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
;
299 variable_storage
*find_variable_storage(ir_variable
*var
);
301 int add_constant(gl_register_file file
, gl_constant_value values
[4],
302 int size
, int datatype
, GLuint
*swizzle_out
);
304 function_entry
*get_function_signature(ir_function_signature
*sig
);
306 st_src_reg
get_temp(const glsl_type
*type
);
307 void reladdr_to_temp(ir_instruction
*ir
, st_src_reg
*reg
, int *num_reladdr
);
309 st_src_reg
st_src_reg_for_float(float val
);
310 st_src_reg
st_src_reg_for_int(int val
);
311 st_src_reg
st_src_reg_for_type(int type
, int val
);
314 * \name Visit methods
316 * As typical for the visitor pattern, there must be one \c visit method for
317 * each concrete subclass of \c ir_instruction. Virtual base classes within
318 * the hierarchy should not have \c visit methods.
321 virtual void visit(ir_variable
*);
322 virtual void visit(ir_loop
*);
323 virtual void visit(ir_loop_jump
*);
324 virtual void visit(ir_function_signature
*);
325 virtual void visit(ir_function
*);
326 virtual void visit(ir_expression
*);
327 virtual void visit(ir_swizzle
*);
328 virtual void visit(ir_dereference_variable
*);
329 virtual void visit(ir_dereference_array
*);
330 virtual void visit(ir_dereference_record
*);
331 virtual void visit(ir_assignment
*);
332 virtual void visit(ir_constant
*);
333 virtual void visit(ir_call
*);
334 virtual void visit(ir_return
*);
335 virtual void visit(ir_discard
*);
336 virtual void visit(ir_texture
*);
337 virtual void visit(ir_if
*);
342 /** List of variable_storage */
345 /** List of immediate_storage */
346 exec_list immediates
;
349 /** List of function_entry */
350 exec_list function_signatures
;
351 int next_signature_id
;
353 /** List of glsl_to_tgsi_instruction */
354 exec_list instructions
;
356 glsl_to_tgsi_instruction
*emit(ir_instruction
*ir
, unsigned op
);
358 glsl_to_tgsi_instruction
*emit(ir_instruction
*ir
, unsigned op
,
359 st_dst_reg dst
, st_src_reg src0
);
361 glsl_to_tgsi_instruction
*emit(ir_instruction
*ir
, unsigned op
,
362 st_dst_reg dst
, st_src_reg src0
, st_src_reg src1
);
364 glsl_to_tgsi_instruction
*emit(ir_instruction
*ir
, unsigned op
,
366 st_src_reg src0
, st_src_reg src1
, st_src_reg src2
);
368 unsigned get_opcode(ir_instruction
*ir
, unsigned op
,
370 st_src_reg src0
, st_src_reg src1
);
373 * Emit the correct dot-product instruction for the type of arguments
375 void emit_dp(ir_instruction
*ir
,
381 void emit_scalar(ir_instruction
*ir
, unsigned op
,
382 st_dst_reg dst
, st_src_reg src0
);
384 void emit_scalar(ir_instruction
*ir
, unsigned op
,
385 st_dst_reg dst
, st_src_reg src0
, st_src_reg src1
);
387 void emit_arl(ir_instruction
*ir
, st_dst_reg dst
, st_src_reg src0
);
389 void emit_scs(ir_instruction
*ir
, unsigned op
,
390 st_dst_reg dst
, const st_src_reg
&src
);
392 GLboolean
try_emit_mad(ir_expression
*ir
,
394 GLboolean
try_emit_sat(ir_expression
*ir
);
396 void emit_swz(ir_expression
*ir
);
398 bool process_move_condition(ir_rvalue
*ir
);
400 void remove_output_reads(gl_register_file type
);
401 void simplify_cmp(void);
403 void rename_temp_register(int index
, int new_index
);
404 int get_first_temp_read(int index
);
405 int get_first_temp_write(int index
);
406 int get_last_temp_read(int index
);
407 int get_last_temp_write(int index
);
409 void copy_propagate(void);
410 void eliminate_dead_code(void);
411 int eliminate_dead_code_advanced(void);
412 void merge_registers(void);
413 void renumber_registers(void);
418 static st_src_reg undef_src
= st_src_reg(PROGRAM_UNDEFINED
, 0, GLSL_TYPE_ERROR
);
420 static st_dst_reg undef_dst
= st_dst_reg(PROGRAM_UNDEFINED
, SWIZZLE_NOOP
, GLSL_TYPE_ERROR
);
422 static st_dst_reg address_reg
= st_dst_reg(PROGRAM_ADDRESS
, WRITEMASK_X
, GLSL_TYPE_FLOAT
);
425 fail_link(struct gl_shader_program
*prog
, const char *fmt
, ...) PRINTFLIKE(2, 3);
428 fail_link(struct gl_shader_program
*prog
, const char *fmt
, ...)
432 ralloc_vasprintf_append(&prog
->InfoLog
, fmt
, args
);
435 prog
->LinkStatus
= GL_FALSE
;
439 swizzle_for_size(int size
)
441 int size_swizzles
[4] = {
442 MAKE_SWIZZLE4(SWIZZLE_X
, SWIZZLE_X
, SWIZZLE_X
, SWIZZLE_X
),
443 MAKE_SWIZZLE4(SWIZZLE_X
, SWIZZLE_Y
, SWIZZLE_Y
, SWIZZLE_Y
),
444 MAKE_SWIZZLE4(SWIZZLE_X
, SWIZZLE_Y
, SWIZZLE_Z
, SWIZZLE_Z
),
445 MAKE_SWIZZLE4(SWIZZLE_X
, SWIZZLE_Y
, SWIZZLE_Z
, SWIZZLE_W
),
448 assert((size
>= 1) && (size
<= 4));
449 return size_swizzles
[size
- 1];
453 is_tex_instruction(unsigned opcode
)
455 const tgsi_opcode_info
* info
= tgsi_get_opcode_info(opcode
);
460 num_inst_dst_regs(unsigned opcode
)
462 const tgsi_opcode_info
* info
= tgsi_get_opcode_info(opcode
);
463 return info
->num_dst
;
467 num_inst_src_regs(unsigned opcode
)
469 const tgsi_opcode_info
* info
= tgsi_get_opcode_info(opcode
);
470 return info
->is_tex
? info
->num_src
- 1 : info
->num_src
;
473 glsl_to_tgsi_instruction
*
474 glsl_to_tgsi_visitor::emit(ir_instruction
*ir
, unsigned op
,
476 st_src_reg src0
, st_src_reg src1
, st_src_reg src2
)
478 glsl_to_tgsi_instruction
*inst
= new(mem_ctx
) glsl_to_tgsi_instruction();
479 int num_reladdr
= 0, i
;
481 op
= get_opcode(ir
, op
, dst
, src0
, src1
);
483 /* If we have to do relative addressing, we want to load the ARL
484 * reg directly for one of the regs, and preload the other reladdr
485 * sources into temps.
487 num_reladdr
+= dst
.reladdr
!= NULL
;
488 num_reladdr
+= src0
.reladdr
!= NULL
;
489 num_reladdr
+= src1
.reladdr
!= NULL
;
490 num_reladdr
+= src2
.reladdr
!= NULL
;
492 reladdr_to_temp(ir
, &src2
, &num_reladdr
);
493 reladdr_to_temp(ir
, &src1
, &num_reladdr
);
494 reladdr_to_temp(ir
, &src0
, &num_reladdr
);
497 emit_arl(ir
, address_reg
, *dst
.reladdr
);
500 assert(num_reladdr
== 0);
510 inst
->function
= NULL
;
512 if (op
== TGSI_OPCODE_ARL
)
513 this->num_address_regs
= 1;
515 /* Update indirect addressing status used by TGSI */
518 case PROGRAM_TEMPORARY
:
519 this->indirect_addr_temps
= true;
521 case PROGRAM_LOCAL_PARAM
:
522 case PROGRAM_ENV_PARAM
:
523 case PROGRAM_STATE_VAR
:
524 case PROGRAM_NAMED_PARAM
:
525 case PROGRAM_CONSTANT
:
526 case PROGRAM_UNIFORM
:
527 this->indirect_addr_consts
= true;
529 case PROGRAM_IMMEDIATE
:
530 assert(!"immediates should not have indirect addressing");
537 for (i
=0; i
<3; i
++) {
538 if(inst
->src
[i
].reladdr
) {
539 switch(inst
->src
[i
].file
) {
540 case PROGRAM_TEMPORARY
:
541 this->indirect_addr_temps
= true;
543 case PROGRAM_LOCAL_PARAM
:
544 case PROGRAM_ENV_PARAM
:
545 case PROGRAM_STATE_VAR
:
546 case PROGRAM_NAMED_PARAM
:
547 case PROGRAM_CONSTANT
:
548 case PROGRAM_UNIFORM
:
549 this->indirect_addr_consts
= true;
551 case PROGRAM_IMMEDIATE
:
552 assert(!"immediates should not have indirect addressing");
561 this->instructions
.push_tail(inst
);
567 glsl_to_tgsi_instruction
*
568 glsl_to_tgsi_visitor::emit(ir_instruction
*ir
, unsigned op
,
569 st_dst_reg dst
, st_src_reg src0
, st_src_reg src1
)
571 return emit(ir
, op
, dst
, src0
, src1
, undef_src
);
574 glsl_to_tgsi_instruction
*
575 glsl_to_tgsi_visitor::emit(ir_instruction
*ir
, unsigned op
,
576 st_dst_reg dst
, st_src_reg src0
)
578 assert(dst
.writemask
!= 0);
579 return emit(ir
, op
, dst
, src0
, undef_src
, undef_src
);
582 glsl_to_tgsi_instruction
*
583 glsl_to_tgsi_visitor::emit(ir_instruction
*ir
, unsigned op
)
585 return emit(ir
, op
, undef_dst
, undef_src
, undef_src
, undef_src
);
589 * Determines whether to use an integer, unsigned integer, or float opcode
590 * based on the operands and input opcode, then emits the result.
592 * TODO: type checking for remaining TGSI opcodes
595 glsl_to_tgsi_visitor::get_opcode(ir_instruction
*ir
, unsigned op
,
597 st_src_reg src0
, st_src_reg src1
)
599 int type
= GLSL_TYPE_FLOAT
;
601 if (src0
.type
== GLSL_TYPE_FLOAT
|| src1
.type
== GLSL_TYPE_FLOAT
)
602 type
= GLSL_TYPE_FLOAT
;
603 else if (glsl_version
>= 130)
606 #define case4(c, f, i, u) \
607 case TGSI_OPCODE_##c: \
608 if (type == GLSL_TYPE_INT) op = TGSI_OPCODE_##i; \
609 else if (type == GLSL_TYPE_UINT) op = TGSI_OPCODE_##u; \
610 else op = TGSI_OPCODE_##f; \
612 #define case3(f, i, u) case4(f, f, i, u)
613 #define case2fi(f, i) case4(f, f, i, i)
614 #define case2iu(i, u) case4(i, LAST, i, u)
620 case3(DIV
, IDIV
, UDIV
);
621 case3(MAX
, IMAX
, UMAX
);
622 case3(MIN
, IMIN
, UMIN
);
627 case3(SGE
, ISGE
, USGE
);
628 case3(SLT
, ISLT
, USLT
);
640 assert(op
!= TGSI_OPCODE_LAST
);
645 glsl_to_tgsi_visitor::emit_dp(ir_instruction
*ir
,
646 st_dst_reg dst
, st_src_reg src0
, st_src_reg src1
,
649 static const unsigned dot_opcodes
[] = {
650 TGSI_OPCODE_DP2
, TGSI_OPCODE_DP3
, TGSI_OPCODE_DP4
653 emit(ir
, dot_opcodes
[elements
- 2], dst
, src0
, src1
);
657 * Emits TGSI scalar opcodes to produce unique answers across channels.
659 * Some TGSI opcodes are scalar-only, like ARB_fp/vp. The src X
660 * channel determines the result across all channels. So to do a vec4
661 * of this operation, we want to emit a scalar per source channel used
662 * to produce dest channels.
665 glsl_to_tgsi_visitor::emit_scalar(ir_instruction
*ir
, unsigned op
,
667 st_src_reg orig_src0
, st_src_reg orig_src1
)
670 int done_mask
= ~dst
.writemask
;
672 /* TGSI RCP is a scalar operation splatting results to all channels,
673 * like ARB_fp/vp. So emit as many RCPs as necessary to cover our
676 for (i
= 0; i
< 4; i
++) {
677 GLuint this_mask
= (1 << i
);
678 glsl_to_tgsi_instruction
*inst
;
679 st_src_reg src0
= orig_src0
;
680 st_src_reg src1
= orig_src1
;
682 if (done_mask
& this_mask
)
685 GLuint src0_swiz
= GET_SWZ(src0
.swizzle
, i
);
686 GLuint src1_swiz
= GET_SWZ(src1
.swizzle
, i
);
687 for (j
= i
+ 1; j
< 4; j
++) {
688 /* If there is another enabled component in the destination that is
689 * derived from the same inputs, generate its value on this pass as
692 if (!(done_mask
& (1 << j
)) &&
693 GET_SWZ(src0
.swizzle
, j
) == src0_swiz
&&
694 GET_SWZ(src1
.swizzle
, j
) == src1_swiz
) {
695 this_mask
|= (1 << j
);
698 src0
.swizzle
= MAKE_SWIZZLE4(src0_swiz
, src0_swiz
,
699 src0_swiz
, src0_swiz
);
700 src1
.swizzle
= MAKE_SWIZZLE4(src1_swiz
, src1_swiz
,
701 src1_swiz
, src1_swiz
);
703 inst
= emit(ir
, op
, dst
, src0
, src1
);
704 inst
->dst
.writemask
= this_mask
;
705 done_mask
|= this_mask
;
710 glsl_to_tgsi_visitor::emit_scalar(ir_instruction
*ir
, unsigned op
,
711 st_dst_reg dst
, st_src_reg src0
)
713 st_src_reg undef
= undef_src
;
715 undef
.swizzle
= SWIZZLE_XXXX
;
717 emit_scalar(ir
, op
, dst
, src0
, undef
);
721 glsl_to_tgsi_visitor::emit_arl(ir_instruction
*ir
,
722 st_dst_reg dst
, st_src_reg src0
)
724 st_src_reg tmp
= get_temp(glsl_type::float_type
);
726 if (src0
.type
== GLSL_TYPE_INT
)
727 emit(NULL
, TGSI_OPCODE_I2F
, st_dst_reg(tmp
), src0
);
728 else if (src0
.type
== GLSL_TYPE_UINT
)
729 emit(NULL
, TGSI_OPCODE_U2F
, st_dst_reg(tmp
), src0
);
733 emit(NULL
, TGSI_OPCODE_ARL
, dst
, tmp
);
737 * Emit an TGSI_OPCODE_SCS instruction
739 * The \c SCS opcode functions a bit differently than the other TGSI opcodes.
740 * Instead of splatting its result across all four components of the
741 * destination, it writes one value to the \c x component and another value to
742 * the \c y component.
744 * \param ir IR instruction being processed
745 * \param op Either \c TGSI_OPCODE_SIN or \c TGSI_OPCODE_COS depending
746 * on which value is desired.
747 * \param dst Destination register
748 * \param src Source register
751 glsl_to_tgsi_visitor::emit_scs(ir_instruction
*ir
, unsigned op
,
753 const st_src_reg
&src
)
755 /* Vertex programs cannot use the SCS opcode.
757 if (this->prog
->Target
== GL_VERTEX_PROGRAM_ARB
) {
758 emit_scalar(ir
, op
, dst
, src
);
762 const unsigned component
= (op
== TGSI_OPCODE_SIN
) ? 0 : 1;
763 const unsigned scs_mask
= (1U << component
);
764 int done_mask
= ~dst
.writemask
;
767 assert(op
== TGSI_OPCODE_SIN
|| op
== TGSI_OPCODE_COS
);
769 /* If there are compnents in the destination that differ from the component
770 * that will be written by the SCS instrution, we'll need a temporary.
772 if (scs_mask
!= unsigned(dst
.writemask
)) {
773 tmp
= get_temp(glsl_type::vec4_type
);
776 for (unsigned i
= 0; i
< 4; i
++) {
777 unsigned this_mask
= (1U << i
);
778 st_src_reg src0
= src
;
780 if ((done_mask
& this_mask
) != 0)
783 /* The source swizzle specified which component of the source generates
784 * sine / cosine for the current component in the destination. The SCS
785 * instruction requires that this value be swizzle to the X component.
786 * Replace the current swizzle with a swizzle that puts the source in
789 unsigned src0_swiz
= GET_SWZ(src
.swizzle
, i
);
791 src0
.swizzle
= MAKE_SWIZZLE4(src0_swiz
, src0_swiz
,
792 src0_swiz
, src0_swiz
);
793 for (unsigned j
= i
+ 1; j
< 4; j
++) {
794 /* If there is another enabled component in the destination that is
795 * derived from the same inputs, generate its value on this pass as
798 if (!(done_mask
& (1 << j
)) &&
799 GET_SWZ(src0
.swizzle
, j
) == src0_swiz
) {
800 this_mask
|= (1 << j
);
804 if (this_mask
!= scs_mask
) {
805 glsl_to_tgsi_instruction
*inst
;
806 st_dst_reg tmp_dst
= st_dst_reg(tmp
);
808 /* Emit the SCS instruction.
810 inst
= emit(ir
, TGSI_OPCODE_SCS
, tmp_dst
, src0
);
811 inst
->dst
.writemask
= scs_mask
;
813 /* Move the result of the SCS instruction to the desired location in
816 tmp
.swizzle
= MAKE_SWIZZLE4(component
, component
,
817 component
, component
);
818 inst
= emit(ir
, TGSI_OPCODE_SCS
, dst
, tmp
);
819 inst
->dst
.writemask
= this_mask
;
821 /* Emit the SCS instruction to write directly to the destination.
823 glsl_to_tgsi_instruction
*inst
= emit(ir
, TGSI_OPCODE_SCS
, dst
, src0
);
824 inst
->dst
.writemask
= scs_mask
;
827 done_mask
|= this_mask
;
832 glsl_to_tgsi_visitor::add_constant(gl_register_file file
,
833 gl_constant_value values
[4], int size
, int datatype
,
836 if (file
== PROGRAM_CONSTANT
) {
837 return _mesa_add_typed_unnamed_constant(this->prog
->Parameters
, values
,
838 size
, datatype
, swizzle_out
);
841 immediate_storage
*entry
;
842 assert(file
== PROGRAM_IMMEDIATE
);
844 /* Search immediate storage to see if we already have an identical
845 * immediate that we can use instead of adding a duplicate entry.
847 foreach_iter(exec_list_iterator
, iter
, this->immediates
) {
848 entry
= (immediate_storage
*)iter
.get();
850 if (entry
->size
== size
&&
851 entry
->type
== datatype
&&
852 !memcmp(entry
->values
, values
, size
* sizeof(gl_constant_value
))) {
858 /* Add this immediate to the list. */
859 entry
= new(mem_ctx
) immediate_storage(values
, size
, datatype
);
860 this->immediates
.push_tail(entry
);
861 this->num_immediates
++;
867 glsl_to_tgsi_visitor::st_src_reg_for_float(float val
)
869 st_src_reg
src(PROGRAM_IMMEDIATE
, -1, GLSL_TYPE_FLOAT
);
870 union gl_constant_value uval
;
873 src
.index
= add_constant(src
.file
, &uval
, 1, GL_FLOAT
, &src
.swizzle
);
879 glsl_to_tgsi_visitor::st_src_reg_for_int(int val
)
881 st_src_reg
src(PROGRAM_IMMEDIATE
, -1, GLSL_TYPE_INT
);
882 union gl_constant_value uval
;
884 assert(glsl_version
>= 130);
887 src
.index
= add_constant(src
.file
, &uval
, 1, GL_INT
, &src
.swizzle
);
893 glsl_to_tgsi_visitor::st_src_reg_for_type(int type
, int val
)
895 if (glsl_version
>= 130)
896 return type
== GLSL_TYPE_FLOAT
? st_src_reg_for_float(val
) :
897 st_src_reg_for_int(val
);
899 return st_src_reg_for_float(val
);
903 type_size(const struct glsl_type
*type
)
908 switch (type
->base_type
) {
911 case GLSL_TYPE_FLOAT
:
913 if (type
->is_matrix()) {
914 return type
->matrix_columns
;
916 /* Regardless of size of vector, it gets a vec4. This is bad
917 * packing for things like floats, but otherwise arrays become a
918 * mess. Hopefully a later pass over the code can pack scalars
919 * down if appropriate.
