2 * Copyright (C) 2005-2007 Brian Paul All Rights Reserved.
3 * Copyright (C) 2008 VMware, Inc. All Rights Reserved.
4 * Copyright © 2010 Intel Corporation
5 * Copyright © 2011 Bryan Cain
7 * Permission is hereby granted, free of charge, to any person obtaining a
8 * copy of this software and associated documentation files (the "Software"),
9 * to deal in the Software without restriction, including without limitation
10 * the rights to use, copy, modify, merge, publish, distribute, sublicense,
11 * and/or sell copies of the Software, and to permit persons to whom the
12 * Software is furnished to do so, subject to the following conditions:
14 * The above copyright notice and this permission notice (including the next
15 * paragraph) shall be included in all copies or substantial portions of the
18 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
19 * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
20 * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL
21 * THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
22 * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING
23 * FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER
24 * DEALINGS IN THE SOFTWARE.
28 * \file glsl_to_tgsi.cpp
30 * Translate GLSL IR to TGSI.
34 #include "main/compiler.h"
36 #include "ir_visitor.h"
37 #include "ir_print_visitor.h"
38 #include "ir_expression_flattening.h"
39 #include "glsl_types.h"
40 #include "glsl_parser_extras.h"
41 #include "../glsl/program.h"
42 #include "ir_optimization.h"
46 #include "main/mtypes.h"
47 #include "main/shaderapi.h"
48 #include "main/shaderobj.h"
49 #include "main/uniforms.h"
50 #include "program/hash_table.h"
51 #include "program/prog_instruction.h"
52 #include "program/prog_optimize.h"
53 #include "program/prog_print.h"
54 #include "program/program.h"
55 #include "program/prog_uniform.h"
56 #include "program/prog_parameter.h"
57 #include "program/sampler.h"
59 #include "pipe/p_compiler.h"
60 #include "pipe/p_context.h"
61 #include "pipe/p_screen.h"
62 #include "pipe/p_shader_tokens.h"
63 #include "pipe/p_state.h"
64 #include "util/u_math.h"
65 #include "tgsi/tgsi_ureg.h"
66 #include "tgsi/tgsi_info.h"
67 #include "st_context.h"
68 #include "st_program.h"
69 #include "st_glsl_to_tgsi.h"
70 #include "st_mesa_to_tgsi.h"
73 #define PROGRAM_IMMEDIATE PROGRAM_FILE_MAX
74 #define PROGRAM_ANY_CONST ((1 << PROGRAM_LOCAL_PARAM) | \
75 (1 << PROGRAM_ENV_PARAM) | \
76 (1 << PROGRAM_STATE_VAR) | \
77 (1 << PROGRAM_NAMED_PARAM) | \
78 (1 << PROGRAM_CONSTANT) | \
79 (1 << PROGRAM_UNIFORM))
81 #define MAX_TEMPS 4096
83 /* will be 4 for GLSL 4.00 */
84 #define MAX_GLSL_TEXTURE_OFFSET 1
89 static int swizzle_for_size(int size
);
92 * This struct is a corresponding struct to TGSI ureg_src.
96 st_src_reg(gl_register_file file
, int index
, const glsl_type
*type
)
100 if (type
&& (type
->is_scalar() || type
->is_vector() || type
->is_matrix()))
101 this->swizzle
= swizzle_for_size(type
->vector_elements
);
103 this->swizzle
= SWIZZLE_XYZW
;
105 this->type
= type
? type
->base_type
: GLSL_TYPE_ERROR
;
106 this->reladdr
= NULL
;
109 st_src_reg(gl_register_file file
, int index
, int type
)
114 this->swizzle
= SWIZZLE_XYZW
;
116 this->reladdr
= NULL
;
121 this->type
= GLSL_TYPE_ERROR
;
122 this->file
= PROGRAM_UNDEFINED
;
126 this->reladdr
= NULL
;
129 explicit st_src_reg(st_dst_reg reg
);
131 gl_register_file file
; /**< PROGRAM_* from Mesa */
132 int index
; /**< temporary index, VERT_ATTRIB_*, FRAG_ATTRIB_*, etc. */
133 GLuint swizzle
; /**< SWIZZLE_XYZWONEZERO swizzles from Mesa. */
134 int negate
; /**< NEGATE_XYZW mask from mesa */
135 int type
; /** GLSL_TYPE_* from GLSL IR (enum glsl_base_type) */
136 /** Register index should be offset by the integer in this reg. */
142 st_dst_reg(gl_register_file file
, int writemask
, int type
)
146 this->writemask
= writemask
;
147 this->cond_mask
= COND_TR
;
148 this->reladdr
= NULL
;
154 this->type
= GLSL_TYPE_ERROR
;
155 this->file
= PROGRAM_UNDEFINED
;
158 this->cond_mask
= COND_TR
;
159 this->reladdr
= NULL
;
162 explicit st_dst_reg(st_src_reg reg
);
164 gl_register_file file
