sysInputs
& (1 << i
)) {
4505 unsigned semName
= mesa_sysval_to_semantic
[i
];
4506 t
->systemValues
[i
] = ureg_DECL_system_value(ureg
, numSys
, semName
, 0);
4508 sysInputs
&= ~(1 << i
);
4513 if (program
->indirect_addr_temps
) {
4514 /* If temps are accessed with indirect addressing, declare temporaries
4515 * in sequential order. Else, we declare them on demand elsewhere.
4516 * (Note: the number of temporaries is equal to program->next_temp)
4518 for (i
= 0; i
< (unsigned)program
->next_temp
; i
++) {
4519 /* XXX use TGSI_FILE_TEMPORARY_ARRAY when it's supported by ureg */
4520 t
->temps
[i
] = ureg_DECL_temporary(t
->ureg
);
4524 /* Emit constants and uniforms. TGSI uses a single index space for these,
4525 * so we put all the translated regs in t->constants.
4527 if (proginfo
->Parameters
) {
4528 t
->constants
= (struct ureg_src
*)CALLOC(proginfo
->Parameters
->NumParameters
* sizeof(t
->constants
[0]));
4529 if (t
->constants
== NULL
) {
4530 ret
= PIPE_ERROR_OUT_OF_MEMORY
;
4534 for (i
= 0; i
< proginfo
->Parameters
->NumParameters
; i
++) {
4535 switch (proginfo
->Parameters
->Parameters
[i
].Type
) {
4536 case PROGRAM_ENV_PARAM
:
4537 case PROGRAM_LOCAL_PARAM
:
4538 case PROGRAM_STATE_VAR
:
4539 case PROGRAM_NAMED_PARAM
:
4540 case PROGRAM_UNIFORM
:
4541 t
->constants
[i
] = ureg_DECL_constant(ureg
, i
);
4544 /* Emit immediates for PROGRAM_CONSTANT only when there's no indirect
4545 * addressing of the const buffer.
4546 * FIXME: Be smarter and recognize param arrays:
4547 * indirect addressing is only valid within the referenced
4550 case PROGRAM_CONSTANT
:
4551 if (program
->indirect_addr_consts
)
4552 t
->constants
[i
] = ureg_DECL_constant(ureg
, i
);
4554 t
->constants
[i
] = emit_immediate(t
,
4555 proginfo
->Parameters
->ParameterValues
[i
],
4556 proginfo
->Parameters
->Parameters
[i
].DataType
,
4565 /* Emit immediate values.
4567 t
->immediates
= (struct ureg_src
*)CALLOC(program
->num_immediates
* sizeof(struct ureg_src
));
4568 if (t
->immediates
== NULL
) {
4569 ret
= PIPE_ERROR_OUT_OF_MEMORY
;
4573 foreach_iter(exec_list_iterator
, iter
, program
->immediates
) {
4574 immediate_storage
*imm
= (immediate_storage
*)iter
.get();
4575 t
->immediates
[i
++] = emit_immediate(t
, imm
->values
, imm
->type
, imm
->size
);
4578 /* texture samplers */
4579 for (i
= 0; i
< ctx
->Const
.MaxTextureImageUnits
; i
++) {
4580 if (program
->samplers_used
& (1 << i
)) {
4581 t
->samplers
[i
] = ureg_DECL_sampler(ureg
, i
);
4585 /* Emit each instruction in turn:
4587 foreach_iter(exec_list_iterator
, iter
, program
->instructions
) {
4588 set_insn_start(t
, ureg_get_instruction_number(ureg
));
4589 compile_tgsi_instruction(t
, (glsl_to_tgsi_instruction
*)iter
.get());
4591 if (t
->prevInstWrotePointSize
&& proginfo
->Id
) {
4592 /* The previous instruction wrote to the (fake) vertex point size
4593 * result register. Now we need to clamp that value to the min/max
4594 * point size range, putting the result into the real point size
4596 * Note that we can't do this easily at the end of program due to
4597 * possible early return.