923 case GLSL_TYPE_ARRAY
:
924 assert(type
->length
> 0);
925 return type_size(type
->fields
.array
) * type
->length
;
926 case GLSL_TYPE_STRUCT
:
928 for (i
= 0; i
< type
->length
; i
++) {
929 size
+= type_size(type
->fields
.structure
[i
].type
);
932 case GLSL_TYPE_SAMPLER
:
933 /* Samplers take up one slot in UNIFORMS[], but they're baked in
944 * In the initial pass of codegen, we assign temporary numbers to
945 * intermediate results. (not SSA -- variable assignments will reuse
949 glsl_to_tgsi_visitor::get_temp(const glsl_type
*type
)
953 src
.type
= glsl_version
>= 130 ? type
->base_type
: GLSL_TYPE_FLOAT
;
954 src
.file
= PROGRAM_TEMPORARY
;
955 src
.index
= next_temp
;
957 next_temp
+= type_size(type
);
959 if (type
->is_array() || type
->is_record()) {
960 src
.swizzle
= SWIZZLE_NOOP
;
962 src
.swizzle
= swizzle_for_size(type
->vector_elements
);
970 glsl_to_tgsi_visitor::find_variable_storage(ir_variable
*var
)
973 variable_storage
*entry
;
975 foreach_iter(exec_list_iterator
, iter
, this->variables
) {
976 entry
= (variable_storage
*)iter
.get();
978 if (entry
->var
== var
)
986 glsl_to_tgsi_visitor::visit(ir_variable
*ir
)
988 if (strcmp(ir
->name
, "gl_FragCoord") == 0) {
989 struct gl_fragment_program
*fp
= (struct gl_fragment_program
*)this->prog
;
991 fp
->OriginUpperLeft
= ir
->origin_upper_left
;
992 fp
->PixelCenterInteger
= ir
->pixel_center_integer
;
994 } else if (strcmp(ir
->name
, "gl_FragDepth") == 0) {
995 struct gl_fragment_program
*fp
= (struct gl_fragment_program
*)this->prog
;
996 switch (ir
->depth_layout
) {
997 case ir_depth_layout_none
:
998 fp
->FragDepthLayout
= FRAG_DEPTH_LAYOUT_NONE
;
1000 case ir_depth_layout_any
:
1001 fp
->FragDepthLayout
= FRAG_DEPTH_LAYOUT_ANY
;
1003 case ir_depth_layout_greater
:
1004 fp
->FragDepthLayout
= FRAG_DEPTH_LAYOUT_GREATER
;
1006 case ir_depth_layout_less
:
1007 fp
->FragDepthLayout
= FRAG_DEPTH_LAYOUT_LESS
;
1009 case ir_depth_layout_unchanged
:
1010 fp
->FragDepthLayout
= FRAG_DEPTH_LAYOUT_UNCHANGED
;
1018 if (ir
->mode
== ir_var_uniform
&& strncmp(ir
->name
, "gl_", 3) == 0) {
1020 const ir_state_slot
*const slots
= ir
->state_slots
;
1021 assert(ir
->state_slots
!= NULL
);
1023 /* Check if this statevar's setup in the STATE file exactly
1024 * matches how we'll want to reference it as a
1025 * struct/array/whatever. If not, then we need to move it into
1026 * temporary storage and hope that it'll get copy-propagated
1029 for (i
= 0; i
< ir
->num_state_slots
; i
++) {
1030 if (slots
[i
].swizzle
!= SWIZZLE_XYZW
) {
1035 struct variable_storage
*storage
;
1037 if (i
== ir
->num_state_slots
) {
1038 /* We'll set the index later. */
1039 storage
= new(mem_ctx
) variable_storage(ir
, PROGRAM_STATE_VAR
, -1);
1040 this->variables
.push_tail(storage
);
1044 /* The variable_storage constructor allocates slots based on the size
1045 * of the type. However, this had better match the number of state
1046 * elements that we're going to copy into the new temporary.
1048 assert((int) ir
->num_state_slots
== type_size(ir
->type
));
1050 storage
= new(mem_ctx
) variable_storage(ir
, PROGRAM_TEMPORARY
,
1052 this->variables
.push_tail(storage
);
1053 this->next_temp
+= type_size(ir
->type
);
1055 dst
= st_dst_reg(st_src_reg(PROGRAM_TEMPORARY
, storage
->index
,
1056 glsl_version
>= 130 ? ir
->type
->base_type
: GLSL_TYPE_FLOAT
));
1060 for (unsigned int i
= 0; i
< ir
->num_state_slots
; i
++) {
1061 int index
= _mesa_add_state_reference(this->prog
->Parameters
,
1062 (gl_state_index
*)slots
[i
].tokens
);
1064 if (storage
->file
== PROGRAM_STATE_VAR
) {
1065 if (storage
->index
== -1) {
1066 storage
->index
= index
;
1068 assert(index
== storage
->index
+ (int)i
);
1071 st_src_reg
src(PROGRAM_STATE_VAR
, index
,
1072 glsl_version
>= 130 ? ir
->type
->base_type
: GLSL_TYPE_FLOAT
);
1073 src
.swizzle
= slots
[i
].swizzle
;
1074 emit(ir
, TGSI_OPCODE_MOV
, dst
, src
);
1075 /* even a float takes up a whole vec4 reg in a struct/array. */
1080 if (storage
->file
== PROGRAM_TEMPORARY
&&
1081 dst
.index
!= storage
->index
+ (int) ir
->num_state_slots
) {
1082 fail_link(this->shader_program
,
1083 "failed to load builtin uniform `%s' (%d/%d regs loaded)\n",
1084 ir
->name
, dst
.index
- storage
->index
,
1085 type_size(ir
->type
));
1091 glsl_to_tgsi_visitor::visit(ir_loop
*ir
)
1093 ir_dereference_variable
*counter
= NULL
;
1095 if (ir
->counter
!= NULL
)
1096 counter
= new(ir
) ir_dereference_variable(ir
->counter
);
1098 if (ir
->from
!= NULL
) {
1099 assert(ir
->counter
!= NULL
);
1101 ir_assignment
*a
= new(ir
) ir_assignment(counter
, ir
->from
, NULL
);
1107 emit(NULL
, TGSI_OPCODE_BGNLOOP
);
1111 new(ir
) ir_expression(ir
->cmp
, glsl_type::bool_type
,
1113 ir_if
*if_stmt
= new(ir
) ir_if(e
);
1115 ir_loop_jump
*brk
= new(ir
) ir_loop_jump(ir_loop_jump::jump_break
);
1117 if_stmt
->then_instructions
.push_tail(brk
);
1119 if_stmt
->accept(this);
1126 visit_exec_list(&ir
->body_instructions
, this);
1128 if (ir
->increment
) {
1130 new(ir
) ir_expression(ir_binop_add
, counter
->type
,
1131 counter
, ir
->increment
);
1133 ir_assignment
*a
= new(ir
) ir_assignment(counter
, e
, NULL
);
1140 emit(NULL
, TGSI_OPCODE_ENDLOOP
);
1144 glsl_to_tgsi_visitor::visit(ir_loop_jump
*ir
)
1147 case ir_loop_jump::jump_break
:
1148 emit(NULL
, TGSI_OPCODE_BRK
);
1150 case ir_loop_jump::jump_continue
:
1151 emit(NULL
, TGSI_OPCODE_CONT
);
1158 glsl_to_tgsi_visitor::visit(ir_function_signature
*ir
)
1165 glsl_to_tgsi_visitor::visit(ir_function
*ir
)
1167 /* Ignore function bodies other than main() -- we shouldn't see calls to
1168 * them since they should all be inlined before we get to glsl_to_tgsi.
1170 if (strcmp(ir
->name
, "main") == 0) {
1171 const ir_function_signature
*sig
;
1174 sig
= ir
->matching_signature(&empty
);
1178 foreach_iter(exec_list_iterator
, iter
, sig
->body
) {
1179 ir_instruction
*ir
= (ir_instruction
*)iter
.get();
1187 glsl_to_tgsi_visitor::try_emit_mad(ir_expression
*ir
, int mul_operand
)
1189 int nonmul_operand
= 1 - mul_operand
;
1191 st_dst_reg result_dst
;
1193 ir_expression
*expr
= ir
->operands
[mul_operand
]->as_expression();
1194 if (!expr
|| expr
->operation
!= ir_binop_mul
)
1197 expr
->operands
[0]->accept(this);
1199 expr
->operands
[1]->accept(this);
1201 ir
->operands
[nonmul_operand
]->accept(this);
1204 this->result
= get_temp(ir
->type
);
1205 result_dst
= st_dst_reg(this->result
);
1206 result_dst
.writemask
= (1 << ir
->type
->vector_elements
) - 1;
1207 emit(ir
, TGSI_OPCODE_MAD
, result_dst
, a
, b
, c
);
1213 glsl_to_tgsi_visitor::try_emit_sat(ir_expression
*ir
)
1215 /* Saturates were only introduced to vertex programs in
1216 * NV_vertex_program3, so don't give them to drivers in the VP.
1218 if (this->prog
->Target
== GL_VERTEX_PROGRAM_ARB
)
1221 ir_rvalue
*sat_src
= ir
->as_rvalue_to_saturate();
1225 sat_src
->accept(this);
1226 st_src_reg src
= this->result
;
1228 /* If we generated an expression instruction into a temporary in
1229 * processing the saturate's operand, apply the saturate to that
1230 * instruction. Otherwise, generate a MOV to do the saturate.
1232 * Note that we have to be careful to only do this optimization if
1233 * the instruction in question was what generated src->result. For
1234 * example, ir_dereference_array might generate a MUL instruction
1235 * to create the reladdr, and return us a src reg using that
1236 * reladdr. That MUL result is not the value we're trying to
1239 ir_expression
*sat_src_expr
= sat_src
->as_expression();
1240 if (sat_src_expr
&& (sat_src_expr
->operation
== ir_binop_mul
||
1241 sat_src_expr
->operation
== ir_binop_add
||
1242 sat_src_expr
->operation
== ir_binop_dot
)) {
1243 glsl_to_tgsi_instruction
*new_inst
;
1244 new_inst
= (glsl_to_tgsi_instruction
*)this->instructions
.get_tail();
1245 new_inst
->saturate
= true;
1247 this->result
= get_temp(ir
->type
);
1248 st_dst_reg result_dst
= st_dst_reg(this->result
);
1249 result_dst
.writemask
= (1 << ir
->type
->vector_elements
) - 1;
1250 glsl_to_tgsi_instruction
*inst
;
1251 inst
= emit(ir
, TGSI_OPCODE_MOV
, result_dst
, src
);
1252 inst
->saturate
= true;
1259 glsl_to_tgsi_visitor::reladdr_to_temp(ir_instruction
*ir
,
1260 st_src_reg
*reg
, int *num_reladdr
)
1265 emit_arl(ir
, address_reg
, *reg
->reladdr
);
1267 if (*num_reladdr
!= 1) {
1268 st_src_reg temp
= get_temp(glsl_type::vec4_type
);
1270 emit(ir
, TGSI_OPCODE_MOV
, st_dst_reg(temp
), *reg
);
1278 glsl_to_tgsi_visitor::visit(ir_expression
*ir
)
1280 unsigned int operand
;
1281 st_src_reg op
[Elements(ir
->operands
)];
1282 st_src_reg result_src
;
1283 st_dst_reg result_dst
;
1285 /* Quick peephole: Emit MAD(a, b, c) instead of ADD(MUL(a, b), c)
1287 if (ir
->operation
== ir_binop_add
) {
1288 if (try_emit_mad(ir
, 1))
1290 if (try_emit_mad(ir
, 0))
1293 if (try_emit_sat(ir
))
1296 if (ir
->operation
== ir_quadop_vector
)
1297 assert(!"ir_quadop_vector should have been lowered");
1299 for (operand
= 0; operand
< ir
->get_num_operands(); operand
++) {
1300 this->result
.file
= PROGRAM_UNDEFINED
;
1301 ir
->operands
[operand
]->accept(this);
1302 if (this->result
.file
== PROGRAM_UNDEFINED
) {
1304 printf("Failed to get tree for expression operand:\n");
1305 ir
->operands
[operand
]->accept(&v
);
1308 op
[operand
] = this->result
;
1310 /* Matrix expression operands should have been broken down to vector
1311 * operations already.
1313 assert(!ir
->operands
[operand
]->type
->is_matrix());
1316 int vector_elements
= ir
->operands
[0]->type
->vector_elements
;
1317 if (ir
->operands
[1]) {
1318 vector_elements
= MAX2(vector_elements
,
1319 ir
->operands
[1]->type
->vector_elements
);
1322 this->result
.file
= PROGRAM_UNDEFINED
;
1324 /* Storage for our result. Ideally for an assignment we'd be using
1325 * the actual storage for the result here, instead.
1327 result_src
= get_temp(ir
->type
);
1328 /* convenience for the emit functions below. */
1329 result_dst
= st_dst_reg(result_src
);
1330 /* Limit writes to the channels that will be used by result_src later.
1331 * This does limit this temp's use as a temporary for multi-instruction
1334 result_dst
.writemask
= (1 << ir
->type
->vector_elements
) - 1;
1336 switch (ir
->operation
) {
1337 case ir_unop_logic_not
:
1338 emit(ir
, TGSI_OPCODE_SEQ
, result_dst
, op
[0], st_src_reg_for_type(result_dst
.type
, 0));
1341 assert(result_dst
.type
== GLSL_TYPE_FLOAT
|| result_dst
.type
== GLSL_TYPE_INT
);
1342 if (result_dst
.type
== GLSL_TYPE_INT
)
1343 emit(ir
, TGSI_OPCODE_INEG
, result_dst
, op
[0]);
1345 op
[0].negate
= ~op
[0].negate
;
1350 assert(result_dst
.type
== GLSL_TYPE_FLOAT
);
1351 emit(ir
, TGSI_OPCODE_ABS
, result_dst
, op
[0]);
1354 emit(ir
, TGSI_OPCODE_SSG
, result_dst
, op
[0]);
1357 emit_scalar(ir
, TGSI_OPCODE_RCP
, result_dst
, op
[0]);
1361 emit_scalar(ir
, TGSI_OPCODE_EX2
, result_dst
, op
[0]);
1365 assert(!"not reached: should be handled by ir_explog_to_explog2");
1368 emit_scalar(ir
, TGSI_OPCODE_LG2
, result_dst
, op
[0]);
1371 emit_scalar(ir
, TGSI_OPCODE_SIN
, result_dst
, op
[0]);
1374 emit_scalar(ir
, TGSI_OPCODE_COS
, result_dst
, op
[0]);
1376 case ir_unop_sin_reduced
:
1377 emit_scs(ir
, TGSI_OPCODE_SIN
, result_dst
, op
[0]);
1379 case ir_unop_cos_reduced
:
1380 emit_scs(ir
, TGSI_OPCODE_COS
, result_dst
, op
[0]);
1384 emit(ir
, TGSI_OPCODE_DDX
, result_dst
, op
[0]);
1387 op
[0].negate
= ~op
[0].negate
;
1388 emit(ir
, TGSI_OPCODE_DDY
, result_dst
, op
[0]);
1391 case ir_unop_noise
: {
1392 /* At some point, a motivated person could add a better
1393 * implementation of noise. Currently not even the nvidia
1394 * binary drivers do anything more than this. In any case, the
1395 * place to do this is in the GL state tracker, not the poor
1398 emit(ir
, TGSI_OPCODE_MOV
, result_dst
, st_src_reg_for_float(0.5));
1403 emit(ir
, TGSI_OPCODE_ADD
, result_dst
, op
[0], op
[1]);
1406 emit(ir
, TGSI_OPCODE_SUB
, result_dst
, op
[0], op
[1]);
1410 emit(ir
, TGSI_OPCODE_MUL
, result_dst
, op
[0], op
[1]);
1413 if (result_dst
.type
== GLSL_TYPE_FLOAT
)
1414 assert(!"not reached: should be handled by ir_div_to_mul_rcp");
1416 emit(ir
, TGSI_OPCODE_DIV
, result_dst
, op
[0], op
[1]);
1419 if (result_dst
.type
== GLSL_TYPE_FLOAT
)
1420 assert(!"ir_binop_mod should have been converted to b * fract(a/b)");
1422 emit(ir
, TGSI_OPCODE_MOD
, result_dst
, op
[0], op
[1]);
1426 emit(ir
, TGSI_OPCODE_SLT
, result_dst
, op
[0], op
[1]);
1428 case ir_binop_greater
:
1429 emit(ir
, TGSI_OPCODE_SGT
, result_dst
, op
[0], op
[1]);
1431 case ir_binop_lequal
:
1432 emit(ir
, TGSI_OPCODE_SLE
, result_dst
, op
[0], op
[1]);
1434 case ir_binop_gequal
:
1435 emit(ir
, TGSI_OPCODE_SGE
, result_dst
, op
[0], op
[1]);
1437 case ir_binop_equal
:
1438 emit(ir
, TGSI_OPCODE_SEQ
, result_dst
, op
[0], op
[1]);
1440 case ir_binop_nequal
:
1441 emit(ir
, TGSI_OPCODE_SNE
, result_dst
, op
[0], op
[1]);
1443 case ir_binop_all_equal
:
1444 /* "==" operator producing a scalar boolean. */
1445 if (ir
->operands
[0]->type
->is_vector() ||
1446 ir
->operands
[1]->type
->is_vector()) {
1447 st_src_reg temp
= get_temp(glsl_version
>= 130 ?
1448 glsl_type::get_instance(ir
->operands
[0]->type
->base_type
, 4, 1) :
1449 glsl_type::vec4_type
);
1450 assert(ir
->operands
[0]->type
->base_type
== GLSL_TYPE_FLOAT
);
1451 emit(ir
, TGSI_OPCODE_SNE
, st_dst_reg(temp
), op
[0], op
[1]);
1452 emit_dp(ir
, result_dst
, temp
, temp
, vector_elements
);
1453 emit(ir
, TGSI_OPCODE_SEQ
, result_dst
, result_src
, st_src_reg_for_float(0.0));
1455 emit(ir
, TGSI_OPCODE_SEQ
, result_dst
, op
[0], op
[1]);
1458 case ir_binop_any_nequal
:
1459 /* "!=" operator producing a scalar boolean. */
1460 if (ir
->operands
[0]->type
->is_vector() ||
1461 ir
->operands
[1]->type
->is_vector()) {
1462 st_src_reg temp
= get_temp(glsl_version
>= 130 ?