; /**< PROGRAM_* from Mesa */
165 int index
; /**< temporary index, VERT_ATTRIB_*, FRAG_ATTRIB_*, etc. */
166 int writemask
; /**< Bitfield of WRITEMASK_[XYZW] */
168 int type
; /** GLSL_TYPE_* from GLSL IR (enum glsl_base_type) */
169 /** Register index should be offset by the integer in this reg. */
173 st_src_reg::st_src_reg(st_dst_reg reg
)
175 this->type
= reg
.type
;
176 this->file
= reg
.file
;
177 this->index
= reg
.index
;
178 this->swizzle
= SWIZZLE_XYZW
;
180 this->reladdr
= reg
.reladdr
;
183 st_dst_reg::st_dst_reg(st_src_reg reg
)
185 this->type
= reg
.type
;
186 this->file
= reg
.file
;
187 this->index
= reg
.index
;
188 this->writemask
= WRITEMASK_XYZW
;
189 this->cond_mask
= COND_TR
;
190 this->reladdr
= reg
.reladdr
;
193 class glsl_to_tgsi_instruction
: public exec_node
{
195 /* Callers of this ralloc-based new need not call delete. It's
196 * easier to just ralloc_free 'ctx' (or any of its ancestors). */
197 static void* operator new(size_t size
, void *ctx
)
201 node
= rzalloc_size(ctx
, size
);
202 assert(node
!= NULL
);
210 /** Pointer to the ir source this tree came from for debugging */
212 GLboolean cond_update
;
214 int sampler
; /**< sampler index */
215 int tex_target
; /**< One of TEXTURE_*_INDEX */
216 GLboolean tex_shadow
;
217 struct tgsi_texture_offset tex_offsets
[MAX_GLSL_TEXTURE_OFFSET
];
218 unsigned tex_offset_num_offset
;
219 int dead_mask
; /**< Used in dead code elimination */
221 class function_entry
*function
; /* Set on TGSI_OPCODE_CAL or TGSI_OPCODE_BGNSUB */
224 class variable_storage
: public exec_node
{
226 variable_storage(ir_variable
*var
, gl_register_file file
, int index
)
227 : file(file
), index(index
), var(var
)
232 gl_register_file file
;
234 ir_variable
*var
; /* variable that maps to this, if any */
237 class immediate_storage
: public exec_node
{
239 immediate_storage(gl_constant_value
*values
, int size
, int type
)
241 memcpy(this->values
, values
, size
* sizeof(gl_constant_value
));
246 gl_constant_value values
[4];
247 int size
; /**< Number of components (1-4) */
248 int type
; /**< GL_FLOAT, GL_INT, GL_BOOL, or GL_UNSIGNED_INT */
251 class function_entry
: public exec_node
{
253 ir_function_signature
*sig
;
256 * identifier of this function signature used by the program.
258 * At the point that TGSI instructions for function calls are
259 * generated, we don't know the address of the first instruction of
260 * the function body. So we make the BranchTarget that is called a
261 * small integer and rewrite them during set_branchtargets().
266 * Pointer to first instruction of the function body.
268 * Set during function body emits after main() is processed.
270 glsl_to_tgsi_instruction
*bgn_inst
;
273 * Index of the first instruction of the function body in actual TGSI.
275 * Set after conversion from glsl_to_tgsi_instruction to TGSI.
279 /** Storage for the return value. */
280 st_src_reg return_reg
;
283 class glsl_to_tgsi_visitor
: public ir_visitor
{
285 glsl_to_tgsi_visitor();
286 ~glsl_to_tgsi_visitor();
288 function_entry
*current_function
;
290 struct gl_context
*ctx
;
291 struct gl_program
*prog
;
292 struct gl_shader_program
*shader_program
;
293 struct gl_shader_compiler_options
*options
;
297 int num_address_regs
;
299 bool indirect_addr_temps
;
300 bool indirect_addr_consts
;
303 bool native_integers
;
305 variable_storage
*find_variable_storage(ir_variable
*var
);
307 int add_constant(gl_register_file file
, gl_constant_value values
[4],
308 int size
, int datatype
, GLuint
*swizzle_out
);
310 function_entry
*get_function_signature(ir_function_signature
*sig
);
312 st_src_reg
get_temp(const glsl_type
*type
);
313 void reladdr_to_temp(ir_instruction
*ir
, st_src_reg
*reg
, int *num_reladdr
);
315 st_src_reg
st_src_reg_for_float(float val
);
316 st_src_reg
st_src_reg_for_int(int val
);
317 st_src_reg
st_src_reg_for_type(int type
, int val
);
320 * \name Visit methods
322 * As typical for the visitor pattern, there must be one \c visit method for
323 * each concrete subclass of \c ir_instruction. Virtual base classes within
324 * the hierarchy should not have \c visit methods.