4599 set_insn_start(t
, ureg_get_instruction_number(ureg
));
4601 ureg_writemask(t
->outputs
[t
->pointSizeOutIndex
], WRITEMASK_X
),
4602 ureg_src(t
->outputs
[t
->pointSizeOutIndex
]),
4603 ureg_swizzle(t
->pointSizeConst
, 1,1,1,1));
4604 ureg_MIN(t
->ureg
, ureg_writemask(t
->pointSizeResult
, WRITEMASK_X
),
4605 ureg_src(t
->outputs
[t
->pointSizeOutIndex
]),
4606 ureg_swizzle(t
->pointSizeConst
, 2,2,2,2));
4608 t
->prevInstWrotePointSize
= GL_FALSE
;
4611 /* Fix up all emitted labels:
4613 for (i
= 0; i
< t
->labels_count
; i
++) {
4614 ureg_fixup_label(ureg
, t
->labels
[i
].token
,
4615 t
->insn
[t
->labels
[i
].branch_target
]);
4622 FREE(t
->immediates
);
4625 debug_printf("%s: translate error flag set\n", __FUNCTION__
);
4630 /* ----------------------------- End TGSI code ------------------------------ */
4633 * Convert a shader's GLSL IR into a Mesa gl_program, although without
4634 * generating Mesa IR.
4636 static struct gl_program
*
4637 get_mesa_program(struct gl_context
*ctx
,
4638 struct gl_shader_program
*shader_program
,
4639 struct gl_shader
*shader
)
4641 glsl_to_tgsi_visitor
* v
= new glsl_to_tgsi_visitor();
4642 struct gl_program
*prog
;
4644 const char *target_string
;
4646 struct gl_shader_compiler_options
*options
=
4647 &ctx
->ShaderCompilerOptions
[_mesa_shader_type_to_index(shader
->Type
)];
4649 switch (shader
->Type
) {
4650 case GL_VERTEX_SHADER
:
4651 target
= GL_VERTEX_PROGRAM_ARB
;
4652 target_string
= "vertex";
4654 case GL_FRAGMENT_SHADER
:
4655 target
= GL_FRAGMENT_PROGRAM_ARB
;
4656 target_string
= "fragment";
4658 case GL_GEOMETRY_SHADER
:
4659 target
= GL_GEOMETRY_PROGRAM_NV
;
4660 target_string
= "geometry";
4663 assert(!"should not be reached");
4667 validate_ir_tree(shader
->ir
);
4669 prog
= ctx
->Driver
.NewProgram(ctx
, target
, shader_program
->Name
);
4672 prog
->Parameters
= _mesa_new_parameter_list();
4673 prog
->Varying
= _mesa_new_parameter_list();
4674 prog
->Attributes
= _mesa_new_parameter_list();
4677 v
->shader_program
= shader_program
;
4678 v
->options
= options
;
4679 v
->glsl_version
= ctx
->Const
.GLSLVersion
;
4680 v
->native_integers
= ctx
->Const
.NativeIntegers
;
4682 add_uniforms_to_parameters_list(shader_program
, shader
, prog
);
4684 /* Emit intermediate IR for main(). */
4685 visit_exec_list(shader
->ir
, v
);
4687 /* Now emit bodies for any functions that were used. */
4689 progress
= GL_FALSE
;
4691 foreach_iter(exec_list_iterator
, iter
, v
->function_signatures
) {
4692 function_entry
*entry
= (function_entry
*)iter
.get();
4694 if (!entry
->bgn_inst
) {
4695 v
->current_function
= entry
;
4697 entry
->bgn_inst
= v
->emit(NULL
, TGSI_OPCODE_BGNSUB
);
4698 entry
->bgn_inst
->function
= entry
;
4700 visit_exec_list(&entry
->sig
->body
, v
);
4702 glsl_to_tgsi_instruction
*last
;
4703 last
= (glsl_to_tgsi_instruction
*)v
->instructions
.get_tail();
4704 if (last
->op
!= TGSI_OPCODE_RET
)
4705 v
->emit(NULL
, TGSI_OPCODE_RET
);
4707 glsl_to_tgsi_instruction
*end
;
4708 end
= v
->emit(NULL
, TGSI_OPCODE_ENDSUB
);
4709 end
->function
= entry
;
4717 /* Print out some information (for debugging purposes) used by the
4718 * optimization passes. */
4719 for (i
=0; i
< v
->next_temp
; i
++) {
4720 int fr
= v
->get_first_temp_read(i
);
4721 int fw
= v
->get_first_temp_write(i
);
4722 int lr
= v
->get_last_temp_read(i
);
4723 int lw
= v
->get_last_temp_write(i
);
4725 printf("Temp %d: FR=%3d FW=%3d LR=%3d LW=%3d\n", i
, fr
, fw
, lr
, lw
);
4730 /* Remove reads to output registers, and to varyings in vertex shaders. */
4731 v
->remove_output_reads(PROGRAM_OUTPUT
);
4732 if (target
== GL_VERTEX_PROGRAM_ARB
)
4733 v
->remove_output_reads(PROGRAM_VARYING
);
4735 /* Perform optimizations on the instructions in the glsl_to_tgsi_visitor. */
4737 v
->copy_propagate();
4738 while (v
->eliminate_dead_code_advanced());
4740 /* FIXME: These passes to optimize temporary registers don't work when there
4741 * is indirect addressing of the temporary register space. We need proper
4742 * array support so that we don't have to give up these passes in every
4743 * shader that uses arrays.