1463 glsl_type::get_instance(ir
->operands
[0]->type
->base_type
, 4, 1) :
1464 glsl_type::vec4_type
);
1465 assert(ir
->operands
[0]->type
->base_type
== GLSL_TYPE_FLOAT
);
1466 emit(ir
, TGSI_OPCODE_SNE
, st_dst_reg(temp
), op
[0], op
[1]);
1467 emit_dp(ir
, result_dst
, temp
, temp
, vector_elements
);
1468 emit(ir
, TGSI_OPCODE_SNE
, result_dst
, result_src
, st_src_reg_for_float(0.0));
1470 emit(ir
, TGSI_OPCODE_SNE
, result_dst
, op
[0], op
[1]);
1475 assert(ir
->operands
[0]->type
->is_vector());
1476 emit_dp(ir
, result_dst
, op
[0], op
[0],
1477 ir
->operands
[0]->type
->vector_elements
);
1478 emit(ir
, TGSI_OPCODE_SNE
, result_dst
, result_src
, st_src_reg_for_float(0.0));
1481 case ir_binop_logic_xor
:
1482 emit(ir
, TGSI_OPCODE_SNE
, result_dst
, op
[0], op
[1]);
1485 case ir_binop_logic_or
:
1486 /* This could be a saturated add and skip the SNE. */
1487 emit(ir
, TGSI_OPCODE_ADD
, result_dst
, op
[0], op
[1]);
1488 emit(ir
, TGSI_OPCODE_SNE
, result_dst
, result_src
, st_src_reg_for_float(0.0));
1491 case ir_binop_logic_and
:
1492 /* the bool args are stored as float 0.0 or 1.0, so "mul" gives us "and". */
1493 emit(ir
, TGSI_OPCODE_MUL
, result_dst
, op
[0], op
[1]);
1497 assert(ir
->operands
[0]->type
->is_vector());
1498 assert(ir
->operands
[0]->type
== ir
->operands
[1]->type
);
1499 emit_dp(ir
, result_dst
, op
[0], op
[1],
1500 ir
->operands
[0]->type
->vector_elements
);
1504 /* sqrt(x) = x * rsq(x). */
1505 emit_scalar(ir
, TGSI_OPCODE_RSQ
, result_dst
, op
[0]);
1506 emit(ir
, TGSI_OPCODE_MUL
, result_dst
, result_src
, op
[0]);
1507 /* For incoming channels <= 0, set the result to 0. */
1508 op
[0].negate
= ~op
[0].negate
;
1509 emit(ir
, TGSI_OPCODE_CMP
, result_dst
,
1510 op
[0], result_src
, st_src_reg_for_float(0.0));
1513 emit_scalar(ir
, TGSI_OPCODE_RSQ
, result_dst
, op
[0]);
1517 if (glsl_version
>= 130) {
1518 emit(ir
, TGSI_OPCODE_I2F
, result_dst
, op
[0]);
1523 /* Converting between signed and unsigned integers is a no-op. */
1525 /* Booleans are stored as integers (or floats in GLSL 1.20 and lower). */
1529 if (glsl_version
>= 130)
1530 emit(ir
, TGSI_OPCODE_F2I
, result_dst
, op
[0]);
1532 emit(ir
, TGSI_OPCODE_TRUNC
, result_dst
, op
[0]);
1536 emit(ir
, TGSI_OPCODE_SNE
, result_dst
, op
[0],
1537 st_src_reg_for_type(result_dst
.type
, 0));
1540 emit(ir
, TGSI_OPCODE_TRUNC
, result_dst
, op
[0]);
1543 op
[0].negate
= ~op
[0].negate
;
1544 emit(ir
, TGSI_OPCODE_FLR
, result_dst
, op
[0]);
1545 result_src
.negate
= ~result_src
.negate
;
1548 emit(ir
, TGSI_OPCODE_FLR
, result_dst
, op
[0]);
1551 emit(ir
, TGSI_OPCODE_FRC
, result_dst
, op
[0]);
1555 emit(ir
, TGSI_OPCODE_MIN
, result_dst
, op
[0], op
[1]);
1558 emit(ir
, TGSI_OPCODE_MAX
, result_dst
, op
[0], op
[1]);
1561 emit_scalar(ir
, TGSI_OPCODE_POW
, result_dst
, op
[0], op
[1]);
1564 case ir_unop_bit_not
:
1565 if (glsl_version
>= 130) {
1566 emit(ir
, TGSI_OPCODE_NOT
, result_dst
, op
[0]);
1570 if (glsl_version
>= 130) {
1571 emit(ir
, TGSI_OPCODE_U2F
, result_dst
, op
[0]);
1574 case ir_binop_lshift
:
1575 if (glsl_version
>= 130) {
1576 emit(ir
, TGSI_OPCODE_SHL
, result_dst
, op
[0]);
1579 case ir_binop_rshift
:
1580 if (glsl_version
>= 130) {
1581 emit(ir
, TGSI_OPCODE_ISHR
, result_dst
, op
[0]);
1584 case ir_binop_bit_and
:
1585 if (glsl_version
>= 130) {
1586 emit(ir
, TGSI_OPCODE_AND
, result_dst
, op
[0]);
1589 case ir_binop_bit_xor
:
1590 if (glsl_version
>= 130) {
1591 emit(ir
, TGSI_OPCODE_XOR
, result_dst
, op
[0]);
1594 case ir_binop_bit_or
:
1595 if (glsl_version
>= 130) {
1596 emit(ir
, TGSI_OPCODE_OR
, result_dst
, op
[0]);
1599 case ir_unop_round_even
:
1600 assert(!"GLSL 1.30 features unsupported");
1603 case ir_quadop_vector
:
1604 /* This operation should have already been handled.
1606 assert(!"Should not get here.");
1610 this->result
= result_src
;
1615 glsl_to_tgsi_visitor::visit(ir_swizzle
*ir
)
1621 /* Note that this is only swizzles in expressions, not those on the left
1622 * hand side of an assignment, which do write masking. See ir_assignment
1626 ir
->val
->accept(this);
1628 assert(src
.file
!= PROGRAM_UNDEFINED
);
1630 for (i
= 0; i
< 4; i
++) {
1631 if (i
< ir
->type
->vector_elements
) {
1634 swizzle
[i
] = GET_SWZ(src
.swizzle
, ir
->mask
.x
);
1637 swizzle
[i
] = GET_SWZ(src
.swizzle
, ir
->mask
.y
);
1640 swizzle
[i
] = GET_SWZ(src
.swizzle
, ir
->mask
.z
);
1643 swizzle
[i
] = GET_SWZ(src
.swizzle
, ir
->mask
.w
);
1647 /* If the type is smaller than a vec4, replicate the last
1650 swizzle
[i
] = swizzle
[ir
->type
->vector_elements
- 1];
1654 src
.swizzle
= MAKE_SWIZZLE4(swizzle
[0], swizzle
[1], swizzle
[2], swizzle
[3]);
1660 glsl_to_tgsi_visitor::visit(ir_dereference_variable
*ir
)
1662 variable_storage
*entry
= find_variable_storage(ir
->var
);
1663 ir_variable
*var
= ir
->var
;
1666 switch (var
->mode
) {
1667 case ir_var_uniform
:
1668 entry
= new(mem_ctx
) variable_storage(var
, PROGRAM_UNIFORM
,
1670 this->variables
.push_tail(entry
);
1674 /* The linker assigns locations for varyings and attributes,
1675 * including deprecated builtins (like gl_Color), user-assign
1676 * generic attributes (glBindVertexLocation), and
1677 * user-defined varyings.
1679 * FINISHME: We would hit this path for function arguments. Fix!
1681 assert(var
->location
!= -1);
1682 entry
= new(mem_ctx
) variable_storage(var
,
1685 if (this->prog
->Target
== GL_VERTEX_PROGRAM_ARB
&&
1686 var
->location
>= VERT_ATTRIB_GENERIC0
) {
1687 _mesa_add_attribute(this->prog
->Attributes
,
1689 _mesa_sizeof_glsl_type(var
->type
->gl_type
),
1691 var
->location
- VERT_ATTRIB_GENERIC0
);
1695 assert(var
->location
!= -1);
1696 entry
= new(mem_ctx
) variable_storage(var
,
1700 case ir_var_system_value
:
1701 entry
= new(mem_ctx
) variable_storage(var
,
1702 PROGRAM_SYSTEM_VALUE
,
1706 case ir_var_temporary
:
1707 entry
= new(mem_ctx
) variable_storage(var
, PROGRAM_TEMPORARY
,
1709 this->variables
.push_tail(entry
);
1711 next_temp
+= type_size(var
->type
);
1716 printf("Failed to make storage for %s\n", var
->name
);
1721 this->result
= st_src_reg(entry
->file
, entry
->index
, var
->type
);
1722 if (glsl_version
<= 120)
1723 this->result
.type
= GLSL_TYPE_FLOAT
;
1727 glsl_to_tgsi_visitor::visit(ir_dereference_array
*ir
)
1731 int element_size
= type_size(ir
->type
);
1733 index
= ir
->array_index
->constant_expression_value();
1735 ir
->array
->accept(this);
1739 src
.index
+= index
->value
.i
[0] * element_size
;
1741 /* Variable index array dereference. It eats the "vec4" of the
1742 * base of the array and an index that offsets the TGSI register
1745 ir
->array_index
->accept(this);
1747 st_src_reg index_reg
;
1749 if (element_size
== 1) {
1750 index_reg
= this->result
;
1752 index_reg
= get_temp(glsl_type::float_type
);
1754 emit(ir
, TGSI_OPCODE_MUL
, st_dst_reg(index_reg
),
1755 this->result
, st_src_reg_for_float(element_size
));
1758 /* If there was already a relative address register involved, add the
1759 * new and the old together to get the new offset.
1761 if (src
.reladdr
!= NULL
) {
1762 st_src_reg accum_reg
= get_temp(glsl_type::float_type
);
1764 emit(ir
, TGSI_OPCODE_ADD
, st_dst_reg(accum_reg
),
1765 index_reg
, *src
.reladdr
);
1767 index_reg
= accum_reg
;
1770 src
.reladdr
= ralloc(mem_ctx
, st_src_reg
);
1771 memcpy(src
.reladdr
, &index_reg
, sizeof(index_reg
));
1774 /* If the type is smaller than a vec4, replicate the last channel out. */
1775 if (ir
->type
->is_scalar() || ir
->type
->is_vector())
1776 src
.swizzle
= swizzle_for_size(ir
->type
->vector_elements
);
1778 src
.swizzle
= SWIZZLE_NOOP
;
1784 glsl_to_tgsi_visitor::visit(ir_dereference_record
*ir
)
1787 const glsl_type
*struct_type
= ir
->record
->type
;
1790 ir
->record
->accept(this);
1792 for (i
= 0; i
< struct_type
->length
; i
++) {
1793 if (strcmp(struct_type
->fields
.structure
[i
].name
, ir
->field
) == 0)
1795 offset
+= type_size(struct_type
->fields
.structure
[i
].type
);
1798 /* If the type is smaller than a vec4, replicate the last channel out. */
1799 if (ir
->type
->is_scalar() || ir
->type
->is_vector())
1800 this->result
.swizzle
= swizzle_for_size(ir
->type
->vector_elements
);
1802 this->result
.swizzle
= SWIZZLE_NOOP
;
1804 this->result
.index
+= offset
;
1808 * We want to be careful in assignment setup to hit the actual storage
1809 * instead of potentially using a temporary like we might with the
1810 * ir_dereference handler.
1813 get_assignment_lhs(ir_dereference
*ir
, glsl_to_tgsi_visitor
*v
)
1815 /* The LHS must be a dereference. If the LHS is a variable indexed array
1816 * access of a vector, it must be separated into a series conditional moves
1817 * before reaching this point (see ir_vec_index_to_cond_assign).
1819 assert(ir
->as_dereference());
1820 ir_dereference_array
*deref_array
= ir
->as_dereference_array();
1822 assert(!deref_array
->array
->type
->is_vector());
1825 /* Use the rvalue deref handler for the most part. We'll ignore
1826 * swizzles in it and write swizzles using writemask, though.
1829 return st_dst_reg(v
->result
);
1833 * Process the condition of a conditional assignment
1835 * Examines the condition of a conditional assignment to generate the optimal
1836 * first operand of a \c CMP instruction. If the condition is a relational
1837 * operator with 0 (e.g., \c ir_binop_less), the value being compared will be
1838 * used as the source for the \c CMP instruction. Otherwise the comparison
1839 * is processed to a boolean result, and the boolean result is used as the
1840 * operand to the CMP instruction.
1843 glsl_to_tgsi_visitor::process_move_condition(ir_rvalue
*ir
)
1845 ir_rvalue
*src_ir
= ir
;
1847 bool switch_order
= false;
1849 ir_expression
*const expr
= ir
->as_expression();
1850 if ((expr
!= NULL
) && (expr
->get_num_operands() == 2)) {
1851 bool zero_on_left
= false;
1853 if (expr
->operands
[0]->is_zero()) {
1854 src_ir
= expr
->operands
[1];
1855 zero_on_left
= true;
1856 } else if (expr
->operands
[1]->is_zero()) {
1857 src_ir
= expr
->operands
[0];
1858 zero_on_left
= false;
1862 * (a < 0) T F F ( a < 0) T F F
1863 * (0 < a) F F T (-a < 0) F F T
1864 * (a <= 0) T T F (-a < 0) F F T (swap order of other operands)
1865 * (0 <= a) F T T ( a < 0) T F F (swap order of other operands)
1866 * (a > 0) F F T (-a < 0) F F T
1867 * (0 > a) T F F ( a < 0) T F F
1868 * (a >= 0) F T T ( a < 0) T F F (swap order of other operands)
1869 * (0 >= a) T T F (-a < 0) F F T (swap order of other operands)
1871 * Note that exchanging the order of 0 and 'a' in the comparison simply
1872 * means that the value of 'a' should be negated.
1875 switch (expr
->operation
) {
1877 switch_order
= false;
1878 negate
= zero_on_left
;
1881 case ir_binop_greater
:
1882 switch_order
= false;
1883 negate
= !zero_on_left
;
1886 case ir_binop_lequal
:
1887 switch_order
= true;
1888 negate
= !zero_on_left
;
1891 case ir_binop_gequal
:
1892 switch_order
= true;
1893 negate
= zero_on_left
;
1897 /* This isn't the right kind of comparison afterall, so make sure
1898 * the whole condition is visited.
1906 src_ir
->accept(this);
1908 /* We use the TGSI_OPCODE_CMP (a < 0 ? b : c) for conditional moves, and the
1909 * condition we produced is 0.0 or 1.0. By flipping the sign, we can
1910 * choose which value TGSI_OPCODE_CMP produces without an extra instruction
1911 * computing the condition.
1914 this->result
.negate
= ~this->result
.negate
;
1916 return switch_order
;
1920 glsl_to_tgsi_visitor::visit(ir_assignment
*ir
)
1926 ir
->rhs
->accept(this);
1929 l
= get_assignment_lhs(ir
->lhs
, this);
1931 /* FINISHME: This should really set to the correct maximal writemask for each
1932 * FINISHME: component written (in the loops below). This case can only
1933 * FINISHME: occur for matrices, arrays, and structures.
1935 if (ir
->write_mask
== 0) {
1936 assert(!ir
->lhs
->type
->is_scalar() && !ir
->lhs
->type
->is_vector());
1937 l
.writemask
= WRITEMASK_XYZW
;
1938 } else if (ir
->lhs
->type
->is_scalar() &&
1939 ir
->lhs
->variable_referenced()->mode
== ir_var_out
) {
1940 /* FINISHME: This hack makes writing to gl_FragDepth, which lives in the
1941 * FINISHME: W component of fragment shader output zero, work correctly.
1943 l
.writemask
= WRITEMASK_XYZW
;
1946 int first_enabled_chan
= 0;
1949 l
.writemask
= ir
->write_mask
;
1951 for (int i
= 0; i
< 4; i
++) {
1952 if (l
.writemask
& (1 << i
)) {
1953 first_enabled_chan
= GET_SWZ(r
.swizzle
, i
);
1958 /* Swizzle a small RHS vector into the channels being written.
1960 * glsl ir treats write_mask as dictating how many channels are
1961 * present on the RHS while TGSI treats write_mask as just
1962 * showing which channels of the vec4 RHS get written.
1964 for (int i
= 0; i
< 4; i
++) {
1965 if (l
.writemask
& (1 << i
))
1966 swizzles
[i
] = GET_SWZ(r
.swizzle
, rhs_chan
++);
1968 swizzles
[i
] = first_enabled_chan
;
1970 r
.swizzle
= MAKE_SWIZZLE4(swizzles
[0], swizzles
[1],
1971 swizzles
[2], swizzles
[3]);
1974 assert(l
.file
!= PROGRAM_UNDEFINED
);
1975 assert(r
.file
!= PROGRAM_UNDEFINED
);
1977 if (ir
->condition
) {
1978 const bool switch_order
= this->process_move_condition(ir
->condition
);
1979 st_src_reg condition
= this->result
;
1981 for (i
= 0; i
< type_size(ir
->lhs
->type
); i
++) {
1982 st_src_reg l_src
= st_src_reg(l
);
1983 l_src
.swizzle
= swizzle_for_size(ir
->lhs
->type
->vector_elements
);
1986 emit(ir
, TGSI_OPCODE_CMP
, l
, condition
, l_src
, r
);
1988 emit(ir
, TGSI_OPCODE_CMP
, l
, condition
, r
, l_src
);
1994 } else if (ir
->rhs
->as_expression() &&
1995 this->instructions
.get_tail() &&
1996 ir
->rhs
== ((glsl_to_tgsi_instruction
*)this->instructions
.get_tail())->ir
&&
1997 type_size(ir
->lhs
->type
) == 1) {
1998 /* To avoid emitting an extra MOV when assigning an expression to a
1999 * variable, emit the last instruction of the expression again, but
2000 * replace the destination register with the target of the assignment.
2001 * Dead code elimination will remove the original instruction.
2003 glsl_to_tgsi_instruction
*inst
;
2004 inst
= (glsl_to_tgsi_instruction
*)this->instructions
.get_tail();
2005 emit(ir
, inst
->op
, l
, inst
->src
[0], inst
->src
[1], inst
->src
[2]);
2007 for (i
= 0; i
< type_size(ir
->lhs
->type
); i
++) {
2008 emit(ir
, TGSI_OPCODE_MOV
, l
, r
);
2017 glsl_to_tgsi_visitor::visit(ir_constant
*ir
)
2020 GLfloat stack_vals
[4] = { 0 };
2021 gl_constant_value
*values
= (gl_constant_value
*) stack_vals
;
2022 GLenum gl_type
= GL_NONE
;
2024 static int in_array
= 0;
2025 gl_register_file file
= in_array
? PROGRAM_CONSTANT
: PROGRAM_IMMEDIATE
;
2027 /* Unfortunately, 4 floats is all we can get into
2028 * _mesa_add_typed_unnamed_constant. So, make a temp to store an
2029 * aggregate constant and move each constant value into it. If we
2030 * get lucky, copy propagation will eliminate the extra moves.
2032 if (ir
->type
->base_type
== GLSL_TYPE_STRUCT
) {
2033 st_src_reg temp_base
= get_temp(ir
->type
);
2034 st_dst_reg temp
= st_dst_reg(temp_base
);
2036 foreach_iter(exec_list_iterator
, iter
, ir
->components
) {
2037 ir_constant
*field_value
= (ir_constant
*)iter
.get();
2038 int size
= type_size(field_value
->type
);
2042 field_value
->accept(this);
2045 for (i
= 0; i
< (unsigned int)size
; i
++) {
2046 emit(ir
, TGSI_OPCODE_MOV
, temp
, src
);
2052 this->result
= temp_base
;
2056 if (ir
->type
->is_array()) {
2057 st_src_reg temp_base
= get_temp(ir
->type
);
2058 st_dst_reg temp
= st_dst_reg(temp_base
);
2059 int size
= type_size(ir
->type
->fields
.array
);
2064 for (i
= 0; i
< ir
->type
->length
; i
++) {
2065 ir
->array_elements
[i
]->accept(this);
2067 for (int j
= 0; j
< size
; j
++) {
2068 emit(ir
, TGSI_OPCODE_MOV
, temp
, src
);
2074 this->result
= temp_base
;
2079 if (ir
->type
->is_matrix()) {
2080 st_src_reg mat
= get_temp(ir
->type
);
2081 st_dst_reg mat_column
= st_dst_reg(mat
);
2083 for (i
= 0; i
< ir
->type
->matrix_columns
; i
++) {
2084 assert(ir
->type
->base_type
== GLSL_TYPE_FLOAT
);
2085 values
= (gl_constant_value
*) &ir
->value
.f
[i
* ir
->type
->vector_elements
];
2087 src
= st_src_reg(file
, -1, ir
->type
->base_type
);
2088 src
.index
= add_constant(file
,
2090 ir
->type
->vector_elements
,
2093 emit(ir
, TGSI_OPCODE_MOV
, mat_column
, src
);
2102 switch (ir
->type
->base_type
) {
2103 case GLSL_TYPE_FLOAT
:
2105 for (i
= 0; i
< ir
->type
->vector_elements
; i
++) {
2106 values
[i
].f
= ir
->value
.f
[i
];
2109 case GLSL_TYPE_UINT
:
2110 gl_type
= glsl_version
>= 130 ? GL_UNSIGNED_INT
: GL_FLOAT
;
2111 for (i
= 0; i
< ir
->type
->vector_elements
; i
++) {
2112 if (glsl_version
>= 130)
2113 values
[i
].u
= ir
->value
.u
[i
];
2115 values
[i
].f
= ir
->value
.u
[i
];
2119 gl_type
= glsl_version
>= 130 ? GL_INT
: GL_FLOAT
;
2120 for (i
= 0; i
< ir
->type
->vector_elements
; i
++) {
2121 if (glsl_version
>= 130)
2122 values
[i
].i
= ir
->value
.i
[i
];
2124 values
[i
].f
= ir
->value
.i
[i
];
2127 case GLSL_TYPE_BOOL
:
2128 gl_type
= glsl_version
>= 130 ? GL_BOOL
: GL_FLOAT
;
2129 for (i
= 0; i
< ir
->type
->vector_elements
; i
++) {
2130 if (glsl_version
>= 130)
2131 values
[i
].b
= ir
->value
.b
[i
];
2133 values
[i
].f
= ir
->value
.b
[i
];
2137 assert(!"Non-float/uint/int/bool constant");
2140 this->result
= st_src_reg(file
, -1, ir
->type
);
2141 this->result
.index
= add_constant(file
,
2143 ir
->type
->vector_elements
,
2145 &this->result
.swizzle
);
2149 glsl_to_tgsi_visitor::get_function_signature(ir_function_signature
*sig
)
2151 function_entry
*entry
;
2153 foreach_iter(exec_list_iterator
, iter
, this->function_signatures
) {
2154 entry
= (function_entry
*)iter
.get();
2156 if (entry
->sig
== sig
)
2160 entry
= ralloc(mem_ctx
, function_entry
);
2162 entry
->sig_id
= this->next_signature_id
++;
2163 entry
->bgn_inst
= NULL
;
2165 /* Allocate storage for all the parameters. */
2166 foreach_iter(exec_list_iterator
, iter
, sig
->parameters
) {
2167 ir_variable
*param
= (ir_variable
*)iter
.get();
2168 variable_storage
*storage
;
2170 storage
= find_variable_storage(param
);
2173 storage
= new(mem_ctx
) variable_storage(param
, PROGRAM_TEMPORARY
,
2175 this->variables
.push_tail(storage
);
2177 this->next_temp
+= type_size(param
->type
);
2180 if (!sig
->return_type
->is_void()) {
2181 entry
->return_reg
= get_temp(sig
->return_type
);
2183 entry
->return_reg
= undef_src
;
2186 this->function_signatures
.push_tail(entry
);
2191 glsl_to_tgsi_visitor::visit(ir_call
*ir
)
2193 glsl_to_tgsi_instruction
*call_inst
;
2194 ir_function_signature
*sig
= ir
->get_callee();
2195 function_entry
*entry
= get_function_signature(sig
);
2198 /* Process in parameters. */
2199 exec_list_iterator sig_iter
= sig
->parameters
.iterator();
2200 foreach_iter(exec_list_iterator
, iter
, *ir
) {
2201 ir_rvalue
*param_rval
= (ir_rvalue
*)iter
.get();
2202 ir_variable
*param
= (ir_variable
*)sig_iter
.get();
2204 if (param
->mode
== ir_var_in
||
2205 param
->mode
== ir_var_inout
) {
2206 variable_storage
*storage
= find_variable_storage(param
);
2209 param_rval
->accept(this);
2210 st_src_reg r
= this->result
;
2213 l
.file
= storage
->file
;
2214 l
.index
= storage
->index
;
2216 l
.writemask
= WRITEMASK_XYZW
;
2217 l
.cond_mask
= COND_TR
;
2219 for (i
= 0; i
< type_size(param
->type
); i
++) {
2220 emit(ir
, TGSI_OPCODE_MOV
, l
, r
);
2228 assert(!sig_iter
.has_next());
2230 /* Emit call instruction */
2231 call_inst
= emit(ir
, TGSI_OPCODE_CAL
);
2232 call_inst
->function
= entry
;
2234 /* Process out parameters. */
2235 sig_iter
= sig
->parameters
.iterator();
2236 foreach_iter(exec_list_iterator
, iter
, *ir
) {
2237 ir_rvalue
*param_rval
= (ir_rvalue
*)iter
.get();
2238 ir_variable
*param
= (ir_variable
*)sig_iter
.get();
2240 if (param
->mode
== ir_var_out
||
2241 param
->mode
== ir_var_inout
) {
2242 variable_storage
*storage
= find_variable_storage(param
);
2246 r
.file
= storage
->file
;
2247 r
.index
= storage
->index
;
2249 r
.swizzle
= SWIZZLE_NOOP
;
2252 param_rval
->accept(this);
2253 st_dst_reg l
= st_dst_reg(this->result
);
2255 for (i
= 0; i
< type_size(param
->type
); i
++) {
2256 emit(ir
, TGSI_OPCODE_MOV
, l
, r
);
2264 assert(!sig_iter
.has_next());
2266 /* Process return value. */
2267 this->result
= entry
->return_reg
;
2271 glsl_to_tgsi_visitor::visit(ir_texture
*ir
)
2273 st_src_reg result_src
, coord
, lod_info
, projector
, dx
, dy
;
2274 st_dst_reg result_dst
, coord_dst
;
2275 glsl_to_tgsi_instruction
*inst
= NULL
;
2276 unsigned opcode
= TGSI_OPCODE_NOP
;
2278 ir
->coordinate
->accept(this);
2280 /* Put our coords in a temp. We'll need to modify them for shadow,
2281 * projection, or LOD, so the only case we'd use it as is is if
2282 * we're doing plain old texturing. The optimization passes on
2283 * glsl_to_tgsi_visitor should handle cleaning up our mess in that case.