327 virtual void visit(ir_variable
*);
328 virtual void visit(ir_loop
*);
329 virtual void visit(ir_loop_jump
*);
330 virtual void visit(ir_function_signature
*);
331 virtual void visit(ir_function
*);
332 virtual void visit(ir_expression
*);
333 virtual void visit(ir_swizzle
*);
334 virtual void visit(ir_dereference_variable
*);
335 virtual void visit(ir_dereference_array
*);
336 virtual void visit(ir_dereference_record
*);
337 virtual void visit(ir_assignment
*);
338 virtual void visit(ir_constant
*);
339 virtual void visit(ir_call
*);
340 virtual void visit(ir_return
*);
341 virtual void visit(ir_discard
*);
342 virtual void visit(ir_texture
*);
343 virtual void visit(ir_if
*);
348 /** List of variable_storage */
351 /** List of immediate_storage */
352 exec_list immediates
;
355 /** List of function_entry */
356 exec_list function_signatures
;
357 int next_signature_id
;
359 /** List of glsl_to_tgsi_instruction */
360 exec_list instructions
;
362 glsl_to_tgsi_instruction
*emit(ir_instruction
*ir
, unsigned op
);
364 glsl_to_tgsi_instruction
*emit(ir_instruction
*ir
, unsigned op
,
365 st_dst_reg dst
, st_src_reg src0
);
367 glsl_to_tgsi_instruction
*emit(ir_instruction
*ir
, unsigned op
,
368 st_dst_reg dst
, st_src_reg src0
, st_src_reg src1
);
370 glsl_to_tgsi_instruction
*emit(ir_instruction
*ir
, unsigned op
,
372 st_src_reg src0
, st_src_reg src1
, st_src_reg src2
);
374 unsigned get_opcode(ir_instruction
*ir
, unsigned op
,
376 st_src_reg src0
, st_src_reg src1
);
379 * Emit the correct dot-product instruction for the type of arguments
381 glsl_to_tgsi_instruction
*emit_dp(ir_instruction
*ir
,
387 void emit_scalar(ir_instruction
*ir
, unsigned op
,
388 st_dst_reg dst
, st_src_reg src0
);
390 void emit_scalar(ir_instruction
*ir
, unsigned op
,
391 st_dst_reg dst
, st_src_reg src0
, st_src_reg src1
);
393 void emit_arl(ir_instruction
*ir
, st_dst_reg dst
, st_src_reg src0
);
395 void emit_scs(ir_instruction
*ir
, unsigned op
,
396 st_dst_reg dst
, const st_src_reg
&src
);
398 bool try_emit_mad(ir_expression
*ir
,
400 bool try_emit_mad_for_and_not(ir_expression
*ir
,
402 bool try_emit_sat(ir_expression
*ir
);
404 void emit_swz(ir_expression
*ir
);
406 bool process_move_condition(ir_rvalue
*ir
);
408 void remove_output_reads(gl_register_file type
);
409 void simplify_cmp(void);
411 void rename_temp_register(int index
, int new_index
);
412 int get_first_temp_read(int index
);
413 int get_first_temp_write(int index
);
414 int get_last_temp_read(int index
);
415 int get_last_temp_write(int index
);
417 void copy_propagate(void);
418 void eliminate_dead_code(void);
419 int eliminate_dead_code_advanced(void);
420 void merge_registers(void);
421 void renumber_registers(void);
426 static st_src_reg undef_src
= st_src_reg(PROGRAM_UNDEFINED
, 0, GLSL_TYPE_ERROR
);
428 static st_dst_reg undef_dst
= st_dst_reg(PROGRAM_UNDEFINED
, SWIZZLE_NOOP
, GLSL_TYPE_ERROR
);
430 static st_dst_reg address_reg
= st_dst_reg(PROGRAM_ADDRESS
, WRITEMASK_X
, GLSL_TYPE_FLOAT
);
433 fail_link(struct gl_shader_program
*prog
, const char *fmt
, ...) PRINTFLIKE(2, 3);
436 fail_link(struct gl_shader_program
*prog
, const char *fmt
, ...)
440 ralloc_vasprintf_append(&prog
->InfoLog
, fmt
, args
);
443 prog
->LinkStatus
= GL_FALSE
;
447 swizzle_for_size(int size
)
449 int size_swizzles
[4] = {
450 MAKE_SWIZZLE4(SWIZZLE_X
, SWIZZLE_X
, SWIZZLE_X
, SWIZZLE_X
),
451 MAKE_SWIZZLE4(SWIZZLE_X
, SWIZZLE_Y
, SWIZZLE_Y
, SWIZZLE_Y
),
452 MAKE_SWIZZLE4(SWIZZLE_X
, SWIZZLE_Y
, SWIZZLE_Z
, SWIZZLE_Z
),
453 MAKE_SWIZZLE4(SWIZZLE_X
, SWIZZLE_Y
, SWIZZLE_Z
, SWIZZLE_W
),
456 assert((size
>= 1) && (size
<= 4));
457 return size_swizzles
[size
- 1];
461 is_tex_instruction(unsigned opcode
)
463 const tgsi_opcode_info
* info
= tgsi_get_opcode_info(opcode
);
468 num_inst_dst_regs(unsigned opcode
)
470 const tgsi_opcode_info
* info
= tgsi_get_opcode_info(opcode
);
471 return info
->num_dst
;
475 num_inst_src_regs(unsigned opcode
)
477 const tgsi_opcode_info
* info
= tgsi_get_opcode_info(opcode
);
478 return info
->is_tex
? info
->num_src
- 1 : info
->num_src
;
481 glsl_to_tgsi_instruction
*
482 glsl_to_tgsi_visitor::emit(ir_instruction
*ir
, unsigned op
,
484 st_src_reg src0
, st_src_reg src1
, st_src_reg src2
)
486 glsl_to_tgsi_instruction
*inst
= new(mem_ctx
) glsl_to_tgsi_instruction();
487 int num_reladdr
= 0, i
;
489 op
= get_opcode(ir
, op
, dst
, src0
, src1
);
491 /* If we have to do relative addressing, we want to load the ARL
492 * reg directly for one of the regs, and preload the other reladdr
493 * sources into temps.