4745 if (!v
->indirect_addr_temps
) {
4746 v
->eliminate_dead_code();
4747 v
->merge_registers();
4748 v
->renumber_registers();
4751 /* Write the END instruction. */
4752 v
->emit(NULL
, TGSI_OPCODE_END
);
4754 if (ctx
->Shader
.Flags
& GLSL_DUMP
) {
4756 printf("GLSL IR for linked %s program %d:\n", target_string
,
4757 shader_program
->Name
);
4758 _mesa_print_ir(shader
->ir
, NULL
);
4763 prog
->Instructions
= NULL
;
4764 prog
->NumInstructions
= 0;
4766 do_set_program_inouts(shader
->ir
, prog
);
4767 count_resources(v
, prog
);
4769 check_resources(ctx
, shader_program
, v
, prog
);
4771 _mesa_reference_program(ctx
, &shader
->Program
, prog
);
4773 struct st_vertex_program
*stvp
;
4774 struct st_fragment_program
*stfp
;
4775 struct st_geometry_program
*stgp
;
4777 switch (shader
->Type
) {
4778 case GL_VERTEX_SHADER
:
4779 stvp
= (struct st_vertex_program
*)prog
;
4780 stvp
->glsl_to_tgsi
= v
;
4782 case GL_FRAGMENT_SHADER
:
4783 stfp
= (struct st_fragment_program
*)prog
;
4784 stfp
->glsl_to_tgsi
= v
;
4786 case GL_GEOMETRY_SHADER
:
4787 stgp
= (struct st_geometry_program
*)prog
;
4788 stgp
->glsl_to_tgsi
= v
;
4791 assert(!"should not be reached");
4801 st_new_shader(struct gl_context
*ctx
, GLuint name
, GLuint type
)
4803 struct gl_shader
*shader
;
4804 assert(type
== GL_FRAGMENT_SHADER
|| type
== GL_VERTEX_SHADER
||
4805 type
== GL_GEOMETRY_SHADER_ARB
);
4806 shader
= rzalloc(NULL
, struct gl_shader
);
4808 shader
->Type
= type
;
4809 shader
->Name
= name
;
4810 _mesa_init_shader(ctx
, shader
);
4815 struct gl_shader_program
*
4816 st_new_shader_program(struct gl_context
*ctx
, GLuint name
)
4818 struct gl_shader_program
*shProg
;
4819 shProg
= rzalloc(NULL
, struct gl_shader_program
);
4821 shProg
->Name
= name
;
4822 _mesa_init_shader_program(ctx
, shProg
);
4829 * Called via ctx->Driver.LinkShader()
4830 * This actually involves converting GLSL IR into an intermediate TGSI-like IR
4831 * with code lowering and other optimizations.
4834 st_link_shader(struct gl_context
*ctx
, struct gl_shader_program
*prog
)
4836 assert(prog
->LinkStatus
);
4838 for (unsigned i
= 0; i
< MESA_SHADER_TYPES
; i
++) {
4839 if (prog
->_LinkedShaders
[i
] == NULL
)
4843 exec_list
*ir
= prog
->_LinkedShaders
[i
]->ir
;
4844 const struct gl_shader_compiler_options
*options
=
4845 &ctx
->ShaderCompilerOptions
[_mesa_shader_type_to_index(prog
->_LinkedShaders
[i
]->Type
)];
4851 do_mat_op_to_vec(ir
);
4852 lower_instructions(ir
, (MOD_TO_FRACT
| DIV_TO_MUL_RCP
| EXP_TO_EXP2
4854 | ((options
->EmitNoPow
) ? POW_TO_EXP2
: 0)));
4856 progress
= do_lower_jumps(ir
, true, true, options
->EmitNoMainReturn
, options
->EmitNoCont
, options
->EmitNoLoops
) || progress
;
4858 progress
= do_common_optimization(ir
, true, options
->MaxUnrollIterations
) || progress
;
4860 progress
= lower_quadop_vector(ir
, false) || progress
;
4862 if (options
->EmitNoIfs
) {
4863 progress
= lower_discard(ir
) || progress
;
4864 progress
= lower_if_to_cond_assign(ir
) || progress
;
4867 if (options
->EmitNoNoise
)
4868 progress
= lower_noise(ir
) || progress
;
4870 /* If there are forms of indirect addressing that the driver
4871 * cannot handle, perform the lowering pass.