2285 coord
= get_temp(glsl_type::vec4_type
);
2286 coord_dst
= st_dst_reg(coord
);
2287 emit(ir
, TGSI_OPCODE_MOV
, coord_dst
, this->result
);
2289 if (ir
->projector
) {
2290 ir
->projector
->accept(this);
2291 projector
= this->result
;
2294 /* Storage for our result. Ideally for an assignment we'd be using
2295 * the actual storage for the result here, instead.
2297 result_src
= get_temp(glsl_type::vec4_type
);
2298 result_dst
= st_dst_reg(result_src
);
2302 opcode
= TGSI_OPCODE_TEX
;
2305 opcode
= TGSI_OPCODE_TXB
;
2306 ir
->lod_info
.bias
->accept(this);
2307 lod_info
= this->result
;
2310 opcode
= TGSI_OPCODE_TXL
;
2311 ir
->lod_info
.lod
->accept(this);
2312 lod_info
= this->result
;
2315 opcode
= TGSI_OPCODE_TXD
;
2316 ir
->lod_info
.grad
.dPdx
->accept(this);
2318 ir
->lod_info
.grad
.dPdy
->accept(this);
2321 case ir_txf
: /* TODO: use TGSI_OPCODE_TXF here */
2322 assert(!"GLSL 1.30 features unsupported");
2326 if (ir
->projector
) {
2327 if (opcode
== TGSI_OPCODE_TEX
) {
2328 /* Slot the projector in as the last component of the coord. */
2329 coord_dst
.writemask
= WRITEMASK_W
;
2330 emit(ir
, TGSI_OPCODE_MOV
, coord_dst
, projector
);
2331 coord_dst
.writemask
= WRITEMASK_XYZW
;
2332 opcode
= TGSI_OPCODE_TXP
;
2334 st_src_reg coord_w
= coord
;
2335 coord_w
.swizzle
= SWIZZLE_WWWW
;
2337 /* For the other TEX opcodes there's no projective version
2338 * since the last slot is taken up by LOD info. Do the
2339 * projective divide now.
2341 coord_dst
.writemask
= WRITEMASK_W
;
2342 emit(ir
, TGSI_OPCODE_RCP
, coord_dst
, projector
);
2344 /* In the case where we have to project the coordinates "by hand,"
2345 * the shadow comparator value must also be projected.
2347 st_src_reg tmp_src
= coord
;
2348 if (ir
->shadow_comparitor
) {
2349 /* Slot the shadow value in as the second to last component of the
2352 ir
->shadow_comparitor
->accept(this);
2354 tmp_src
= get_temp(glsl_type::vec4_type
);
2355 st_dst_reg tmp_dst
= st_dst_reg(tmp_src
);
2357 tmp_dst
.writemask
= WRITEMASK_Z
;
2358 emit(ir
, TGSI_OPCODE_MOV
, tmp_dst
, this->result
);
2360 tmp_dst
.writemask
= WRITEMASK_XY
;
2361 emit(ir
, TGSI_OPCODE_MOV
, tmp_dst
, coord
);
2364 coord_dst
.writemask
= WRITEMASK_XYZ
;
2365 emit(ir
, TGSI_OPCODE_MUL
, coord_dst
, tmp_src
, coord_w
);
2367 coord_dst
.writemask
= WRITEMASK_XYZW
;
2368 coord
.swizzle
= SWIZZLE_XYZW
;
2372 /* If projection is done and the opcode is not TGSI_OPCODE_TXP, then the shadow
2373 * comparator was put in the correct place (and projected) by the code,
2374 * above, that handles by-hand projection.
2376 if (ir
->shadow_comparitor
&& (!ir
->projector
|| opcode
== TGSI_OPCODE_TXP
)) {
2377 /* Slot the shadow value in as the second to last component of the
2380 ir
->shadow_comparitor
->accept(this);
2381 coord_dst
.writemask
= WRITEMASK_Z
;
2382 emit(ir
, TGSI_OPCODE_MOV
, coord_dst
, this->result
);
2383 coord_dst
.writemask
= WRITEMASK_XYZW
;
2386 if (opcode
== TGSI_OPCODE_TXL
|| opcode
== TGSI_OPCODE_TXB
) {
2387 /* TGSI stores LOD or LOD bias in the last channel of the coords. */
2388 coord_dst
.writemask
= WRITEMASK_W
;
2389 emit(ir
, TGSI_OPCODE_MOV
, coord_dst
, lod_info
);
2390 coord_dst
.writemask
= WRITEMASK_XYZW
;
2393 if (opcode
== TGSI_OPCODE_TXD
)
2394 inst
= emit(ir
, opcode
, result_dst
, coord
, dx
, dy
);
2396 inst
= emit(ir
, opcode
, result_dst
, coord
);
2398 if (ir
->shadow_comparitor
)
2399 inst
->tex_shadow
= GL_TRUE
;
2401 inst
->sampler
= _mesa_get_sampler_uniform_value(ir
->sampler
,
2402 this->shader_program
,
2405 const glsl_type
*sampler_type
= ir
->sampler
->type
;
2407 switch (sampler_type
->sampler_dimensionality
) {
2408 case GLSL_SAMPLER_DIM_1D
:
2409 inst
->tex_target
= (sampler_type
->sampler_array
)
2410 ? TEXTURE_1D_ARRAY_INDEX
: TEXTURE_1D_INDEX
;
2412 case GLSL_SAMPLER_DIM_2D
:
2413 inst
->tex_target
= (sampler_type
->sampler_array
)
2414 ? TEXTURE_2D_ARRAY_INDEX
: TEXTURE_2D_INDEX
;
2416 case GLSL_SAMPLER_DIM_3D
:
2417 inst
->tex_target
= TEXTURE_3D_INDEX
;
2419 case GLSL_SAMPLER_DIM_CUBE
:
2420 inst
->tex_target
= TEXTURE_CUBE_INDEX
;
2422 case GLSL_SAMPLER_DIM_RECT
:
2423 inst
->tex_target
= TEXTURE_RECT_INDEX
;
2425 case GLSL_SAMPLER_DIM_BUF
:
2426 assert(!"FINISHME: Implement ARB_texture_buffer_object");
2429 assert(!"Should not get here.");
2432 this->result
= result_src
;
2436 glsl_to_tgsi_visitor::visit(ir_return
*ir
)
2438 if (ir
->get_value()) {
2442 assert(current_function
);
2444 ir
->get_value()->accept(this);
2445 st_src_reg r
= this->result
;
2447 l
= st_dst_reg(current_function
->return_reg
);
2449 for (i
= 0; i
< type_size(current_function
->sig
->return_type
); i
++) {
2450 emit(ir
, TGSI_OPCODE_MOV
, l
, r
);
2456 emit(ir
, TGSI_OPCODE_RET
);
2460 glsl_to_tgsi_visitor::visit(ir_discard
*ir
)
2462 struct gl_fragment_program
*fp
= (struct gl_fragment_program
*)this->prog
;
2464 if (ir
->condition
) {
2465 ir
->condition
->accept(this);
2466 this->result
.negate
= ~this->result
.negate
;
2467 emit(ir
, TGSI_OPCODE_KIL
, undef_dst
, this->result
);
2469 emit(ir
, TGSI_OPCODE_KILP
);
2472 fp
->UsesKill
= GL_TRUE
;
2476 glsl_to_tgsi_visitor::visit(ir_if
*ir
)
2478 glsl_to_tgsi_instruction
*cond_inst
, *if_inst
;
2479 glsl_to_tgsi_instruction
*prev_inst
;
2481 prev_inst
= (glsl_to_tgsi_instruction
*)this->instructions
.get_tail();
2483 ir
->condition
->accept(this);
2484 assert(this->result
.file
!= PROGRAM_UNDEFINED
);
2486 if (this->options
->EmitCondCodes
) {
2487 cond_inst
= (glsl_to_tgsi_instruction
*)this->instructions
.get_tail();
2489 /* See if we actually generated any instruction for generating
2490 * the condition. If not, then cook up a move to a temp so we
2491 * have something to set cond_update on.
2493 if (cond_inst
== prev_inst
) {
2494 st_src_reg temp
= get_temp(glsl_type::bool_type
);
2495 cond_inst
= emit(ir
->condition
, TGSI_OPCODE_MOV
, st_dst_reg(temp
), result
);
2497 cond_inst
->cond_update
= GL_TRUE
;
2499 if_inst
= emit(ir
->condition
, TGSI_OPCODE_IF
);
2500 if_inst
->dst
.cond_mask
= COND_NE
;
2502 if_inst
= emit(ir
->condition
, TGSI_OPCODE_IF
, undef_dst
, this->result
);
2505 this->instructions
.push_tail(if_inst
);
2507 visit_exec_list(&ir
->then_instructions
, this);
2509 if (!ir
->else_instructions
.is_empty()) {
2510 emit(ir
->condition
, TGSI_OPCODE_ELSE
);
2511 visit_exec_list(&ir
->else_instructions
, this);
2514 if_inst
= emit(ir
->condition
, TGSI_OPCODE_ENDIF
);
2517 glsl_to_tgsi_visitor::glsl_to_tgsi_visitor()
2519 result
.file
= PROGRAM_UNDEFINED
;
2521 next_signature_id
= 1;
2523 current_function
= NULL
;
2524 num_address_regs
= 0;
2525 indirect_addr_temps
= false;
2526 indirect_addr_consts
= false;
2527 mem_ctx
= ralloc_context(NULL
);
2530 glsl_to_tgsi_visitor::~glsl_to_tgsi_visitor()
2532 ralloc_free(mem_ctx
);
2535 extern "C" void free_glsl_to_tgsi_visitor(glsl_to_tgsi_visitor
*v
)
2542 * Count resources used by the given gpu program (number of texture
2546 count_resources(glsl_to_tgsi_visitor
*v
, gl_program
*prog
)
2548 v
->samplers_used
= 0;
2550 foreach_iter(exec_list_iterator
, iter
, v
->instructions
) {
2551 glsl_to_tgsi_instruction
*inst
= (glsl_to_tgsi_instruction
*)iter
.get();
2553 if (is_tex_instruction(inst
->op
)) {
2554 v
->samplers_used
|= 1 << inst
->sampler
;
2556 prog
->SamplerTargets
[inst
->sampler
] =
2557 (gl_texture_index
)inst
->tex_target
;
2558 if (inst
->tex_shadow
) {
2559 prog
->ShadowSamplers
|= 1 << inst
->sampler
;
2564 prog
->SamplersUsed
= v
->samplers_used
;
2565 _mesa_update_shader_textures_used(prog
);
2570 * Check if the given vertex/fragment/shader program is within the
2571 * resource limits of the context (number of texture units, etc).
2572 * If any of those checks fail, record a linker error.
2574 * XXX more checks are needed...
2577 check_resources(const struct gl_context
*ctx
,
2578 struct gl_shader_program
*shader_program
,
2579 glsl_to_tgsi_visitor
*prog
,
2580 struct gl_program
*proginfo
)
2582 switch (proginfo
->Target
) {
2583 case GL_VERTEX_PROGRAM_ARB
:
2584 if (_mesa_bitcount(prog
->samplers_used
) >
2585 ctx
->Const
.MaxVertexTextureImageUnits
) {
2586 fail_link(shader_program
, "Too many vertex shader texture samplers");
2588 if (proginfo
->Parameters
->NumParameters
> MAX_UNIFORMS
) {
2589 fail_link(shader_program
, "Too many vertex shader constants");
2592 case MESA_GEOMETRY_PROGRAM
:
2593 if (_mesa_bitcount(prog
->samplers_used
) >
2594 ctx
->Const
.MaxGeometryTextureImageUnits
) {
2595 fail_link(shader_program
, "Too many geometry shader texture samplers");
2597 if (proginfo
->Parameters
->NumParameters
>
2598 MAX_GEOMETRY_UNIFORM_COMPONENTS
/ 4) {
2599 fail_link(shader_program
, "Too many geometry shader constants");
2602 case GL_FRAGMENT_PROGRAM_ARB
:
2603 if (_mesa_bitcount(prog
->samplers_used
) >
2604 ctx
->Const
.MaxTextureImageUnits
) {
2605 fail_link(shader_program
, "Too many fragment shader texture samplers");
2607 if (proginfo
->Parameters
->NumParameters
> MAX_UNIFORMS
) {
2608 fail_link(shader_program
, "Too many fragment shader constants");
2612 _mesa_problem(ctx
, "unexpected program type in check_resources()");
2618 struct uniform_sort
{
2619 struct gl_uniform
*u
;
2623 /* The shader_program->Uniforms list is almost sorted in increasing
2624 * uniform->{Frag,Vert}Pos locations, but not quite when there are
2625 * uniforms shared between targets. We need to add parameters in
2626 * increasing order for the targets.
2629 sort_uniforms(const void *a
, const void *b
)
2631 struct uniform_sort
*u1
= (struct uniform_sort
*)a
;
2632 struct uniform_sort
*u2
= (struct uniform_sort
*)b
;
2634 return u1
->pos
- u2
->pos
;
2637 /* Add the uniforms to the parameters. The linker chose locations
2638 * in our parameters lists (which weren't created yet), which the
2639 * uniforms code will use to poke values into our parameters list
2640 * when uniforms are updated.
2643 add_uniforms_to_parameters_list(struct gl_shader_program
*shader_program
,
2644 struct gl_shader
*shader
,
2645 struct gl_program
*prog
)
2648 unsigned int next_sampler
= 0, num_uniforms
= 0;
2649 struct uniform_sort
*sorted_uniforms
;
2651 sorted_uniforms
= ralloc_array(NULL
, struct uniform_sort
,
2652 shader_program
->Uniforms
->NumUniforms
);
2654 for (i
= 0; i
< shader_program
->Uniforms
->NumUniforms
; i
++) {
2655 struct gl_uniform
*uniform
= shader_program
->Uniforms
->Uniforms
+ i
;
2656 int parameter_index
= -1;
2658 switch (shader
->Type
) {
2659 case GL_VERTEX_SHADER
:
2660 parameter_index
= uniform
->VertPos
;
2662 case GL_FRAGMENT_SHADER
:
2663 parameter_index
= uniform
->FragPos
;
2665 case GL_GEOMETRY_SHADER
:
2666 parameter_index
= uniform
->GeomPos
;
2670 /* Only add uniforms used in our target. */
2671 if (parameter_index
!= -1) {
2672 sorted_uniforms
[num_uniforms
].pos
= parameter_index
;
2673 sorted_uniforms
[num_uniforms
].u
= uniform
;
2678 qsort(sorted_uniforms
, num_uniforms
, sizeof(struct uniform_sort
),
2681 for (i
= 0; i
< num_uniforms
; i
++) {
2682 struct gl_uniform
*uniform
= sorted_uniforms
[i
].u
;
2683 int parameter_index
= sorted_uniforms
[i
].pos
;
2684 const glsl_type
*type
= uniform
->Type
;
2687 if (type
->is_vector() ||
2688 type
->is_scalar()) {
2689 size
= type
->vector_elements
;
2691 size
= type_size(type
) * 4;
2694 gl_register_file file
;
2695 if (type
->is_sampler() ||
2696 (type
->is_array() && type
->fields
.array
->is_sampler())) {
2697 file
= PROGRAM_SAMPLER
;
2699 file
= PROGRAM_UNIFORM
;
2702 GLint index
= _mesa_lookup_parameter_index(prog
->Parameters
, -1,
2706 index
= _mesa_add_parameter(prog
->Parameters
, file
,
2707 uniform
->Name
, size
, type
->gl_type
,
2710 /* Sampler uniform values are stored in prog->SamplerUnits,
2711 * and the entry in that array is selected by this index we
2712 * store in ParameterValues[].
2714 if (file
== PROGRAM_SAMPLER
) {
2715 for (unsigned int j
= 0; j
< size
/ 4; j
++)
2716 prog
->Parameters
->ParameterValues
[index
+ j
][0].f
= next_sampler
++;
2719 /* The location chosen in the Parameters list here (returned
2720 * from _mesa_add_uniform) has to match what the linker chose.