495 num_reladdr
+= dst
.reladdr
!= NULL
;
496 num_reladdr
+= src0
.reladdr
!= NULL
;
497 num_reladdr
+= src1
.reladdr
!= NULL
;
498 num_reladdr
+= src2
.reladdr
!= NULL
;
500 reladdr_to_temp(ir
, &src2
, &num_reladdr
);
501 reladdr_to_temp(ir
, &src1
, &num_reladdr
);
502 reladdr_to_temp(ir
, &src0
, &num_reladdr
);
505 emit_arl(ir
, address_reg
, *dst
.reladdr
);
508 assert(num_reladdr
== 0);
518 inst
->function
= NULL
;
520 if (op
== TGSI_OPCODE_ARL
)
521 this->num_address_regs
= 1;
523 /* Update indirect addressing status used by TGSI */
526 case PROGRAM_TEMPORARY
:
527 this->indirect_addr_temps
= true;
529 case PROGRAM_LOCAL_PARAM
:
530 case PROGRAM_ENV_PARAM
:
531 case PROGRAM_STATE_VAR
:
532 case PROGRAM_NAMED_PARAM
:
533 case PROGRAM_CONSTANT
:
534 case PROGRAM_UNIFORM
:
535 this->indirect_addr_consts
= true;
537 case PROGRAM_IMMEDIATE
:
538 assert(!"immediates should not have indirect addressing");
545 for (i
=0; i
<3; i
++) {
546 if(inst
->src
[i
].reladdr
) {
547 switch(inst
->src
[i
].file
) {
548 case PROGRAM_TEMPORARY
:
549 this->indirect_addr_temps
= true;
551 case PROGRAM_LOCAL_PARAM
:
552 case PROGRAM_ENV_PARAM
:
553 case PROGRAM_STATE_VAR
:
554 case PROGRAM_NAMED_PARAM
:
555 case PROGRAM_CONSTANT
:
556 case PROGRAM_UNIFORM
:
557 this->indirect_addr_consts
= true;
559 case PROGRAM_IMMEDIATE
:
560 assert(!"immediates should not have indirect addressing");
569 this->instructions
.push_tail(inst
);
575 glsl_to_tgsi_instruction
*
576 glsl_to_tgsi_visitor::emit(ir_instruction
*ir
, unsigned op
,
577 st_dst_reg dst
, st_src_reg src0
, st_src_reg src1
)
579 return emit(ir
, op
, dst
, src0
, src1
, undef_src
);
582 glsl_to_tgsi_instruction
*
583 glsl_to_tgsi_visitor::emit(ir_instruction
*ir
, unsigned op
,
584 st_dst_reg dst
, st_src_reg src0
)
586 assert(dst
.writemask
!= 0);
587 return emit(ir
, op
, dst
, src0
, undef_src
, undef_src
);
590 glsl_to_tgsi_instruction
*
591 glsl_to_tgsi_visitor::emit(ir_instruction
*ir
, unsigned op
)
593 return emit(ir
, op
, undef_dst
, undef_src
, undef_src
, undef_src
);
597 * Determines whether to use an integer, unsigned integer, or float opcode
598 * based on the operands and input opcode, then emits the result.
600 * TODO: type checking for remaining TGSI opcodes
603 glsl_to_tgsi_visitor::get_opcode(ir_instruction
*ir
, unsigned op
,
605 st_src_reg src0
, st_src_reg src1
)
607 int type
= GLSL_TYPE_FLOAT
;
609 if (src0
.type
== GLSL_TYPE_FLOAT
|| src1
.type
== GLSL_TYPE_FLOAT
)
610 type
= GLSL_TYPE_FLOAT
;
611 else if (native_integers
)
614 #define case4(c, f, i, u) \
615 case TGSI_OPCODE_##c: \
616 if (type == GLSL_TYPE_INT) op = TGSI_OPCODE_##i; \
617 else if (type == GLSL_TYPE_UINT) op = TGSI_OPCODE_##u; \
618 else op = TGSI_OPCODE_##f; \
620 #define case3(f, i, u) case4(f, f, i, u)
621 #define case2fi(f, i) case4(f, f, i, i)
622 #define case2iu(i, u) case4(i, LAST, i, u)
628 case3(DIV
, IDIV
, UDIV
);
629 case3(MAX
, IMAX
, UMAX
);
630 case3(MIN
, IMIN
, UMIN
);
635 case3(SGE
, ISGE
, USGE
);
636 case3(SLT
, ISLT
, USLT
);
648 assert(op
!= TGSI_OPCODE_LAST
);
652 glsl_to_tgsi_instruction
*
653 glsl_to_tgsi_visitor::emit_dp(ir_instruction
*ir
,
654 st_dst_reg dst
, st_src_reg src0
, st_src_reg src1
,
657 static const unsigned dot_opcodes
[] = {
658 TGSI_OPCODE_DP2
, TGSI_OPCODE_DP3
, TGSI_OPCODE_DP4
661 return emit(ir
, dot_opcodes
[elements
- 2], dst
, src0
, src1
);
665 * Emits TGSI scalar opcodes to produce unique answers across channels.