4873 if (options
->EmitNoIndirectInput
|| options
->EmitNoIndirectOutput
4874 || options
->EmitNoIndirectTemp
|| options
->EmitNoIndirectUniform
)
4876 lower_variable_index_to_cond_assign(ir
,
4877 options
->EmitNoIndirectInput
,
4878 options
->EmitNoIndirectOutput
,
4879 options
->EmitNoIndirectTemp
,
4880 options
->EmitNoIndirectUniform
)
4883 progress
= do_vec_index_to_cond_assign(ir
) || progress
;
4886 validate_ir_tree(ir
);
4889 for (unsigned i
= 0; i
< MESA_SHADER_TYPES
; i
++) {
4890 struct gl_program
*linked_prog
;
4892 if (prog
->_LinkedShaders
[i
] == NULL
)
4895 linked_prog
= get_mesa_program(ctx
, prog
, prog
->_LinkedShaders
[i
]);
4900 switch (prog
->_LinkedShaders
[i
]->Type
) {
4901 case GL_VERTEX_SHADER
:
4902 _mesa_reference_vertprog(ctx
, &prog
->VertexProgram
,
4903 (struct gl_vertex_program
*)linked_prog
);
4904 ok
= ctx
->Driver
.ProgramStringNotify(ctx
, GL_VERTEX_PROGRAM_ARB
,
4907 case GL_FRAGMENT_SHADER
:
4908 _mesa_reference_fragprog(ctx
, &prog
->FragmentProgram
,
4909 (struct gl_fragment_program
*)linked_prog
);
4910 ok
= ctx
->Driver
.ProgramStringNotify(ctx
, GL_FRAGMENT_PROGRAM_ARB
,
4913 case GL_GEOMETRY_SHADER
:
4914 _mesa_reference_geomprog(ctx
, &prog
->GeometryProgram
,
4915 (struct gl_geometry_program
*)linked_prog
);
4916 ok
= ctx
->Driver
.ProgramStringNotify(ctx
, GL_GEOMETRY_PROGRAM_NV
,
4925 _mesa_reference_program(ctx
, &linked_prog
, NULL
);
4933 * Link a GLSL shader program. Called via glLinkProgram().
4936 st_glsl_link_shader(struct gl_context
*ctx
, struct gl_shader_program
*prog
)
4940 _mesa_clear_shader_program_data(ctx
, prog
);
4942 prog
->LinkStatus
= GL_TRUE
;
4944 for (i
= 0; i
< prog
->NumShaders
; i
++) {
4945 if (!prog
->Shaders
[i
]->CompileStatus
) {
4946 fail_link(prog
, "linking with uncompiled shader");
4947 prog
->LinkStatus
= GL_FALSE
;
4951 prog
->Varying
= _mesa_new_parameter_list();
4952 _mesa_reference_vertprog(ctx
, &prog
->VertexProgram
, NULL
);
4953 _mesa_reference_fragprog(ctx
, &prog
->FragmentProgram
, NULL
);
4954 _mesa_reference_geomprog(ctx
, &prog
->GeometryProgram
, NULL
);
4956 if (prog
->LinkStatus
) {
4957 link_shaders(ctx
, prog
);
4960 if (prog
->LinkStatus
) {
4961 if (!ctx
->Driver
.LinkShader(ctx
, prog
)) {
4962 prog
->LinkStatus
= GL_FALSE
;
4966 set_uniform_initializers(ctx
, prog
);
4968 if (ctx
->Shader
.Flags
& GLSL_DUMP
) {
4969 if (!prog
->LinkStatus
) {
4970 printf("GLSL shader program %d failed to link\n", prog
->Name
);
4973 if (prog
->InfoLog
&& prog
->InfoLog
[0] != 0) {
4974 printf("GLSL shader program %d info log:\n", prog
->Name
);
4975 printf("%s\n", prog
->InfoLog
);