2722 if (index
!= parameter_index
) {
2723 fail_link(shader_program
, "Allocation of uniform `%s' to target "
2724 "failed (%d vs %d)\n",
2725 uniform
->Name
, index
, parameter_index
);
2730 ralloc_free(sorted_uniforms
);
2734 set_uniform_initializer(struct gl_context
*ctx
, void *mem_ctx
,
2735 struct gl_shader_program
*shader_program
,
2736 const char *name
, const glsl_type
*type
,
2739 if (type
->is_record()) {
2740 ir_constant
*field_constant
;
2742 field_constant
= (ir_constant
*)val
->components
.get_head();
2744 for (unsigned int i
= 0; i
< type
->length
; i
++) {
2745 const glsl_type
*field_type
= type
->fields
.structure
[i
].type
;
2746 const char *field_name
= ralloc_asprintf(mem_ctx
, "%s.%s", name
,
2747 type
->fields
.structure
[i
].name
);
2748 set_uniform_initializer(ctx
, mem_ctx
, shader_program
, field_name
,
2749 field_type
, field_constant
);
2750 field_constant
= (ir_constant
*)field_constant
->next
;
2755 int loc
= _mesa_get_uniform_location(ctx
, shader_program
, name
);
2758 fail_link(shader_program
,
2759 "Couldn't find uniform for initializer %s\n", name
);
2763 for (unsigned int i
= 0; i
< (type
->is_array() ? type
->length
: 1); i
++) {
2764 ir_constant
*element
;
2765 const glsl_type
*element_type
;
2766 if (type
->is_array()) {
2767 element
= val
->array_elements
[i
];
2768 element_type
= type
->fields
.array
;
2771 element_type
= type
;
2776 if (element_type
->base_type
== GLSL_TYPE_BOOL
) {
2777 int *conv
= ralloc_array(mem_ctx
, int, element_type
->components());
2778 for (unsigned int j
= 0; j
< element_type
->components(); j
++) {
2779 conv
[j
] = element
->value
.b
[j
];
2781 values
= (void *)conv
;
2782 element_type
= glsl_type::get_instance(GLSL_TYPE_INT
,
2783 element_type
->vector_elements
,
2786 values
= &element
->value
;
2789 if (element_type
->is_matrix()) {
2790 _mesa_uniform_matrix(ctx
, shader_program
,
2791 element_type
->matrix_columns
,
2792 element_type
->vector_elements
,
2793 loc
, 1, GL_FALSE
, (GLfloat
*)values
);
2794 loc
+= element_type
->matrix_columns
;
2796 _mesa_uniform(ctx
, shader_program
, loc
, element_type
->matrix_columns
,
2797 values
, element_type
->gl_type
);
2798 loc
+= type_size(element_type
);
2804 set_uniform_initializers(struct gl_context
*ctx
,
2805 struct gl_shader_program
*shader_program
)
2807 void *mem_ctx
= NULL
;
2809 for (unsigned int i
= 0; i
< MESA_SHADER_TYPES
; i
++) {
2810 struct gl_shader
*shader
= shader_program
->_LinkedShaders
[i
];
2815 foreach_iter(exec_list_iterator
, iter
, *shader
->ir
) {
2816 ir_instruction
*ir
= (ir_instruction
*)iter
.get();
2817 ir_variable
*var
= ir
->as_variable();
2819 if (!var
|| var
->mode
!= ir_var_uniform
|| !var
->constant_value
)
2823 mem_ctx
= ralloc_context(NULL
);
2825 set_uniform_initializer(ctx
, mem_ctx
, shader_program
, var
->name
,
2826 var
->type
, var
->constant_value
);
2830 ralloc_free(mem_ctx
);
2834 * Scan/rewrite program to remove reads of custom (output) registers.
2835 * The passed type has to be either PROGRAM_OUTPUT or PROGRAM_VARYING
2836 * (for vertex shaders).
2837 * In GLSL shaders, varying vars can be read and written.
2838 * On some hardware, trying to read an output register causes trouble.
2839 * So, rewrite the program to use a temporary register in this case.
2841 * Based on _mesa_remove_output_reads from programopt.c.
2844 glsl_to_tgsi_visitor::remove_output_reads(gl_register_file type
)
2847 GLint outputMap
[VERT_RESULT_MAX
];
2848 GLint outputTypes
[VERT_RESULT_MAX
];
2849 GLuint numVaryingReads
= 0;
2850 GLboolean usedTemps
[MAX_TEMPS
];
2851 GLuint firstTemp
= 0;
2853 _mesa_find_used_registers(prog
, PROGRAM_TEMPORARY
,
2854 usedTemps
, MAX_TEMPS
);
2856 assert(type
== PROGRAM_VARYING
|| type
== PROGRAM_OUTPUT
);
2857 assert(prog
->Target
== GL_VERTEX_PROGRAM_ARB
|| type
!= PROGRAM_VARYING
);
2859 for (i
= 0; i
< VERT_RESULT_MAX
; i
++)
2862 /* look for instructions which read from varying vars */
2863 foreach_iter(exec_list_iterator
, iter
, this->instructions
) {
2864 glsl_to_tgsi_instruction
*inst
= (glsl_to_tgsi_instruction
*)iter
.get();
2865 const GLuint numSrc
= num_inst_src_regs(inst
->op
);
2867 for (j
= 0; j
< numSrc
; j
++) {
2868 if (inst
->src
[j
].file
== type
) {
2869 /* replace the read with a temp reg */
2870 const GLuint var
= inst
->src
[j
].index
;
2871 if (outputMap
[var
] == -1) {
2873 outputMap
[var
] = _mesa_find_free_register(usedTemps
,
2876 outputTypes
[var
] = inst
->src
[j
].type
;
2877 firstTemp
= outputMap
[var
] + 1;
2879 inst
->src
[j
].file
= PROGRAM_TEMPORARY
;
2880 inst
->src
[j
].index
= outputMap
[var
];
2885 if (numVaryingReads
== 0)
2886 return; /* nothing to be done */
2888 /* look for instructions which write to the varying vars identified above */
2889 foreach_iter(exec_list_iterator
, iter
, this->instructions
) {
2890 glsl_to_tgsi_instruction
*inst
= (glsl_to_tgsi_instruction
*)iter
.get();
2891 if (inst
->dst
.file
== type
&& outputMap
[inst
->dst
.index
] >= 0) {
2892 /* change inst to write to the temp reg, instead of the varying */
2893 inst
->dst
.file
= PROGRAM_TEMPORARY
;
2894 inst
->dst
.index
= outputMap
[inst
->dst
.index
];
2898 /* insert new MOV instructions at the end */
2899 for (i
= 0; i
< VERT_RESULT_MAX
; i
++) {
2900 if (outputMap
[i
] >= 0) {
2901 /* MOV VAR[i], TEMP[tmp]; */
2902 st_src_reg src
= st_src_reg(PROGRAM_TEMPORARY
, outputMap
[i
], outputTypes
[i
]);
2903 st_dst_reg dst
= st_dst_reg(type
, WRITEMASK_XYZW
, outputTypes
[i
]);
2905 this->emit(NULL
, TGSI_OPCODE_MOV
, dst
, src
);
2911 * Returns the mask of channels (bitmask of WRITEMASK_X,Y,Z,W) which
2912 * are read from the given src in this instruction
2915 get_src_arg_mask(st_dst_reg dst
, st_src_reg src
)
2917 int read_mask
= 0, comp
;
2919 /* Now, given the src swizzle and the written channels, find which
2920 * components are actually read
2922 for (comp
= 0; comp
< 4; ++comp
) {
2923 const unsigned coord
= GET_SWZ(src
.swizzle
, comp
);
2925 if (dst
.writemask
& (1 << comp
) && coord
<= SWIZZLE_W
)
2926 read_mask
|= 1 << coord
;
2933 * This pass replaces CMP T0, T1 T2 T0 with MOV T0, T2 when the CMP
2934 * instruction is the first instruction to write to register T0. There are
2935 * several lowering passes done in GLSL IR (e.g. branches and
2936 * relative addressing) that create a large number of conditional assignments
2937 * that ir_to_mesa converts to CMP instructions like the one mentioned above.
2939 * Here is why this conversion is safe:
2940 * CMP T0, T1 T2 T0 can be expanded to:
2946 * If (T1 < 0.0) evaluates to true then our replacement MOV T0, T2 is the same
2947 * as the original program. If (T1 < 0.0) evaluates to false, executing
2948 * MOV T0, T0 will store a garbage value in T0 since T0 is uninitialized.
2949 * Therefore, it doesn't matter that we are replacing MOV T0, T0 with MOV T0, T2
2950 * because any instruction that was going to read from T0 after this was going
2951 * to read a garbage value anyway.
2954 glsl_to_tgsi_visitor::simplify_cmp(void)
2956 unsigned tempWrites
[MAX_TEMPS
];
2957 unsigned outputWrites
[MAX_PROGRAM_OUTPUTS
];
2959 memset(tempWrites
, 0, sizeof(tempWrites
));
2960 memset(outputWrites
, 0, sizeof(outputWrites
));
2962 foreach_iter(exec_list_iterator
, iter
, this->instructions
) {
2963 glsl_to_tgsi_instruction
*inst
= (glsl_to_tgsi_instruction
*)iter
.get();
2964 unsigned prevWriteMask
= 0;
2966 /* Give up if we encounter relative addressing or flow control. */
2967 if (inst
->dst
.reladdr
||
2968 tgsi_get_opcode_info(inst
->op
)->is_branch
||
2969 inst
->op
== TGSI_OPCODE_BGNSUB
||
2970 inst
->op
== TGSI_OPCODE_CONT
||
2971 inst
->op
== TGSI_OPCODE_END
||
2972 inst
->op
== TGSI_OPCODE_ENDSUB
||
2973 inst
->op
== TGSI_OPCODE_RET
) {
2977 if (inst
->dst
.file
== PROGRAM_OUTPUT
) {
2978 assert(inst
->dst
.index
< MAX_PROGRAM_OUTPUTS
);
2979 prevWriteMask
= outputWrites
[inst
->dst
.index
];
2980 outputWrites
[inst
->dst
.index
] |= inst
->dst
.writemask
;
2981 } else if (inst
->dst
.file
== PROGRAM_TEMPORARY
) {
2982 assert(inst
->dst
.index
< MAX_TEMPS
);
2983 prevWriteMask
= tempWrites
[inst
->dst
.index
];
2984 tempWrites
[inst
->dst
.index
] |= inst
->dst
.writemask
;
2987 /* For a CMP to be considered a conditional write, the destination
2988 * register and source register two must be the same. */
2989 if (inst
->op
== TGSI_OPCODE_CMP
2990 && !(inst
->dst
.writemask
& prevWriteMask
)
2991 && inst
->src
[2].file
== inst
->dst
.file
2992 && inst
->src
[2].index
== inst
->dst
.index
2993 && inst
->dst
.writemask
== get_src_arg_mask(inst
->dst
, inst
->src
[2])) {
2995 inst
->op
= TGSI_OPCODE_MOV
;
2996 inst
->src
[0] = inst
->src
[1];
3001 /* Replaces all references to a temporary register index with another index. */
3003 glsl_to_tgsi_visitor::rename_temp_register(int index
, int new_index
)
3005 foreach_iter(exec_list_iterator
, iter
, this->instructions
) {
3006 glsl_to_tgsi_instruction
*inst
= (glsl_to_tgsi_instruction
*)iter
.get();
3009 for (j
=0; j
< num_inst_src_regs(inst
->op
); j
++) {
3010 if (inst
->src
[j
].file
== PROGRAM_TEMPORARY
&&
3011 inst
->src
[j
].index
== index
) {
3012 inst
->src
[j
].index
= new_index
;
3016 if (inst
->dst
.file
== PROGRAM_TEMPORARY
&& inst
->dst
.index
== index
) {
3017 inst
->dst
.index
= new_index
;
3023 glsl_to_tgsi_visitor::get_first_temp_read(int index
)
3025 int depth
= 0; /* loop depth */
3026 int loop_start
= -1; /* index of the first active BGNLOOP (if any) */
3029 foreach_iter(exec_list_iterator
, iter
, this->instructions
) {
3030 glsl_to_tgsi_instruction
*inst
= (glsl_to_tgsi_instruction
*)iter
.get();
3032 for (j
=0; j
< num_inst_src_regs(inst
->op
); j
++) {
3033 if (inst
->src
[j
].file
== PROGRAM_TEMPORARY
&&
3034 inst
->src
[j
].index
== index
) {
3035 return (depth
== 0) ? i
: loop_start
;
3039 if (inst
->op
== TGSI_OPCODE_BGNLOOP
) {
3042 } else if (inst
->op
== TGSI_OPCODE_ENDLOOP
) {
3055 glsl_to_tgsi_visitor::get_first_temp_write(int index
)
3057 int depth
= 0; /* loop depth */
3058 int loop_start
= -1; /* index of the first active BGNLOOP (if any) */
3061 foreach_iter(exec_list_iterator
, iter
, this->instructions
) {
3062 glsl_to_tgsi_instruction
*inst
= (glsl_to_tgsi_instruction
*)iter
.get();
3064 if (inst
->dst
.file
== PROGRAM_TEMPORARY
&& inst
->dst
.index
== index
) {
3065 return (depth
== 0) ? i
: loop_start
;
3068 if (inst
->op
== TGSI_OPCODE_BGNLOOP
) {
3071 } else if (inst
->op
== TGSI_OPCODE_ENDLOOP
) {
3084 glsl_to_tgsi_visitor::get_last_temp_read(int index
)
3086 int depth
= 0; /* loop depth */
3087 int last
= -1; /* index of last instruction that reads the temporary */
3090 foreach_iter(exec_list_iterator
, iter
, this->instructions
) {
3091 glsl_to_tgsi_instruction
*inst
= (glsl_to_tgsi_instruction
*)iter
.get();
3093 for (j
=0; j
< num_inst_src_regs(inst
->op
); j
++) {
3094 if (inst
->src
[j
].file
== PROGRAM_TEMPORARY
&&
3095 inst
->src
[j
].index
== index
) {
3096 last
= (depth
== 0) ? i
: -2;
3100 if (inst
->op
== TGSI_OPCODE_BGNLOOP
)
3102 else if (inst
->op
== TGSI_OPCODE_ENDLOOP
)
3103 if (--depth
== 0 && last
== -2)
3115 glsl_to_tgsi_visitor::get_last_temp_write(int index
)
3117 int depth
= 0; /* loop depth */
3118 int last
= -1; /* index of last instruction that writes to the temporary */
3121 foreach_iter(exec_list_iterator
, iter
, this->instructions
) {
3122 glsl_to_tgsi_instruction
*inst
= (glsl_to_tgsi_instruction
*)iter
.get();
3124 if (inst
->dst
.file
== PROGRAM_TEMPORARY
&& inst
->dst
.index
== index
)
3125 last
= (depth
== 0) ? i
: -2;
3127 if (inst
->op
== TGSI_OPCODE_BGNLOOP
)
3129 else if (inst
->op
== TGSI_OPCODE_ENDLOOP
)
3130 if (--depth
== 0 && last
== -2)
3142 * On a basic block basis, tracks available PROGRAM_TEMPORARY register
3143 * channels for copy propagation and updates following instructions to
3144 * use the original versions.
3146 * The glsl_to_tgsi_visitor lazily produces code assuming that this pass
3147 * will occur. As an example, a TXP production before this pass:
3149 * 0: MOV TEMP[1], INPUT[4].xyyy;
3150 * 1: MOV TEMP[1].w, INPUT[4].wwww;
3151 * 2: TXP TEMP[2], TEMP[1], texture[0], 2D;
3155 * 0: MOV TEMP[1], INPUT[4].xyyy;
3156 * 1: MOV TEMP[1].w, INPUT[4].wwww;
3157 * 2: TXP TEMP[2], INPUT[4].xyyw, texture[0], 2D;
3159 * which allows for dead code elimination on TEMP[1]'s writes.
3162 glsl_to_tgsi_visitor::copy_propagate(void)
3164 glsl_to_tgsi_instruction
**acp
= rzalloc_array(mem_ctx
,
3165 glsl_to_tgsi_instruction
*,
3166 this->next_temp
* 4);
3167 int *acp_level
= rzalloc_array(mem_ctx
, int, this->next_temp
* 4);
3170 foreach_iter(exec_list_iterator
, iter
, this->instructions
) {
3171 glsl_to_tgsi_instruction
*inst
= (glsl_to_tgsi_instruction
*)iter
.get();
3173 assert(inst
->dst
.file
!= PROGRAM_TEMPORARY
3174 || inst
->dst
.index
< this->next_temp
);
3176 /* First, do any copy propagation possible into the src regs. */
3177 for (int r
= 0; r
< 3; r
++) {
3178 glsl_to_tgsi_instruction
*first
= NULL
;
3180 int acp_base
= inst
->src
[r
].index
* 4;
3182 if (inst
->src
[r
].file
!= PROGRAM_TEMPORARY
||
3183 inst
->src
[r
].reladdr
)
3186 /* See if we can find entries in the ACP consisting of MOVs
3187 * from the same src register for all the swizzled channels
3188 * of this src register reference.
3190 for (int i
= 0; i
< 4; i
++) {
3191 int src_chan
= GET_SWZ(inst
->src
[r
].swizzle
, i
);
3192 glsl_to_tgsi_instruction
*copy_chan
= acp
[acp_base
+ src_chan
];
3199 assert(acp_level
[acp_base
+ src_chan
] <= level
);
3204 if (first
->src
[0].file
!= copy_chan
->src
[0].file
||
3205 first
->src
[0].index
!= copy_chan
->src
[0].index
) {
3213 /* We've now validated that we can copy-propagate to
3214 * replace this src register reference. Do it.
3216 inst
->src
[r
].file
= first
->src
[0].file
;
3217 inst
->src
[r
].index
= first
->src
[0].index
;
3220 for (int i
= 0; i
< 4; i
++) {
3221 int src_chan
= GET_SWZ(inst
->src
[r
].swizzle
, i
);
3222 glsl_to_tgsi_instruction
*copy_inst
= acp
[acp_base
+ src_chan
];
3223 swizzle
|= (GET_SWZ(copy_inst
->src
[0].swizzle
, src_chan
) <<
3226 inst
->src
[r
].swizzle
= swizzle
;
3231 case TGSI_OPCODE_BGNLOOP
:
3232 case TGSI_OPCODE_ENDLOOP
:
3233 /* End of a basic block, clear the ACP entirely. */
3234 memset(acp
, 0, sizeof(*acp
) * this->next_temp
* 4);
3237 case TGSI_OPCODE_IF
:
3241 case TGSI_OPCODE_ENDIF
:
3242 case TGSI_OPCODE_ELSE
:
3243 /* Clear all channels written inside the block from the ACP, but
3244 * leaving those that were not touched.
3246 for (int r
= 0; r
< this->next_temp
; r
++) {
3247 for (int c
= 0; c
< 4; c
++) {
3248 if (!acp
[4 * r
+ c
])
3251 if (acp_level
[4 * r
+ c
] >= level
)
3252 acp
[4 * r
+ c
] = NULL
;
3255 if (inst
->op
== TGSI_OPCODE_ENDIF
)
3260 /* Continuing the block, clear any written channels from
3263 if (inst
->dst
.file
== PROGRAM_TEMPORARY
&& inst
->dst
.reladdr
) {
3264 /* Any temporary might be written, so no copy propagation
3265 * across this instruction.
3267 memset(acp
, 0, sizeof(*acp
) * this->next_temp
* 4);
3268 } else if (inst
->dst
.file
== PROGRAM_OUTPUT
&&
3269 inst
->dst
.reladdr
) {
3270 /* Any output might be written, so no copy propagation
3271 * from outputs across this instruction.
3273 for (int r
= 0; r
< this->next_temp
; r
++) {
3274 for (int c
= 0; c
< 4; c
++) {
3275 if (!acp
[4 * r
+ c
])
3278 if (acp
[4 * r
+ c
]->src
[0].file
== PROGRAM_OUTPUT
)
3279 acp
[4 * r
+ c
] = NULL
;
3282 } else if (inst
->dst
.file
== PROGRAM_TEMPORARY
||
3283 inst
->dst
.file
== PROGRAM_OUTPUT
) {
3284 /* Clear where it's used as dst. */
3285 if (inst
->dst
.file
== PROGRAM_TEMPORARY
) {
3286 for (int c
= 0; c
< 4; c
++) {
3287 if (inst
->dst
.writemask
& (1 << c
)) {
3288 acp
[4 * inst
->dst
.index
+ c
] = NULL
;
3293 /* Clear where it's used as src. */
3294 for (int r
= 0; r
< this->next_temp
; r
++) {
3295 for (int c
= 0; c
< 4; c
++) {
3296 if (!acp
[4 * r
+ c
])
3299 int src_chan
= GET_SWZ(acp
[4 * r
+ c
]->src
[0].swizzle
, c
);
3301 if (acp
[4 * r
+ c
]->src
[0].file
== inst
->dst
.file
&&
3302 acp
[4 * r
+ c
]->src
[0].index
== inst
->dst
.index
&&
3303 inst
->dst
.writemask
& (1 << src_chan
))
3305 acp
[4 * r
+ c
] = NULL
;
3313 /* If this is a copy, add it to the ACP. */
3314 if (inst
->op
== TGSI_OPCODE_MOV
&&
3315 inst
->dst
.file
== PROGRAM_TEMPORARY
&&
3316 !inst
->dst
.reladdr
&&
3318 !inst
->src
[0].reladdr
&&
3319 !inst
->src
[0].negate
) {
3320 for (int i
= 0; i
< 4; i
++) {
3321 if (inst
->dst
.writemask
& (1 << i
)) {
3322 acp
[4 * inst
->dst
.index
+ i
] = inst
;
3323 acp_level
[4 * inst
->dst
.index
+ i
] = level
;
3329 ralloc_free(acp_level
);
3334 * Tracks available PROGRAM_TEMPORARY registers for dead code elimination.