667 * Some TGSI opcodes are scalar-only, like ARB_fp/vp. The src X
668 * channel determines the result across all channels. So to do a vec4
669 * of this operation, we want to emit a scalar per source channel used
670 * to produce dest channels.
673 glsl_to_tgsi_visitor::emit_scalar(ir_instruction
*ir
, unsigned op
,
675 st_src_reg orig_src0
, st_src_reg orig_src1
)
678 int done_mask
= ~dst
.writemask
;
680 /* TGSI RCP is a scalar operation splatting results to all channels,
681 * like ARB_fp/vp. So emit as many RCPs as necessary to cover our
684 for (i
= 0; i
< 4; i
++) {
685 GLuint this_mask
= (1 << i
);
686 glsl_to_tgsi_instruction
*inst
;
687 st_src_reg src0
= orig_src0
;
688 st_src_reg src1
= orig_src1
;
690 if (done_mask
& this_mask
)
693 GLuint src0_swiz
= GET_SWZ(src0
.swizzle
, i
);
694 GLuint src1_swiz
= GET_SWZ(src1
.swizzle
, i
);
695 for (j
= i
+ 1; j
< 4; j
++) {
696 /* If there is another enabled component in the destination that is
697 * derived from the same inputs, generate its value on this pass as
700 if (!(done_mask
& (1 << j
)) &&
701 GET_SWZ(src0
.swizzle
, j
) == src0_swiz
&&
702 GET_SWZ(src1
.swizzle
, j
) == src1_swiz
) {
703 this_mask
|= (1 << j
);
706 src0
.swizzle
= MAKE_SWIZZLE4(src0_swiz
, src0_swiz
,
707 src0_swiz
, src0_swiz
);
708 src1
.swizzle
= MAKE_SWIZZLE4(src1_swiz
, src1_swiz
,
709 src1_swiz
, src1_swiz
);
711 inst
= emit(ir
, op
, dst
, src0
, src1
);
712 inst
->dst
.writemask
= this_mask
;
713 done_mask
|= this_mask
;
718 glsl_to_tgsi_visitor::emit_scalar(ir_instruction
*ir
, unsigned op
,
719 st_dst_reg dst
, st_src_reg src0
)
721 st_src_reg undef
= undef_src
;
723 undef
.swizzle
= SWIZZLE_XXXX
;
725 emit_scalar(ir
, op
, dst
, src0
, undef
);
729 glsl_to_tgsi_visitor::emit_arl(ir_instruction
*ir
,
730 st_dst_reg dst
, st_src_reg src0
)
732 st_src_reg tmp
= get_temp(glsl_type::float_type
);
734 if (src0
.type
== GLSL_TYPE_INT
)
735 emit(NULL
, TGSI_OPCODE_I2F
, st_dst_reg(tmp
), src0
);
736 else if (src0
.type
== GLSL_TYPE_UINT
)
737 emit(NULL
, TGSI_OPCODE_U2F
, st_dst_reg(tmp
), src0
);
741 emit(NULL
, TGSI_OPCODE_ARL
, dst
, tmp
);
745 * Emit an TGSI_OPCODE_SCS instruction
747 * The \c SCS opcode functions a bit differently than the other TGSI opcodes.
748 * Instead of splatting its result across all four components of the
749 * destination, it writes one value to the \c x component and another value to
750 * the \c y component.
752 * \param ir IR instruction being processed
753 * \param op Either \c TGSI_OPCODE_SIN or \c TGSI_OPCODE_COS depending
754 * on which value is desired.
755 * \param dst Destination register
756 * \param src Source register
759 glsl_to_tgsi_visitor::emit_scs(ir_instruction
*ir
, unsigned op
,
761 const st_src_reg
&src
)
763 /* Vertex programs cannot use the SCS opcode.
765 if (this->prog
->Target
== GL_VERTEX_PROGRAM_ARB
) {
766 emit_scalar(ir
, op
, dst
, src
);
770 const unsigned component
= (op
== TGSI_OPCODE_SIN
) ? 0 : 1;
771 const unsigned scs_mask
= (1U << component
);
772 int done_mask
= ~dst
.writemask
;
775 assert(op
== TGSI_OPCODE_SIN
|| op
== TGSI_OPCODE_COS
);
777 /* If there are compnents in the destination that differ from the component
778 * that will be written by the SCS instrution, we'll need a temporary.
780 if (scs_mask
!= unsigned(dst
.writemask
)) {
781 tmp
= get_temp(glsl_type::vec4_type
);
784 for (unsigned i
= 0; i
< 4; i
++) {
785 unsigned this_mask
= (1U << i
);
786 st_src_reg src0
= src
;
788 if ((done_mask
& this_mask
) != 0)
791 /* The source swizzle specified which component of the source generates
792 * sine / cosine for the current component in the destination. The SCS
793 * instruction requires that this value be swizzle to the X component.
794 * Replace the current swizzle with a swizzle that puts the source in
797 unsigned src0_swiz
= GET_SWZ(src
.swizzle
, i
);
799 src0
.swizzle
= MAKE_SWIZZLE4(src0_swiz
, src0_swiz
,
800 src0_swiz
, src0_swiz
);
801 for (unsigned j
= i
+ 1; j
< 4; j
++) {
802 /* If there is another enabled component in the destination that is
803 * derived from the same inputs, generate its value on this pass as
806 if (!(done_mask
& (1 << j
)) &&
807 GET_SWZ(src0
.swizzle
, j
) == src0_swiz
) {
808 this_mask
|= (1 << j
);
812 if (this_mask
!= scs_mask
) {
813 glsl_to_tgsi_instruction
*inst
;
814 st_dst_reg tmp_dst
= st_dst_reg(tmp
);
816 /* Emit the SCS instruction.
818 inst
= emit(ir
, TGSI_OPCODE_SCS
, tmp_dst
, src0
);
819 inst
->dst
.writemask
= scs_mask
;
821 /* Move the result of the SCS instruction to the desired location in
824 tmp
.swizzle
= MAKE_SWIZZLE4(component
, component
,
825 component
, component
);
826 inst
= emit(ir
, TGSI_OPCODE_SCS
, dst
, tmp
);
827 inst
->dst
.writemask
= this_mask
;
829 /* Emit the SCS instruction to write directly to the destination.
831 glsl_to_tgsi_instruction
*inst
= emit(ir
, TGSI_OPCODE_SCS
, dst
, src0
);
832 inst
->dst
.writemask
= scs_mask
;
835 done_mask
|= this_mask
;
840 glsl_to_tgsi_visitor::add_constant(gl_register_file file
,
841 gl_constant_value values
[4], int size
, int datatype
,
844 if (file
== PROGRAM_CONSTANT
) {
845 return _mesa_add_typed_unnamed_constant(this->prog
->Parameters
, values
,
846 size
, datatype
, swizzle_out
);
849 immediate_storage
*entry
;
850 assert(file
== PROGRAM_IMMEDIATE
);
852 /* Search immediate storage to see if we already have an identical
853 * immediate that we can use instead of adding a duplicate entry.
855 foreach_iter(exec_list_iterator
, iter
, this->immediates
) {
856 entry
= (immediate_storage
*)iter
.get();
858 if (entry
->size
== size
&&
859 entry
->type
== datatype
&&
860 !memcmp(entry
->values
, values
, size
* sizeof(gl_constant_value
))) {
866 /* Add this immediate to the list. */
867 entry
= new(mem_ctx
) immediate_storage(values
, size
, datatype
);
868 this->immediates
.push_tail(entry
);
869 this->num_immediates
++;
875 glsl_to_tgsi_visitor::st_src_reg_for_float(float val
)
877 st_src_reg
src(PROGRAM_IMMEDIATE
, -1, GLSL_TYPE_FLOAT
);
878 union gl_constant_value uval
;
881 src
.index
= add_constant(src
.file
, &uval
, 1, GL_FLOAT
, &src
.swizzle
);
887 glsl_to_tgsi_visitor::st_src_reg_for_int(int val
)
889 st_src_reg
src(PROGRAM_IMMEDIATE
, -1, GLSL_TYPE_INT
);
890 union gl_constant_value uval
;
892 assert(native_integers
);
895 src
.index
= add_constant(src
.file
, &uval
, 1, GL_INT
, &src
.swizzle
);
901 glsl_to_tgsi_visitor::st_src_reg_for_type(int type
, int val
)
904 return type
== GLSL_TYPE_FLOAT
? st_src_reg_for_float(val
) :
905 st_src_reg_for_int(val
);
907 return st_src_reg_for_float(val
);
911 type_size(const struct glsl_type
*type
)
916 switch (type
->base_type
) {
919 case GLSL_TYPE_FLOAT
:
921 if (type
->is_matrix()) {
922 return type
->matrix_columns
;
924 /* Regardless of size of vector, it gets a vec4. This is bad
925 * packing for things like floats, but otherwise arrays become a
926 * mess. Hopefully a later pass over the code can pack scalars
927 * down if appropriate.