3336 * The glsl_to_tgsi_visitor lazily produces code assuming that this pass
3337 * will occur. As an example, a TXP production after copy propagation but
3340 * 0: MOV TEMP[1], INPUT[4].xyyy;
3341 * 1: MOV TEMP[1].w, INPUT[4].wwww;
3342 * 2: TXP TEMP[2], INPUT[4].xyyw, texture[0], 2D;
3344 * and after this pass:
3346 * 0: TXP TEMP[2], INPUT[4].xyyw, texture[0], 2D;
3348 * FIXME: assumes that all functions are inlined (no support for BGNSUB/ENDSUB)
3349 * FIXME: doesn't eliminate all dead code inside of loops; it steps around them
3352 glsl_to_tgsi_visitor::eliminate_dead_code(void)
3356 for (i
=0; i
< this->next_temp
; i
++) {
3357 int last_read
= get_last_temp_read(i
);
3360 foreach_iter(exec_list_iterator
, iter
, this->instructions
) {
3361 glsl_to_tgsi_instruction
*inst
= (glsl_to_tgsi_instruction
*)iter
.get();
3363 if (inst
->dst
.file
== PROGRAM_TEMPORARY
&& inst
->dst
.index
== i
&&
3376 * On a basic block basis, tracks available PROGRAM_TEMPORARY registers for dead
3377 * code elimination. This is less primitive than eliminate_dead_code(), as it
3378 * is per-channel and can detect consecutive writes without a read between them
3379 * as dead code. However, there is some dead code that can be eliminated by
3380 * eliminate_dead_code() but not this function - for example, this function
3381 * cannot eliminate an instruction writing to a register that is never read and
3382 * is the only instruction writing to that register.
3384 * The glsl_to_tgsi_visitor lazily produces code assuming that this pass
3388 glsl_to_tgsi_visitor::eliminate_dead_code_advanced(void)
3390 glsl_to_tgsi_instruction
**writes
= rzalloc_array(mem_ctx
,
3391 glsl_to_tgsi_instruction
*,
3392 this->next_temp
* 4);
3393 int *write_level
= rzalloc_array(mem_ctx
, int, this->next_temp
* 4);
3397 foreach_iter(exec_list_iterator
, iter
, this->instructions
) {
3398 glsl_to_tgsi_instruction
*inst
= (glsl_to_tgsi_instruction
*)iter
.get();
3400 assert(inst
->dst
.file
!= PROGRAM_TEMPORARY
3401 || inst
->dst
.index
< this->next_temp
);
3404 case TGSI_OPCODE_BGNLOOP
:
3405 case TGSI_OPCODE_ENDLOOP
:
3406 /* End of a basic block, clear the write array entirely.
3407 * FIXME: This keeps us from killing dead code when the writes are
3408 * on either side of a loop, even when the register isn't touched
3411 memset(writes
, 0, sizeof(*writes
) * this->next_temp
* 4);
3414 case TGSI_OPCODE_ENDIF
:
3418 case TGSI_OPCODE_ELSE
:
3419 /* Clear all channels written inside the preceding if block from the
3420 * write array, but leave those that were not touched.
3422 * FIXME: This destroys opportunities to remove dead code inside of
3423 * IF blocks that are followed by an ELSE block.
3425 for (int r
= 0; r
< this->next_temp
; r
++) {
3426 for (int c
= 0; c
< 4; c
++) {
3427 if (!writes
[4 * r
+ c
])
3430 if (write_level
[4 * r
+ c
] >= level
)
3431 writes
[4 * r
+ c
] = NULL
;
3436 case TGSI_OPCODE_IF
:
3438 /* fallthrough to default case to mark the condition as read */
3441 /* Continuing the block, clear any channels from the write array that
3442 * are read by this instruction.
3444 for (int i
= 0; i
< 4; i
++) {
3445 if (inst
->src
[i
].file
== PROGRAM_TEMPORARY
&& inst
->src
[i
].reladdr
){
3446 /* Any temporary might be read, so no dead code elimination
3447 * across this instruction.
3449 memset(writes
, 0, sizeof(*writes
) * this->next_temp
* 4);
3450 } else if (inst
->src
[i
].file
== PROGRAM_TEMPORARY
) {
3451 /* Clear where it's used as src. */
3452 int src_chans
= 1 << GET_SWZ(inst
->src
[i
].swizzle
, 0);
3453 src_chans
|= 1 << GET_SWZ(inst
->src
[i
].swizzle
, 1);
3454 src_chans
|= 1 << GET_SWZ(inst
->src
[i
].swizzle
, 2);
3455 src_chans
|= 1 << GET_SWZ(inst
->src
[i
].swizzle
, 3);
3457 for (int c
= 0; c
< 4; c
++) {
3458 if (src_chans
& (1 << c
)) {
3459 writes
[4 * inst
->src
[i
].index
+ c
] = NULL
;
3467 /* If this instruction writes to a temporary, add it to the write array.
3468 * If there is already an instruction in the write array for one or more
3469 * of the channels, flag that channel write as dead.
3471 if (inst
->dst
.file
== PROGRAM_TEMPORARY
&&
3472 !inst
->dst
.reladdr
&&
3474 for (int c
= 0; c
< 4; c
++) {
3475 if (inst
->dst
.writemask
& (1 << c
)) {
3476 if (writes
[4 * inst
->dst
.index
+ c
]) {
3477 if (write_level
[4 * inst
->dst
.index
+ c
] < level
)
3480 writes
[4 * inst
->dst
.index
+ c
]->dead_mask
|= (1 << c
);
3482 writes
[4 * inst
->dst
.index
+ c
] = inst
;
3483 write_level
[4 * inst
->dst
.index
+ c
] = level
;
3489 /* Anything still in the write array at this point is dead code. */
3490 for (int r
= 0; r
< this->next_temp
; r
++) {
3491 for (int c
= 0; c
< 4; c
++) {
3492 glsl_to_tgsi_instruction
*inst
= writes
[4 * r
+ c
];
3494 inst
->dead_mask
|= (1 << c
);
3498 /* Now actually remove the instructions that are completely dead and update
3499 * the writemask of other instructions with dead channels.
3501 foreach_iter(exec_list_iterator
, iter
, this->instructions
) {
3502 glsl_to_tgsi_instruction
*inst
= (glsl_to_tgsi_instruction
*)iter
.get();
3504 if (!inst
->dead_mask
|| !inst
->dst
.writemask
)
3506 else if (inst
->dead_mask
== inst
->dst
.writemask
) {
3511 inst
->dst
.writemask
&= ~(inst
->dead_mask
);
3514 ralloc_free(write_level
);
3515 ralloc_free(writes
);
3520 /* Merges temporary registers together where possible to reduce the number of
3521 * registers needed to run a program.
3523 * Produces optimal code only after copy propagation and dead code elimination
3526 glsl_to_tgsi_visitor::merge_registers(void)
3528 int *last_reads
= rzalloc_array(mem_ctx
, int, this->next_temp
);
3529 int *first_writes
= rzalloc_array(mem_ctx
, int, this->next_temp
);
3532 /* Read the indices of the last read and first write to each temp register
3533 * into an array so that we don't have to traverse the instruction list as
3535 for (i
=0; i
< this->next_temp
; i
++) {
3536 last_reads
[i
] = get_last_temp_read(i
);
3537 first_writes
[i
] = get_first_temp_write(i
);
3540 /* Start looking for registers with non-overlapping usages that can be
3541 * merged together. */
3542 for (i
=0; i
< this->next_temp
; i
++) {
3543 /* Don't touch unused registers. */
3544 if (last_reads
[i
] < 0 || first_writes
[i
] < 0) continue;
3546 for (j
=0; j
< this->next_temp
; j
++) {
3547 /* Don't touch unused registers. */
3548 if (last_reads
[j
] < 0 || first_writes
[j
] < 0) continue;
3550 /* We can merge the two registers if the first write to j is after or
3551 * in the same instruction as the last read from i. Note that the
3552 * register at index i will always be used earlier or at the same time
3553 * as the register at index j. */
3554 if (first_writes
[i
] <= first_writes
[j
] &&
3555 last_reads
[i
] <= first_writes
[j
])
3557 rename_temp_register(j
, i
); /* Replace all references to j with i.*/
3559 /* Update the first_writes and last_reads arrays with the new
3560 * values for the merged register index, and mark the newly unused
3561 * register index as such. */
3562 last_reads
[i
] = last_reads
[j
];
3563 first_writes
[j
] = -1;
3569 ralloc_free(last_reads
);
3570 ralloc_free(first_writes
);
3573 /* Reassign indices to temporary registers by reusing unused indices created
3574 * by optimization passes. */
3576 glsl_to_tgsi_visitor::renumber_registers(void)
3581 for (i
=0; i
< this->next_temp
; i
++) {
3582 if (get_first_temp_read(i
) < 0) continue;
3584 rename_temp_register(i
, new_index
);
3588 this->next_temp
= new_index
;
3592 * Returns a fragment program which implements the current pixel transfer ops.
3593 * Based on get_pixel_transfer_program in st_atom_pixeltransfer.c.
3596 get_pixel_transfer_visitor(struct st_fragment_program
*fp
,
3597 glsl_to_tgsi_visitor
*original
,
3598 int scale_and_bias
, int pixel_maps
)
3600 glsl_to_tgsi_visitor
*v
= new glsl_to_tgsi_visitor();
3601 struct st_context
*st
= st_context(original
->ctx
);
3602 struct gl_program
*prog
= &fp
->Base
.Base
;
3603 struct gl_program_parameter_list
*params
= _mesa_new_parameter_list();
3604 st_src_reg coord
, src0
;
3606 glsl_to_tgsi_instruction
*inst
;
3608 /* Copy attributes of the glsl_to_tgsi_visitor in the original shader. */
3609 v
->ctx
= original
->ctx
;
3611 v
->glsl_version
= original
->glsl_version
;
3612 v
->options
= original
->options
;
3613 v
->next_temp
= original
->next_temp
;
3614 v
->num_address_regs
= original
->num_address_regs
;
3615 v
->samplers_used
= prog
->SamplersUsed
= original
->samplers_used
;
3616 v
->indirect_addr_temps
= original
->indirect_addr_temps
;
3617 v
->indirect_addr_consts
= original
->indirect_addr_consts
;
3618 memcpy(&v
->immediates
, &original
->immediates
, sizeof(v
->immediates
));
3621 * Get initial pixel color from the texture.
3622 * TEX colorTemp, fragment.texcoord[0], texture[0], 2D;
3624 coord
= st_src_reg(PROGRAM_INPUT
, FRAG_ATTRIB_TEX0
, glsl_type::vec2_type
);
3625 src0
= v
->get_temp(glsl_type::vec4_type
);
3626 dst0
= st_dst_reg(src0
);
3627 inst
= v
->emit(NULL
, TGSI_OPCODE_TEX
, dst0
, coord
);
3629 inst
->tex_target
= TEXTURE_2D_INDEX
;
3631 prog
->InputsRead
|= (1 << FRAG_ATTRIB_TEX0
);
3632 prog
->SamplersUsed
|= (1 << 0); /* mark sampler 0 as used */
3633 v
->samplers_used
|= (1 << 0);
3635 if (scale_and_bias
) {
3636 static const gl_state_index scale_state
[STATE_LENGTH
] =
3637 { STATE_INTERNAL
, STATE_PT_SCALE
,
3638 (gl_state_index
) 0, (gl_state_index
) 0, (gl_state_index
) 0 };
3639 static const gl_state_index bias_state
[STATE_LENGTH
] =
3640 { STATE_INTERNAL
, STATE_PT_BIAS
,
3641 (gl_state_index
) 0, (gl_state_index
) 0, (gl_state_index
) 0 };
3642 GLint scale_p
, bias_p
;
3643 st_src_reg scale
, bias
;
3645 scale_p
= _mesa_add_state_reference(params
, scale_state
);
3646 bias_p
= _mesa_add_state_reference(params
, bias_state
);
3648 /* MAD colorTemp, colorTemp, scale, bias; */
3649 scale
= st_src_reg(PROGRAM_STATE_VAR
, scale_p
, GLSL_TYPE_FLOAT
);
3650 bias
= st_src_reg(PROGRAM_STATE_VAR
, bias_p
, GLSL_TYPE_FLOAT
);
3651 inst
= v
->emit(NULL
, TGSI_OPCODE_MAD
, dst0
, src0
, scale
, bias
);
3655 st_src_reg temp
= v
->get_temp(glsl_type::vec4_type
);
3656 st_dst_reg temp_dst
= st_dst_reg(temp
);
3658 assert(st
->pixel_xfer
.pixelmap_texture
);
3660 /* With a little effort, we can do four pixel map look-ups with
3661 * two TEX instructions:
3664 /* TEX temp.rg, colorTemp.rgba, texture[1], 2D; */
3665 temp_dst
.writemask
= WRITEMASK_XY
; /* write R,G */
3666 inst
= v
->emit(NULL
, TGSI_OPCODE_TEX
, temp_dst
, src0
);
3668 inst
->tex_target
= TEXTURE_2D_INDEX
;
3670 /* TEX temp.ba, colorTemp.baba, texture[1], 2D; */
3671 src0
.swizzle
= MAKE_SWIZZLE4(SWIZZLE_Z
, SWIZZLE_W
, SWIZZLE_Z
, SWIZZLE_W
);
3672 temp_dst
.writemask
= WRITEMASK_ZW
; /* write B,A */
3673 inst
= v
->emit(NULL
, TGSI_OPCODE_TEX
, temp_dst
, src0
);
3675 inst
->tex_target
= TEXTURE_2D_INDEX
;
3677 prog
->SamplersUsed
|= (1 << 1); /* mark sampler 1 as used */
3678 v
->samplers_used
|= (1 << 1);
3680 /* MOV colorTemp, temp; */
3681 inst
= v
->emit(NULL
, TGSI_OPCODE_MOV
, dst0
, temp
);
3684 /* Now copy the instructions from the original glsl_to_tgsi_visitor into the
3686 foreach_iter(exec_list_iterator
, iter
, original
->instructions
) {
3687 glsl_to_tgsi_instruction
*inst
= (glsl_to_tgsi_instruction
*)iter
.get();
3688 st_src_reg src_regs
[3];
3690 if (inst
->dst
.file
== PROGRAM_OUTPUT
)
3691 prog
->OutputsWritten
|= BITFIELD64_BIT(inst
->dst
.index
);
3693 for (int i
=0; i
<3; i
++) {
3694 src_regs
[i
] = inst
->src
[i
];
3695 if (src_regs
[i
].file
== PROGRAM_INPUT
&&
3696 src_regs
[i
].index
== FRAG_ATTRIB_COL0
)
3698 src_regs
[i
].file
= PROGRAM_TEMPORARY
;
3699 src_regs
[i
].index
= src0
.index
;
3701 else if (src_regs
[i
].file
== PROGRAM_INPUT
)
3702 prog
->InputsRead
|= (1 << src_regs
[i
].index
);
3705 v
->emit(NULL
, inst
->op
, inst
->dst
, src_regs
[0], src_regs
[1], src_regs
[2]);
3708 /* Make modifications to fragment program info. */
3709 prog
->Parameters
= _mesa_combine_parameter_lists(params
,
3710 original
->prog
->Parameters
);
3711 prog
->Attributes
= _mesa_clone_parameter_list(original
->prog
->Attributes
);
3712 prog
->Varying
= _mesa_clone_parameter_list(original
->prog
->Varying
);
3713 _mesa_free_parameter_list(params
);
3714 count_resources(v
, prog
);
3715 fp
->glsl_to_tgsi
= v
;
3719 * Make fragment program for glBitmap:
3720 * Sample the texture and kill the fragment if the bit is 0.
3721 * This program will be combined with the user's fragment program.
3723 * Based on make_bitmap_fragment_program in st_cb_bitmap.c.
3726 get_bitmap_visitor(struct st_fragment_program
*fp
,
3727 glsl_to_tgsi_visitor
*original
, int samplerIndex
)
3729 glsl_to_tgsi_visitor
*v
= new glsl_to_tgsi_visitor();
3730 struct st_context
*st
= st_context(original
->ctx
);
3731 struct gl_program
*prog
= &fp
->Base
.Base
;
3732 st_src_reg coord
, src0
;
3734 glsl_to_tgsi_instruction
*inst
;
3736 /* Copy attributes of the glsl_to_tgsi_visitor in the original shader. */
3737 v
->ctx
= original
->ctx
;
3739 v
->glsl_version
= original
->glsl_version
;
3740 v
->options
= original
->options
;
3741 v
->next_temp
= original
->next_temp
;
3742 v
->num_address_regs
= original
->num_address_regs
;
3743 v
->samplers_used
= prog
->SamplersUsed
= original
->samplers_used
;
3744 v
->indirect_addr_temps
= original
->indirect_addr_temps
;
3745 v
->indirect_addr_consts
= original
->indirect_addr_consts
;
3746 memcpy(&v
->immediates
, &original
->immediates
, sizeof(v
->immediates
));
3748 /* TEX tmp0, fragment.texcoord[0], texture[0], 2D; */
3749 coord
= st_src_reg(PROGRAM_INPUT
, FRAG_ATTRIB_TEX0
, glsl_type::vec2_type
);
3750 src0
= v
->get_temp(glsl_type::vec4_type
);
3751 dst0
= st_dst_reg(src0
);
3752 inst
= v
->emit(NULL
, TGSI_OPCODE_TEX
, dst0
, coord
);
3753 inst
->sampler
= samplerIndex
;
3754 inst
->tex_target
= TEXTURE_2D_INDEX
;
3756 prog
->InputsRead
|= (1 << FRAG_ATTRIB_TEX0
);
3757 prog
->SamplersUsed
|= (1 << samplerIndex
); /* mark sampler as used */
3758 v
->samplers_used
|= (1 << samplerIndex
);
3760 /* KIL if -tmp0 < 0 # texel=0 -> keep / texel=0 -> discard */
3761 src0
.negate
= NEGATE_XYZW
;
3762 if (st
->bitmap
.tex_format
== PIPE_FORMAT_L8_UNORM
)
3763 src0
.swizzle
= SWIZZLE_XXXX
;
3764 inst
= v
->emit(NULL
, TGSI_OPCODE_KIL
, undef_dst
, src0
);
3766 /* Now copy the instructions from the original glsl_to_tgsi_visitor into the
3768 foreach_iter(exec_list_iterator
, iter
, original
->instructions
) {
3769 glsl_to_tgsi_instruction
*inst
= (glsl_to_tgsi_instruction
*)iter
.get();
3770 st_src_reg src_regs
[3];
3772 if (inst
->dst
.file
== PROGRAM_OUTPUT
)
3773 prog
->OutputsWritten
|= BITFIELD64_BIT(inst
->dst
.index
);
3775 for (int i
=0; i
<3; i
++) {
3776 src_regs
[i
] = inst
->src
[i
];
3777 if (src_regs
[i
].file
== PROGRAM_INPUT
)
3778 prog
->InputsRead
|= (1 << src_regs
[i
].index
);
3781 v
->emit(NULL
, inst
->op
, inst
->dst
, src_regs
[0], src_regs
[1], src_regs
[2]);
3784 /* Make modifications to fragment program info. */
3785 prog
->Parameters
= _mesa_clone_parameter_list(original
->prog
->Parameters
);
3786 prog
->Attributes
= _mesa_clone_parameter_list(original
->prog
->Attributes
);
3787 prog
->Varying
= _mesa_clone_parameter_list(original
->prog
->Varying
);
3788 count_resources(v
, prog
);
3789 fp
->glsl_to_tgsi
= v
;
3792 /* ------------------------- TGSI conversion stuff -------------------------- */
3794 unsigned branch_target
;
3799 * Intermediate state used during shader translation.
3801 struct st_translate
{
3802 struct ureg_program
*ureg
;
3804 struct ureg_dst temps
[MAX_TEMPS
];
3805 struct ureg_src
*constants
;
3806 struct ureg_src
*immediates
;
3807 struct ureg_dst outputs
[PIPE_MAX_SHADER_OUTPUTS
];
3808 struct ureg_src inputs
[PIPE_MAX_SHADER_INPUTS
];
3809 struct ureg_dst address
[1];
3810 struct ureg_src samplers
[PIPE_MAX_SAMPLERS
];
3811 struct ureg_src systemValues
[SYSTEM_VALUE_MAX
];
3813 /* Extra info for handling point size clamping in vertex shader */
3814 struct ureg_dst pointSizeResult
; /**< Actual point size output register */
3815 struct ureg_src pointSizeConst
; /**< Point size range constant register */
3816 GLint pointSizeOutIndex
; /**< Temp point size output register */
3817 GLboolean prevInstWrotePointSize
;
3819 const GLuint
*inputMapping
;
3820 const GLuint
*outputMapping
;
3822 /* For every instruction that contains a label (eg CALL), keep
3823 * details so that we can go back afterwards and emit the correct
3824 * tgsi instruction number for each label.
3826 struct label
*labels
;
3827 unsigned labels_size
;
3828 unsigned labels_count
;
3830 /* Keep a record of the tgsi instruction number that each mesa
3831 * instruction starts at, will be used to fix up labels after
3836 unsigned insn_count
;
3838 unsigned procType
; /**< TGSI_PROCESSOR_VERTEX/FRAGMENT */
3843 /** Map Mesa's SYSTEM_VALUE_x to TGSI_SEMANTIC_x */
3844 static unsigned mesa_sysval_to_semantic
[SYSTEM_VALUE_MAX
] = {
3846 TGSI_SEMANTIC_INSTANCEID
3850 * Make note of a branch to a label in the TGSI code.