931 case GLSL_TYPE_ARRAY
:
932 assert(type
->length
> 0);
933 return type_size(type
->fields
.array
) * type
->length
;
934 case GLSL_TYPE_STRUCT
:
936 for (i
= 0; i
< type
->length
; i
++) {
937 size
+= type_size(type
->fields
.structure
[i
].type
);
940 case GLSL_TYPE_SAMPLER
:
941 /* Samplers take up one slot in UNIFORMS[], but they're baked in
952 * In the initial pass of codegen, we assign temporary numbers to
953 * intermediate results. (not SSA -- variable assignments will reuse
957 glsl_to_tgsi_visitor::get_temp(const glsl_type
*type
)
961 src
.type
= native_integers
? type
->base_type
: GLSL_TYPE_FLOAT
;
962 src
.file
= PROGRAM_TEMPORARY
;
963 src
.index
= next_temp
;
965 next_temp
+= type_size(type
);
967 if (type
->is_array() || type
->is_record()) {
968 src
.swizzle
= SWIZZLE_NOOP
;
970 src
.swizzle
= swizzle_for_size(type
->vector_elements
);
978 glsl_to_tgsi_visitor::find_variable_storage(ir_variable
*var
)
981 variable_storage
*entry
;
983 foreach_iter(exec_list_iterator
, iter
, this->variables
) {
984 entry
= (variable_storage
*)iter
.get();
986 if (entry
->var
== var
)
994 glsl_to_tgsi_visitor::visit(ir_variable
*ir
)
996 if (strcmp(ir
->name
, "gl_FragCoord") == 0) {
997 struct gl_fragment_program
*fp
= (struct gl_fragment_program
*)this->prog
;
999 fp
->OriginUpperLeft
= ir
->origin_upper_left
;
1000 fp
->PixelCenterInteger
= ir
->pixel_center_integer
;
1002 } else if (strcmp(ir
->name
, "gl_FragDepth") == 0) {
1003 struct gl_fragment_program
*fp
= (struct gl_fragment_program
*)this->prog
;
1004 switch (ir
->depth_layout
) {
1005 case ir_depth_layout_none
:
1006 fp
->FragDepthLayout
= FRAG_DEPTH_LAYOUT_NONE
;
1008 case ir_depth_layout_any
:
1009 fp
->FragDepthLayout
= FRAG_DEPTH_LAYOUT_ANY
;
1011 case ir_depth_layout_greater
:
1012 fp
->FragDepthLayout
= FRAG_DEPTH_LAYOUT_GREATER
;
1014 case ir_depth_layout_less
:
1015 fp
->FragDepthLayout
= FRAG_DEPTH_LAYOUT_LESS
;
1017 case ir_depth_layout_unchanged
:
1018 fp
->FragDepthLayout
= FRAG_DEPTH_LAYOUT_UNCHANGED
;
1026 if (ir
->mode
== ir_var_uniform
&& strncmp(ir
->name
, "gl_", 3) == 0) {
1028 const ir_state_slot
*const slots
= ir
->state_slots
;
1029 assert(ir
->state_slots
!= NULL
);
1031 /* Check if this statevar's setup in the STATE file exactly
1032 * matches how we'll want to reference it as a
1033 * struct/array/whatever. If not, then we need to move it into
1034 * temporary storage and hope that it'll get copy-propagated
1037 for (i
= 0; i
< ir
->num_state_slots
; i
++) {
1038 if (slots
[i
].swizzle
!= SWIZZLE_XYZW
) {
1043 variable_storage
*storage
;
1045 if (i
== ir
->num_state_slots
) {
1046 /* We'll set the index later. */
1047 storage
= new(mem_ctx
) variable_storage(ir
, PROGRAM_STATE_VAR
, -1);
1048 this->variables
.push_tail(storage
);
1052 /* The variable_storage constructor allocates slots based on the size
1053 * of the type. However, this had better match the number of state
1054 * elements that we're going to copy into the new temporary.
1056 assert((int) ir
->num_state_slots
== type_size(ir
->type
));
1058 storage
= new(mem_ctx
) variable_storage(ir
, PROGRAM_TEMPORARY
,
1060 this->variables
.push_tail(storage
);
1061 this->next_temp
+= type_size(ir
->type
);
1063 dst
= st_dst_reg(st_src_reg(PROGRAM_TEMPORARY
, storage
->index
,
1064 native_integers
? ir
->type
->base_type
: GLSL_TYPE_FLOAT
));
1068 for (unsigned int i
= 0; i
< ir
->num_state_slots
; i
++) {
1069 int index
= _mesa_add_state_reference(this->prog
->Parameters
,
1070 (gl_state_index
*)slots
[i
].tokens
);
1072 if (storage
->file
== PROGRAM_STATE_VAR
) {
1073 if (storage
->index
== -1) {
1074 storage
->index
= index
;
1076 assert(index
== storage
->index
+ (int)i
);
1079 st_src_reg
src(PROGRAM_STATE_VAR
, index
,
1080 native_integers
? ir
->type
->base_type
: GLSL_TYPE_FLOAT
);
1081 src
.swizzle
= slots
[i
].swizzle
;
1082 emit(ir
, TGSI_OPCODE_MOV
, dst
, src
);
1083 /* even a float takes up a whole vec4 reg in a struct/array. */
1088 if (storage
->file
== PROGRAM_TEMPORARY
&&
1089 dst
.index
!= storage
->index
+ (int) ir
->num_state_slots
) {
1090 fail_link(this->shader_program
,
1091 "failed to load builtin uniform `%s' (%d/%d regs loaded)\n",
1092 ir
->name
, dst
.index
- storage
->index
,
1093 type_size(ir
->type
));
1099 glsl_to_tgsi_visitor::visit(ir_loop
*ir
)
1101 ir_dereference_variable
*counter
= NULL
;
1103 if (ir
->counter
!= NULL
)
1104 counter
= new(ir
) ir_dereference_variable(ir
->counter
);
1106 if (ir
->from
!= NULL
) {
1107 assert(ir
->counter
!= NULL
);
1109 ir_assignment
*a
= new(ir
) ir_assignment(counter
, ir
->from
, NULL
);
1115 emit(NULL
, TGSI_OPCODE_BGNLOOP
);
1119 new(ir
) ir_expression(ir
->cmp
, glsl_type::bool_type
,
1121 ir_if
*if_stmt
= new(ir
) ir_if(e
);
1123 ir_loop_jump
*brk
= new(ir
) ir_loop_jump(ir_loop_jump::jump_break
);
1125 if_stmt
->then_instructions
.push_tail(brk
);
1127 if_stmt
->accept(this);
1134 visit_exec_list(&ir
->body_instructions
, this);
1136 if (ir
->increment
) {
1138 new(ir
) ir_expression(ir_binop_add
, counter
->type
,
1139 counter
, ir
->increment
);
1141 ir_assignment
*a
= new(ir
) ir_assignment(counter
, e
, NULL
);
1148 emit(NULL
, TGSI_OPCODE_ENDLOOP
);
1152 glsl_to_tgsi_visitor::visit(ir_loop_jump
*ir
)
1155 case ir_loop_jump::jump_break
:
1156 emit(NULL
, TGSI_OPCODE_BRK
);
1158 case ir_loop_jump::jump_continue
:
1159 emit(NULL
, TGSI_OPCODE_CONT
);
1166 glsl_to_tgsi_visitor::visit(ir_function_signature
*ir
)
1173 glsl_to_tgsi_visitor::visit(ir_function
*ir
)
1175 /* Ignore function bodies other than main() -- we shouldn't see calls to
1176 * them since they should all be inlined before we get to glsl_to_tgsi.