3851 * After we've emitted all instructions, we'll go over the list
3852 * of labels built here and patch the TGSI code with the actual
3853 * location of each label.
3855 static unsigned *get_label(struct st_translate
*t
, unsigned branch_target
)
3859 if (t
->labels_count
+ 1 >= t
->labels_size
) {
3860 t
->labels_size
= 1 << (util_logbase2(t
->labels_size
) + 1);
3861 t
->labels
= (struct label
*)realloc(t
->labels
,
3862 t
->labels_size
* sizeof(struct label
));
3863 if (t
->labels
== NULL
) {
3864 static unsigned dummy
;
3870 i
= t
->labels_count
++;
3871 t
->labels
[i
].branch_target
= branch_target
;
3872 return &t
->labels
[i
].token
;
3876 * Called prior to emitting the TGSI code for each instruction.
3877 * Allocate additional space for instructions if needed.
3878 * Update the insn[] array so the next glsl_to_tgsi_instruction points to
3879 * the next TGSI instruction.
3881 static void set_insn_start(struct st_translate
*t
, unsigned start
)
3883 if (t
->insn_count
+ 1 >= t
->insn_size
) {
3884 t
->insn_size
= 1 << (util_logbase2(t
->insn_size
) + 1);
3885 t
->insn
= (unsigned *)realloc(t
->insn
, t
->insn_size
* sizeof(t
->insn
[0]));
3886 if (t
->insn
== NULL
) {
3892 t
->insn
[t
->insn_count
++] = start
;
3896 * Map a glsl_to_tgsi constant/immediate to a TGSI immediate.
3898 static struct ureg_src
3899 emit_immediate(struct st_translate
*t
,
3900 gl_constant_value values
[4],
3903 struct ureg_program
*ureg
= t
->ureg
;
3908 return ureg_DECL_immediate(ureg
, &values
[0].f
, size
);
3910 return ureg_DECL_immediate_int(ureg
, &values
[0].i
, size
);
3911 case GL_UNSIGNED_INT
:
3913 return ureg_DECL_immediate_uint(ureg
, &values
[0].u
, size
);
3915 assert(!"should not get here - type must be float, int, uint, or bool");
3916 return ureg_src_undef();
3921 * Map a glsl_to_tgsi dst register to a TGSI ureg_dst register.
3923 static struct ureg_dst
3924 dst_register(struct st_translate
*t
,
3925 gl_register_file file
,
3929 case PROGRAM_UNDEFINED
:
3930 return ureg_dst_undef();
3932 case PROGRAM_TEMPORARY
:
3933 if (ureg_dst_is_undef(t
->temps
[index
]))
3934 t
->temps
[index
] = ureg_DECL_temporary(t
->ureg
);
3936 return t
->temps
[index
];
3938 case PROGRAM_OUTPUT
:
3939 if (t
->procType
== TGSI_PROCESSOR_VERTEX
&& index
== VERT_RESULT_PSIZ
)
3940 t
->prevInstWrotePointSize
= GL_TRUE
;
3942 if (t
->procType
== TGSI_PROCESSOR_VERTEX
)
3943 assert(index
< VERT_RESULT_MAX
);
3944 else if (t
->procType
== TGSI_PROCESSOR_FRAGMENT
)
3945 assert(index
< FRAG_RESULT_MAX
);
3947 assert(index
< GEOM_RESULT_MAX
);
3949 assert(t
->outputMapping
[index
] < Elements(t
->outputs
));
3951 return t
->outputs
[t
->outputMapping
[index
]];
3953 case PROGRAM_ADDRESS
:
3954 return t
->address
[index
];
3957 assert(!"unknown dst register file");
3958 return ureg_dst_undef();
3963 * Map a glsl_to_tgsi src register to a TGSI ureg_src register.
3965 static struct ureg_src
3966 src_register(struct st_translate
*t
,
3967 gl_register_file file
,
3971 case PROGRAM_UNDEFINED
:
3972 return ureg_src_undef();
3974 case PROGRAM_TEMPORARY
:
3976 assert(index
< Elements(t
->temps
));
3977 if (ureg_dst_is_undef(t
->temps
[index
]))
3978 t
->temps
[index
] = ureg_DECL_temporary(t
->ureg
);
3979 return ureg_src(t
->temps
[index
]);
3981 case PROGRAM_NAMED_PARAM
:
3982 case PROGRAM_ENV_PARAM
:
3983 case PROGRAM_LOCAL_PARAM
:
3984 case PROGRAM_UNIFORM
:
3986 return t
->constants
[index
];
3987 case PROGRAM_STATE_VAR
:
3988 case PROGRAM_CONSTANT
: /* ie, immediate */
3990 return ureg_DECL_constant(t
->ureg
, 0);
3992 return t
->constants
[index
];
3994 case PROGRAM_IMMEDIATE
:
3995 return t
->immediates
[index
];
3998 assert(t
->inputMapping
[index
] < Elements(t
->inputs
));
3999 return t
->inputs
[t
->inputMapping
[index
]];
4001 case PROGRAM_OUTPUT
:
4002 assert(t
->outputMapping
[index
] < Elements(t
->outputs
));
4003 return ureg_src(t
->outputs
[t
->outputMapping
[index
]]); /* not needed? */
4005 case PROGRAM_ADDRESS
:
4006 return ureg_src(t
->address
[index
]);
4008 case PROGRAM_SYSTEM_VALUE
:
4009 assert(index
< Elements(t
->systemValues
));
4010 return t
->systemValues
[index
];
4013 assert(!"unknown src register file");
4014 return ureg_src_undef();
4019 * Create a TGSI ureg_dst register from an st_dst_reg.
4021 static struct ureg_dst
4022 translate_dst(struct st_translate
*t
,
4023 const st_dst_reg
*dst_reg
,
4026 struct ureg_dst dst
= dst_register(t
,
4030 dst
= ureg_writemask(dst
, dst_reg
->writemask
);
4033 dst
= ureg_saturate(dst
);
4035 if (dst_reg
->reladdr
!= NULL
)
4036 dst
= ureg_dst_indirect(dst
, ureg_src(t
->address
[0]));
4042 * Create a TGSI ureg_src register from an st_src_reg.
4044 static struct ureg_src
4045 translate_src(struct st_translate
*t
, const st_src_reg
*src_reg
)
4047 struct ureg_src src
= src_register(t
, src_reg
->file
, src_reg
->index
);
4049 src
= ureg_swizzle(src
,
4050 GET_SWZ(src_reg
->swizzle
, 0) & 0x3,
4051 GET_SWZ(src_reg
->swizzle
, 1) & 0x3,
4052 GET_SWZ(src_reg
->swizzle
, 2) & 0x3,
4053 GET_SWZ(src_reg
->swizzle
, 3) & 0x3);
4055 if ((src_reg
->negate
& 0xf) == NEGATE_XYZW
)
4056 src
= ureg_negate(src
);
4058 if (src_reg
->reladdr
!= NULL
) {
4059 /* Normally ureg_src_indirect() would be used here, but a stupid compiler
4060 * bug in g++ makes ureg_src_indirect (an inline C function) erroneously
4061 * set the bit for src.Negate. So we have to do the operation manually
4062 * here to work around the compiler's problems. */
4063 /*src = ureg_src_indirect(src, ureg_src(t->address[0]));*/
4064 struct ureg_src addr
= ureg_src(t
->address
[0]);
4066 src
.IndirectFile
= addr
.File
;
4067 src
.IndirectIndex
= addr
.Index
;
4068 src
.IndirectSwizzle
= addr
.SwizzleX
;
4070 if (src_reg
->file
!= PROGRAM_INPUT
&&
4071 src_reg
->file
!= PROGRAM_OUTPUT
) {
4072 /* If src_reg->index was negative, it was set to zero in
4073 * src_register(). Reassign it now. But don't do this
4074 * for input/output regs since they get remapped while
4075 * const buffers don't.
4077 src
.Index
= src_reg
->index
;
4085 compile_tgsi_instruction(struct st_translate
*t
,
4086 const struct glsl_to_tgsi_instruction
*inst
)
4088 struct ureg_program
*ureg
= t
->ureg
;
4090 struct ureg_dst dst
[1];
4091 struct ureg_src src
[4];
4095 num_dst
= num_inst_dst_regs(inst
->op
);
4096 num_src
= num_inst_src_regs(inst
->op
);
4099 dst
[0] = translate_dst(t
,
4103 for (i
= 0; i
< num_src
; i
++)
4104 src
[i
] = translate_src(t
, &inst
->src
[i
]);
4107 case TGSI_OPCODE_BGNLOOP
:
4108 case TGSI_OPCODE_CAL
:
4109 case TGSI_OPCODE_ELSE
:
4110 case TGSI_OPCODE_ENDLOOP
:
4111 case TGSI_OPCODE_IF
:
4112 assert(num_dst
== 0);
4113 ureg_label_insn(ureg
,
4117 inst
->op
== TGSI_OPCODE_CAL
? inst
->function
->sig_id
: 0));
4120 case TGSI_OPCODE_TEX
:
4121 case TGSI_OPCODE_TXB
:
4122 case TGSI_OPCODE_TXD
:
4123 case TGSI_OPCODE_TXL
:
4124 case TGSI_OPCODE_TXP
:
4125 src
[num_src
++] = t
->samplers
[inst
->sampler
];
4129 translate_texture_target(inst
->tex_target
, inst
->tex_shadow
),
4133 case TGSI_OPCODE_SCS
:
4134 dst
[0] = ureg_writemask(dst
[0], TGSI_WRITEMASK_XY
);
4135 ureg_insn(ureg
, inst
->op
, dst
, num_dst
, src
, num_src
);
4148 * Emit the TGSI instructions to adjust the WPOS pixel center convention
4149 * Basically, add (adjX, adjY) to the fragment position.
4152 emit_adjusted_wpos(struct st_translate
*t
,
4153 const struct gl_program
*program
,
4154 float adjX
, float adjY
)
4156 struct ureg_program
*ureg
= t
->ureg
;
4157 struct ureg_dst wpos_temp
= ureg_DECL_temporary(ureg
);
4158 struct ureg_src wpos_input
= t
->inputs
[t
->inputMapping
[FRAG_ATTRIB_WPOS
]];
4160 /* Note that we bias X and Y and pass Z and W through unchanged.
4161 * The shader might also use gl_FragCoord.w and .z.
4163 ureg_ADD(ureg
, wpos_temp
, wpos_input
,
4164 ureg_imm4f(ureg
, adjX
, adjY
, 0.0f
, 0.0f
));
4166 t
->inputs
[t
->inputMapping
[FRAG_ATTRIB_WPOS
]] = ureg_src(wpos_temp
);
4171 * Emit the TGSI instructions for inverting the WPOS y coordinate.
4172 * This code is unavoidable because it also depends on whether
4173 * a FBO is bound (STATE_FB_WPOS_Y_TRANSFORM).
4176 emit_wpos_inversion(struct st_translate
*t
,
4177 const struct gl_program
*program
,
4180 struct ureg_program
*ureg
= t
->ureg
;
4182 /* Fragment program uses fragment position input.
4183 * Need to replace instances of INPUT[WPOS] with temp T
4184 * where T = INPUT[WPOS] by y is inverted.
4186 static const gl_state_index wposTransformState
[STATE_LENGTH
]
4187 = { STATE_INTERNAL
, STATE_FB_WPOS_Y_TRANSFORM
,
4188 (gl_state_index
)0, (gl_state_index
)0, (gl_state_index
)0 };
4190 /* XXX: note we are modifying the incoming shader here! Need to
4191 * do this before emitting the constant decls below, or this
4194 unsigned wposTransConst
= _mesa_add_state_reference(program
->Parameters
,
4195 wposTransformState
);
4197 struct ureg_src wpostrans
= ureg_DECL_constant(ureg
, wposTransConst
);
4198 struct ureg_dst wpos_temp
;
4199 struct ureg_src wpos_input
= t
->inputs
[t
->inputMapping
[FRAG_ATTRIB_WPOS
]];
4201 /* MOV wpos_temp, input[wpos]
4203 if (wpos_input
.File
== TGSI_FILE_TEMPORARY
)
4204 wpos_temp
= ureg_dst(wpos_input
);
4206 wpos_temp
= ureg_DECL_temporary(ureg
);
4207 ureg_MOV(ureg
, wpos_temp
, wpos_input
);
4211 /* MAD wpos_temp.y, wpos_input, wpostrans.xxxx, wpostrans.yyyy
4214 ureg_writemask(wpos_temp
, TGSI_WRITEMASK_Y
),
4216 ureg_scalar(wpostrans
, 0),
4217 ureg_scalar(wpostrans
, 1));
4219 /* MAD wpos_temp.y, wpos_input, wpostrans.zzzz, wpostrans.wwww
4222 ureg_writemask(wpos_temp
, TGSI_WRITEMASK_Y
),
4224 ureg_scalar(wpostrans
, 2),
4225 ureg_scalar(wpostrans
, 3));
4228 /* Use wpos_temp as position input from here on:
4230 t
->inputs
[t
->inputMapping
[FRAG_ATTRIB_WPOS
]] = ureg_src(wpos_temp
);
4235 * Emit fragment position/ooordinate code.
4238 emit_wpos(struct st_context
*st
,
4239 struct st_translate
*t
,
4240 const struct gl_program
*program
,
4241 struct ureg_program
*ureg
)
4243 const struct gl_fragment_program
*fp
=
4244 (const struct gl_fragment_program
*) program
;
4245 struct pipe_screen
*pscreen
= st
->pipe
->screen
;
4246 boolean invert
= FALSE
;
4248 if (fp
->OriginUpperLeft
) {
4249 /* Fragment shader wants origin in upper-left */
4250 if (pscreen
->get_param(pscreen
, PIPE_CAP_TGSI_FS_COORD_ORIGIN_UPPER_LEFT
)) {
4251 /* the driver supports upper-left origin */
4253 else if (pscreen
->get_param(pscreen
, PIPE_CAP_TGSI_FS_COORD_ORIGIN_LOWER_LEFT
)) {
4254 /* the driver supports lower-left origin, need to invert Y */
4255 ureg_property_fs_coord_origin(ureg
, TGSI_FS_COORD_ORIGIN_LOWER_LEFT
);
4262 /* Fragment shader wants origin in lower-left */
4263 if (pscreen
->get_param(pscreen
, PIPE_CAP_TGSI_FS_COORD_ORIGIN_LOWER_LEFT
))
4264 /* the driver supports lower-left origin */
4265 ureg_property_fs_coord_origin(ureg
, TGSI_FS_COORD_ORIGIN_LOWER_LEFT
);
4266 else if (pscreen
->get_param(pscreen
, PIPE_CAP_TGSI_FS_COORD_ORIGIN_UPPER_LEFT
))
4267 /* the driver supports upper-left origin, need to invert Y */
4273 if (fp
->PixelCenterInteger
) {
4274 /* Fragment shader wants pixel center integer */
4275 if (pscreen
->get_param(pscreen
, PIPE_CAP_TGSI_FS_COORD_PIXEL_CENTER_INTEGER
))
4276 /* the driver supports pixel center integer */
4277 ureg_property_fs_coord_pixel_center(ureg
, TGSI_FS_COORD_PIXEL_CENTER_INTEGER
);
4278 else if (pscreen
->get_param(pscreen
, PIPE_CAP_TGSI_FS_COORD_PIXEL_CENTER_HALF_INTEGER
))
4279 /* the driver supports pixel center half integer, need to bias X,Y */
4280 emit_adjusted_wpos(t
, program
, 0.5f
, invert
? 0.5f
: -0.5f
);
4285 /* Fragment shader wants pixel center half integer */
4286 if (pscreen
->get_param(pscreen
, PIPE_CAP_TGSI_FS_COORD_PIXEL_CENTER_HALF_INTEGER
)) {
4287 /* the driver supports pixel center half integer */
4289 else if (pscreen
->get_param(pscreen
, PIPE_CAP_TGSI_FS_COORD_PIXEL_CENTER_INTEGER
)) {
4290 /* the driver supports pixel center integer, need to bias X,Y */
4291 ureg_property_fs_coord_pixel_center(ureg
, TGSI_FS_COORD_PIXEL_CENTER_INTEGER
);
4292 emit_adjusted_wpos(t
, program
, 0.5f
, invert
? -0.5f
: 0.5f
);
4298 /* we invert after adjustment so that we avoid the MOV to temporary,
4299 * and reuse the adjustment ADD instead */
4300 emit_wpos_inversion(t
, program
, invert
);
4304 * OpenGL's fragment gl_FrontFace input is 1 for front-facing, 0 for back.
4305 * TGSI uses +1 for front, -1 for back.
4306 * This function converts the TGSI value to the GL value. Simply clamping/
4307 * saturating the value to [0,1] does the job.
4310 emit_face_var(struct st_translate
*t
)
4312 struct ureg_program
*ureg
= t
->ureg
;
4313 struct ureg_dst face_temp
= ureg_DECL_temporary(ureg
);
4314 struct ureg_src face_input
= t
->inputs
[t
->inputMapping
[FRAG_ATTRIB_FACE
]];
4316 /* MOV_SAT face_temp, input[face] */
4317 face_temp
= ureg_saturate(face_temp
);
4318 ureg_MOV(ureg
, face_temp
, face_input
);
4320 /* Use face_temp as face input from here on: */
4321 t
->inputs
[t
->inputMapping
[FRAG_ATTRIB_FACE
]] = ureg_src(face_temp
);
4325 emit_edgeflags(struct st_translate
*t
)
4327 struct ureg_program
*ureg
= t
->ureg
;
4328 struct ureg_dst edge_dst
= t
->outputs
[t
->outputMapping
[VERT_RESULT_EDGE
]];
4329 struct ureg_src edge_src
= t
->inputs
[t
->inputMapping
[VERT_ATTRIB_EDGEFLAG
]];
4331 ureg_MOV(ureg
, edge_dst
, edge_src
);
4335 * Translate intermediate IR (glsl_to_tgsi_instruction) to TGSI format.
4336 * \param program the program to translate
4337 * \param numInputs number of input registers used
4338 * \param inputMapping maps Mesa fragment program inputs to TGSI generic
4340 * \param inputSemanticName the TGSI_SEMANTIC flag for each input
4341 * \param inputSemanticIndex the semantic index (ex: which texcoord) for
4343 * \param interpMode the TGSI_INTERPOLATE_LINEAR/PERSP mode for each input
4344 * \param numOutputs number of output registers used
4345 * \param outputMapping maps Mesa fragment program outputs to TGSI
4347 * \param outputSemanticName the TGSI_SEMANTIC flag for each output
4348 * \param outputSemanticIndex the semantic index (ex: which texcoord) for
4351 * \return PIPE_OK or PIPE_ERROR_OUT_OF_MEMORY
4353 extern "C" enum pipe_error
4354 st_translate_program(
4355 struct gl_context
*ctx
,
4357 struct ureg_program
*ureg
,
4358 glsl_to_tgsi_visitor
*program
,
4359 const struct gl_program
*proginfo
,
4361 const GLuint inputMapping
[],
4362 const ubyte inputSemanticName
[],
4363 const ubyte inputSemanticIndex
[],
4364 const GLuint interpMode
[],
4366 const GLuint outputMapping
[],
4367 const ubyte outputSemanticName
[],
4368 const ubyte outputSemanticIndex
[],
4369 boolean passthrough_edgeflags
)
4371 struct st_translate translate
, *t
;
4373 enum pipe_error ret
= PIPE_OK
;
4375 assert(numInputs
<= Elements(t
->inputs
));
4376 assert(numOutputs
<= Elements(t
->outputs
));
4379 memset(t
, 0, sizeof *t
);
4381 t
->procType
= procType
;
4382 t
->inputMapping
= inputMapping
;
4383 t
->outputMapping
= outputMapping
;
4385 t
->pointSizeOutIndex
= -1;
4386 t
->prevInstWrotePointSize
= GL_FALSE
;
4389 * Declare input attributes.
4391 if (procType
== TGSI_PROCESSOR_FRAGMENT
) {
4392 for (i
= 0; i
< numInputs
; i
++) {
4393 t
->inputs
[i
] = ureg_DECL_fs_input(ureg
,
4394 inputSemanticName
[i
],
4395 inputSemanticIndex
[i
],
4399 if (proginfo
->InputsRead
& FRAG_BIT_WPOS
) {
4400 /* Must do this after setting up t->inputs, and before
4401 * emitting constant references, below:
4403 emit_wpos(st_context(ctx
), t
, proginfo
, ureg
);
4406 if (proginfo
->InputsRead
& FRAG_BIT_FACE
)
4410 * Declare output attributes.
4412 for (i
= 0; i
< numOutputs
; i
++) {
4413 switch (outputSemanticName
[i
]) {
4414 case TGSI_SEMANTIC_POSITION
:
4415 t
->outputs
[i
] = ureg_DECL_output(ureg
,
4416 TGSI_SEMANTIC_POSITION
, /* Z/Depth */
4417 outputSemanticIndex
[i
]);
4418 t
->outputs
[i
] = ureg_writemask(t
->outputs
[i
], TGSI_WRITEMASK_Z
);
4420 case TGSI_SEMANTIC_STENCIL
:
4421 t
->outputs
[i
] = ureg_DECL_output(ureg
,
4422 TGSI_SEMANTIC_STENCIL
, /* Stencil */
4423 outputSemanticIndex
[i
]);
4424 t
->outputs
[i
] = ureg_writemask(t
->outputs
[i
], TGSI_WRITEMASK_Y
);
4426 case TGSI_SEMANTIC_COLOR
:
4427 t
->outputs
[i
] = ureg_DECL_output(ureg
,
4428 TGSI_SEMANTIC_COLOR
,
4429 outputSemanticIndex
[i
]);
4432 assert(!"fragment shader outputs must be POSITION/STENCIL/COLOR");
4433 return PIPE_ERROR_BAD_INPUT
;
4437 else if (procType
== TGSI_PROCESSOR_GEOMETRY
) {
4438 for (i
= 0; i
< numInputs
; i
++) {
4439 t
->inputs
[i
] = ureg_DECL_gs_input(ureg
,
4441 inputSemanticName
[i
],
4442 inputSemanticIndex
[i
]);
4445 for (i
= 0; i
< numOutputs
; i
++) {
4446 t
->outputs
[i
] = ureg_DECL_output(ureg
,
4447 outputSemanticName
[i
],
4448 outputSemanticIndex
[i
]);
4452 assert(procType
== TGSI_PROCESSOR_VERTEX
);
4454 for (i
= 0; i
< numInputs
; i
++) {
4455 t
->inputs
[i
] = ureg_DECL_vs_input(ureg
, i
);
4458 for (i
= 0; i
< numOutputs
; i
++) {
4459 t
->outputs
[i
] = ureg_DECL_output(ureg
,
4460 outputSemanticName
[i
],
4461 outputSemanticIndex
[i
]);
4462 if ((outputSemanticName
[i
] == TGSI_SEMANTIC_PSIZE
) && proginfo
->Id
) {
4463 /* Writing to the point size result register requires special
4464 * handling to implement clamping.
4466 static const gl_state_index pointSizeClampState
[STATE_LENGTH
]
4467 = { STATE_INTERNAL
, STATE_POINT_SIZE_IMPL_CLAMP
, (gl_state_index
)0, (gl_state_index
)0, (gl_state_index
)0 };
4468 /* XXX: note we are modifying the incoming shader here! Need to
4469 * do this before emitting the constant decls below, or this
4472 unsigned pointSizeClampConst
=
4473 _mesa_add_state_reference(proginfo
->Parameters
,
4474 pointSizeClampState
);
4475 struct ureg_dst psizregtemp
= ureg_DECL_temporary(ureg
);
4476 t
->pointSizeConst
= ureg_DECL_constant(ureg
, pointSizeClampConst
);
4477 t
->pointSizeResult
= t
->outputs
[i
];
4478 t
->pointSizeOutIndex
= i
;
4479 t
->outputs
[i
] = psizregtemp
;
4482 if (passthrough_edgeflags
)
4486 /* Declare address register.
4488 if (program
->num_address_regs
> 0) {
4489 assert(program
->num_address_regs
== 1);
4490 t
->address
[0] = ureg_DECL_address(ureg
);
4493 /* Declare misc input registers
4496 GLbitfield sysInputs
= proginfo
->SystemValuesRead
;
4497 unsigned numSys
= 0;
4498 for (i
= 0; sysInputs
; i
++) {
4499 if (sysInputs
& (1 << i
)) {
4500 unsigned semName
= mesa_sysval_to_semantic
[i
];
4501 t
->systemValues
[i
] = ureg_DECL_system_value(ureg
, numSys
, semName
, 0);
4503 sysInputs
&= ~(1 << i
);
4508 if (program
->indirect_addr_temps
) {
4509 /* If temps are accessed with indirect addressing, declare temporaries
4510 * in sequential order. Else, we declare them on demand elsewhere.
4511 * (Note: the number of temporaries is equal to program->next_temp)
4513 for (i
= 0; i
< (unsigned)program
->next_temp
; i
++) {
4514 /* XXX use TGSI_FILE_TEMPORARY_ARRAY when it's supported by ureg */
4515 t
->temps
[i
] = ureg_DECL_temporary(t
->ureg
);
4519 /* Emit constants and uniforms. TGSI uses a single index space for these,
4520 * so we put all the translated regs in t->constants.
4522 if (proginfo
->Parameters
) {
4523 t
->constants
= (struct ureg_src
*)CALLOC(proginfo
->Parameters
->NumParameters
* sizeof(t
->constants
[0]));
4524 if (t
->constants
== NULL
) {
4525 ret
= PIPE_ERROR_OUT_OF_MEMORY
;
4529 for (i
= 0; i
< proginfo
->Parameters
->NumParameters
; i
++) {
4530 switch (proginfo
->Parameters
->Parameters
[i
].Type
) {
4531 case PROGRAM_ENV_PARAM
:
4532 case PROGRAM_LOCAL_PARAM
:
4533 case PROGRAM_STATE_VAR
:
4534 case PROGRAM_NAMED_PARAM
:
4535 case PROGRAM_UNIFORM
:
4536 t
->constants
[i
] = ureg_DECL_constant(ureg
, i
);
4539 /* Emit immediates for PROGRAM_CONSTANT only when there's no indirect
4540 * addressing of the const buffer.
4541 * FIXME: Be smarter and recognize param arrays:
4542 * indirect addressing is only valid within the referenced
4545 case PROGRAM_CONSTANT
:
4546 if (program
->indirect_addr_consts
)
4547 t
->constants
[i
] = ureg_DECL_constant(ureg
, i
);
4549 t
->constants
[i
] = emit_immediate(t
,
4550 proginfo
->Parameters
->ParameterValues
[i
],
4551 proginfo
->Parameters
->Parameters
[i
].DataType
,
4560 /* Emit immediate values.
4562 t
->immediates
= (struct ureg_src
*)CALLOC(program
->num_immediates
* sizeof(struct ureg_src
));
4563 if (t
->immediates
== NULL
) {
4564 ret
= PIPE_ERROR_OUT_OF_MEMORY
;
4568 foreach_iter(exec_list_iterator
, iter
, program
->immediates
) {
4569 immediate_storage
*imm
= (immediate_storage
*)iter
.get();
4570 t
->immediates
[i
++] = emit_immediate(t
, imm
->values
, imm
->type
, imm
->size
);
4573 /* texture samplers */
4574 for (i
= 0; i
< ctx
->Const
.MaxTextureImageUnits
; i
++) {
4575 if (program
->samplers_used
& (1 << i
)) {
4576 t
->samplers
[i
] = ureg_DECL_sampler(ureg
, i
);
4580 /* Emit each instruction in turn:
4582 foreach_iter(exec_list_iterator
, iter
, program
->instructions
) {
4583 set_insn_start(t
, ureg_get_instruction_number(ureg
));
4584 compile_tgsi_instruction(t
, (glsl_to_tgsi_instruction
*)iter
.get());
4586 if (t
->prevInstWrotePointSize
&& proginfo
->Id
) {
4587 /* The previous instruction wrote to the (fake) vertex point size
4588 * result register. Now we need to clamp that value to the min/max
4589 * point size range, putting the result into the real point size
4591 * Note that we can't do this easily at the end of program due to
4592 * possible early return.
4594 set_insn_start(t
, ureg_get_instruction_number(ureg
));
4596 ureg_writemask(t
->outputs
[t
->pointSizeOutIndex
], WRITEMASK_X
),
4597 ureg_src(t
->outputs
[t
->pointSizeOutIndex
]),
4598 ureg_swizzle(t
->pointSizeConst
, 1,1,1,1));
4599 ureg_MIN(t
->ureg
, ureg_writemask(t
->pointSizeResult
, WRITEMASK_X
),
4600 ureg_src(t
->outputs
[t
->pointSizeOutIndex
]),
4601 ureg_swizzle(t
->pointSizeConst
, 2,2,2,2));
4603 t
->prevInstWrotePointSize
= GL_FALSE
;
4606 /* Fix up all emitted labels:
4608 for (i
= 0; i
< t
->labels_count
; i
++) {
4609 ureg_fixup_label(ureg
, t
->labels
[i
].token
,
4610 t
->insn
[t
->labels
[i
].branch_target
]);
4617 FREE(t
->immediates
);
4620 debug_printf("%s: translate error flag set\n", __FUNCTION__
);
4625 /* ----------------------------- End TGSI code ------------------------------ */
4628 * Convert a shader's GLSL IR into a Mesa gl_program, although without
4629 * generating Mesa IR.
4631 static struct gl_program
*
4632 get_mesa_program(struct gl_context
*ctx
,
4633 struct gl_shader_program
*shader_program
,
4634 struct gl_shader
*shader
)
4636 glsl_to_tgsi_visitor
* v
= new glsl_to_tgsi_visitor();
4637 struct gl_program
*prog
;
4639 const char *target_string
;
4641 struct gl_shader_compiler_options
*options
=
4642 &ctx
->ShaderCompilerOptions
[_mesa_shader_type_to_index(shader
->Type
)];
4644 switch (shader
->Type
) {
4645 case GL_VERTEX_SHADER
:
4646 target
= GL_VERTEX_PROGRAM_ARB
;
4647 target_string
= "vertex";
4649 case GL_FRAGMENT_SHADER
:
4650 target
= GL_FRAGMENT_PROGRAM_ARB
;
4651 target_string
= "fragment";
4653 case GL_GEOMETRY_SHADER
:
4654 target
= GL_GEOMETRY_PROGRAM_NV
;
4655 target_string
= "geometry";
4658 assert(!"should not be reached");
4662 validate_ir_tree(shader
->ir
);
4664 prog
= ctx
->Driver
.NewProgram(ctx
, target
, shader_program
->Name
);
4667 prog
->Parameters
= _mesa_new_parameter_list();
4668 prog
->Varying
= _mesa_new_parameter_list();
4669 prog
->Attributes
= _mesa_new_parameter_list();
4672 v
->shader_program
= shader_program
;
4673 v
->options
= options
;
4674 v
->glsl_version
= ctx
->Const
.GLSLVersion
;
4676 add_uniforms_to_parameters_list(shader_program
, shader
, prog
);
4678 /* Emit intermediate IR for main(). */
4679 visit_exec_list(shader
->ir
, v
);
4681 /* Now emit bodies for any functions that were used. */
4683 progress
= GL_FALSE
;
4685 foreach_iter(exec_list_iterator
, iter
, v
->function_signatures
) {
4686 function_entry
*entry
= (function_entry
*)iter
.get();
4688 if (!entry
->bgn_inst
) {
4689 v
->current_function
= entry
;
4691 entry
->bgn_inst
= v
->emit(NULL
, TGSI_OPCODE_BGNSUB
);
4692 entry
->bgn_inst
->function
= entry
;
4694 visit_exec_list(&entry
->sig
->body
, v
);
4696 glsl_to_tgsi_instruction
*last
;
4697 last
= (glsl_to_tgsi_instruction
*)v
->instructions
.get_tail();
4698 if (last
->op
!= TGSI_OPCODE_RET
)
4699 v
->emit(NULL
, TGSI_OPCODE_RET
);
4701 glsl_to_tgsi_instruction
*end
;
4702 end
= v
->emit(NULL
, TGSI_OPCODE_ENDSUB
);
4703 end
->function
= entry
;
4711 /* Print out some information (for debugging purposes) used by the
4712 * optimization passes. */
4713 for (i
=0; i
< v
->next_temp
; i
++) {
4714 int fr
= v
->get_first_temp_read(i
);
4715 int fw
= v
->get_first_temp_write(i
);
4716 int lr
= v
->get_last_temp_read(i
);
4717 int lw
= v
->get_last_temp_write(i
);
4719 printf("Temp %d: FR=%3d FW=%3d LR=%3d LW=%3d\n", i
, fr
, fw
, lr
, lw
);
4724 /* Remove reads to output registers, and to varyings in vertex shaders. */
4725 v
->remove_output_reads(PROGRAM_OUTPUT
);
4726 if (target
== GL_VERTEX_PROGRAM_ARB
)
4727 v
->remove_output_reads(PROGRAM_VARYING
);
4729 /* Perform optimizations on the instructions in the glsl_to_tgsi_visitor. */
4731 v
->copy_propagate();
4732 while (v
->eliminate_dead_code_advanced());
4734 /* FIXME: These passes to optimize temporary registers don't work when there
4735 * is indirect addressing of the temporary register space. We need proper
4736 * array support so that we don't have to give up these passes in every
4737 * shader that uses arrays.
4739 if (!v
->indirect_addr_temps
) {
4740 v
->eliminate_dead_code();
4741 v
->merge_registers();
4742 v
->renumber_registers();
4745 /* Write the END instruction. */
4746 v
->emit(NULL
, TGSI_OPCODE_END
);
4748 if (ctx
->Shader
.Flags
& GLSL_DUMP
) {
4750 printf("GLSL IR for linked %s program %d:\n", target_string
,
4751 shader_program
->Name
);
4752 _mesa_print_ir(shader
->ir
, NULL
);
4757 prog
->Instructions
= NULL
;
4758 prog
->NumInstructions
= 0;
4760 do_set_program_inouts(shader
->ir
, prog
);
4761 count_resources(v
, prog
);
4763 check_resources(ctx
, shader_program
, v
, prog
);
4765 _mesa_reference_program(ctx
, &shader
->Program
, prog
);
4767 struct st_vertex_program
*stvp
;
4768 struct st_fragment_program
*stfp
;
4769 struct st_geometry_program
*stgp
;
4771 switch (shader
->Type
) {
4772 case GL_VERTEX_SHADER
:
4773 stvp
= (struct st_vertex_program
*)prog
;
4774 stvp
->glsl_to_tgsi
= v
;
4776 case GL_FRAGMENT_SHADER
:
4777 stfp
= (struct st_fragment_program
*)prog
;
4778 stfp
->glsl_to_tgsi
= v
;
4780 case GL_GEOMETRY_SHADER
:
4781 stgp
= (struct st_geometry_program
*)prog
;
4782 stgp
->glsl_to_tgsi
= v
;
4785 assert(!"should not be reached");
4795 st_new_shader(struct gl_context
*ctx
, GLuint name
, GLuint type
)
4797 struct gl_shader
*shader
;
4798 assert(type
== GL_FRAGMENT_SHADER
|| type
== GL_VERTEX_SHADER
||
4799 type
== GL_GEOMETRY_SHADER_ARB
);
4800 shader
= rzalloc(NULL
, struct gl_shader
);
4802 shader
->Type
= type
;
4803 shader
->Name
= name
;
4804 _mesa_init_shader(ctx
, shader
);
4809 struct gl_shader_program
*
4810 st_new_shader_program(struct gl_context
*ctx
, GLuint name
)
4812 struct gl_shader_program
*shProg
;
4813 shProg
= rzalloc(NULL
, struct gl_shader_program
);
4815 shProg
->Name
= name
;
4816 _mesa_init_shader_program(ctx
, shProg
);
4823 * Called via ctx->Driver.LinkShader()
4824 * This actually involves converting GLSL IR into an intermediate TGSI-like IR
4825 * with code lowering and other optimizations.
4828 st_link_shader(struct gl_context
*ctx
, struct gl_shader_program
*prog
)
4830 assert(prog
->LinkStatus
);
4832 for (unsigned i
= 0; i
< MESA_SHADER_TYPES
; i
++) {
4833 if (prog
->_LinkedShaders
[i
] == NULL
)
4837 exec_list
*ir
= prog
->_LinkedShaders
[i
]->ir
;
4838 const struct gl_shader_compiler_options
*options
=
4839 &ctx
->ShaderCompilerOptions
[_mesa_shader_type_to_index(prog
->_LinkedShaders
[i
]->Type
)];
4845 do_mat_op_to_vec(ir
);
4846 lower_instructions(ir
, (MOD_TO_FRACT
| DIV_TO_MUL_RCP
| EXP_TO_EXP2
4848 | ((options
->EmitNoPow
) ? POW_TO_EXP2
: 0)));
4850 progress
= do_lower_jumps(ir
, true, true, options
->EmitNoMainReturn
, options
->EmitNoCont
, options
->EmitNoLoops
) || progress
;
4852 progress
= do_common_optimization(ir
, true, options
->MaxUnrollIterations
) || progress
;
4854 progress
= lower_quadop_vector(ir
, false) || progress
;
4856 if (options
->EmitNoIfs
) {
4857 progress
= lower_discard(ir
) || progress
;
4858 progress
= lower_if_to_cond_assign(ir
) || progress
;
4861 if (options
->EmitNoNoise
)
4862 progress
= lower_noise(ir
) || progress
;
4864 /* If there are forms of indirect addressing that the driver
4865 * cannot handle, perform the lowering pass.
4867 if (options
->EmitNoIndirectInput
|| options
->EmitNoIndirectOutput
4868 || options
->EmitNoIndirectTemp
|| options
->EmitNoIndirectUniform
)
4870 lower_variable_index_to_cond_assign(ir
,
4871 options
->EmitNoIndirectInput
,
4872 options
->EmitNoIndirectOutput
,
4873 options
->EmitNoIndirectTemp
,
4874 options
->EmitNoIndirectUniform
)
4877 progress
= do_vec_index_to_cond_assign(ir
) || progress
;
4880 validate_ir_tree(ir
);
4883 for (unsigned i
= 0; i
< MESA_SHADER_TYPES
; i
++) {
4884 struct gl_program
*linked_prog
;
4886 if (prog
->_LinkedShaders
[i
] == NULL
)
4889 linked_prog
= get_mesa_program(ctx
, prog
, prog
->_LinkedShaders
[i
]);
4894 switch (prog
->_LinkedShaders
[i
]->Type
) {
4895 case GL_VERTEX_SHADER
:
4896 _mesa_reference_vertprog(ctx
, &prog
->VertexProgram
,
4897 (struct gl_vertex_program
*)linked_prog
);
4898 ok
= ctx
->Driver
.ProgramStringNotify(ctx
, GL_VERTEX_PROGRAM_ARB
,
4901 case GL_FRAGMENT_SHADER
:
4902 _mesa_reference_fragprog(ctx
, &prog
->FragmentProgram
,
4903 (struct gl_fragment_program
*)linked_prog
);
4904 ok
= ctx
->Driver
.ProgramStringNotify(ctx
, GL_FRAGMENT_PROGRAM_ARB
,
4907 case GL_GEOMETRY_SHADER
:
4908 _mesa_reference_geomprog(ctx
, &prog
->GeometryProgram
,
4909 (struct gl_geometry_program
*)linked_prog
);
4910 ok
= ctx
->Driver
.ProgramStringNotify(ctx
, GL_GEOMETRY_PROGRAM_NV
,
4919 _mesa_reference_program(ctx
, &linked_prog
, NULL
);
4927 * Link a GLSL shader program. Called via glLinkProgram().
4930 st_glsl_link_shader(struct gl_context
*ctx
, struct gl_shader_program
*prog
)
4934 _mesa_clear_shader_program_data(ctx
, prog
);
4936 prog
->LinkStatus
= GL_TRUE
;
4938 for (i
= 0; i
< prog
->NumShaders
; i
++) {
4939 if (!prog
->Shaders
[i
]->CompileStatus
) {
4940 fail_link(prog
, "linking with uncompiled shader");
4941 prog
->LinkStatus
= GL_FALSE
;
4945 prog
->Varying
= _mesa_new_parameter_list();
4946 _mesa_reference_vertprog(ctx
, &prog
->VertexProgram
, NULL
);
4947 _mesa_reference_fragprog(ctx
, &prog
->FragmentProgram
, NULL
);
4948 _mesa_reference_geomprog(ctx
, &prog
->GeometryProgram
, NULL
);
4950 if (prog
->LinkStatus
) {
4951 link_shaders(ctx
, prog
);
4954 if (prog
->LinkStatus
) {
4955 if (!ctx
->Driver
.LinkShader(ctx
, prog
)) {
4956 prog
->LinkStatus
= GL_FALSE
;
4960 set_uniform_initializers(ctx
, prog
);
4962 if (ctx
->Shader
.Flags
& GLSL_DUMP
) {
4963 if (!prog
->LinkStatus
) {
4964 printf("GLSL shader program %d failed to link\n", prog
->Name
);
4967 if (prog
->InfoLog
&& prog
->InfoLog
[0] != 0) {
4968 printf("GLSL shader program %d info log:\n", prog
->Name
);
4969 printf("%s\n", prog
->InfoLog
);