1178 if (strcmp(ir
->name
, "main") == 0) {
1179 const ir_function_signature
*sig
;
1182 sig
= ir
->matching_signature(&empty
);
1186 foreach_iter(exec_list_iterator
, iter
, sig
->body
) {
1187 ir_instruction
*ir
= (ir_instruction
*)iter
.get();
1195 glsl_to_tgsi_visitor::try_emit_mad(ir_expression
*ir
, int mul_operand
)
1197 int nonmul_operand
= 1 - mul_operand
;
1199 st_dst_reg result_dst
;
1201 ir_expression
*expr
= ir
->operands
[mul_operand
]->as_expression();
1202 if (!expr
|| expr
->operation
!= ir_binop_mul
)
1205 expr
->operands
[0]->accept(this);
1207 expr
->operands
[1]->accept(this);
1209 ir
->operands
[nonmul_operand
]->accept(this);
1212 this->result
= get_temp(ir
->type
);
1213 result_dst
= st_dst_reg(this->result
);
1214 result_dst
.writemask
= (1 << ir
->type
->vector_elements
) - 1;
1215 emit(ir
, TGSI_OPCODE_MAD
, result_dst
, a
, b
, c
);
1221 * Emit MAD(a, -b, a) instead of AND(a, NOT(b))
1223 * The logic values are 1.0 for true and 0.0 for false. Logical-and is
1224 * implemented using multiplication, and logical-or is implemented using
1225 * addition. Logical-not can be implemented as (true - x), or (1.0 - x).
1226 * As result, the logical expression (a & !b) can be rewritten as:
1230 * - (a * 1) - (a * b)
1234 * This final expression can be implemented as a single MAD(a, -b, a)
1238 glsl_to_tgsi_visitor::try_emit_mad_for_and_not(ir_expression
*ir
, int try_operand
)
1240 const int other_operand
= 1 - try_operand
;
1243 ir_expression
*expr
= ir
->operands
[try_operand
]->as_expression();
1244 if (!expr
|| expr
->operation
!= ir_unop_logic_not
)
1247 ir
->operands
[other_operand
]->accept(this);
1249 expr
->operands
[0]->accept(this);
1252 b
.negate
= ~b
.negate
;
1254 this->result
= get_temp(ir
->type
);
1255 emit(ir
, TGSI_OPCODE_MAD
, st_dst_reg(this->result
), a
, b
, a
);
1261 glsl_to_tgsi_visitor::try_emit_sat(ir_expression
*ir
)
1263 /* Saturates were only introduced to vertex programs in
1264 * NV_vertex_program3, so don't give them to drivers in the VP.
1266 if (this->prog
->Target
== GL_VERTEX_PROGRAM_ARB
)
1269 ir_rvalue
*sat_src
= ir
->as_rvalue_to_saturate();
1273 sat_src
->accept(this);
1274 st_src_reg src
= this->result
;
1276 /* If we generated an expression instruction into a temporary in
1277 * processing the saturate's operand, apply the saturate to that
1278 * instruction. Otherwise, generate a MOV to do the saturate.
1280 * Note that we have to be careful to only do this optimization if
1281 * the instruction in question was what generated src->result. For
1282 * example, ir_dereference_array might generate a MUL instruction
1283 * to create the reladdr, and return us a src reg using that
1284 * reladdr. That MUL result is not the value we're trying to
1287 ir_expression
*sat_src_expr
= sat_src
->as_expression();
1288 if (sat_src_expr
&& (sat_src_expr
->operation
== ir_binop_mul
||
1289 sat_src_expr
->operation
== ir_binop_add
||
1290 sat_src_expr
->operation
== ir_binop_dot
)) {
1291 glsl_to_tgsi_instruction
*new_inst
;
1292 new_inst
= (glsl_to_tgsi_instruction
*)this->instructions
.get_tail();
1293 new_inst
->saturate
= true;
1295 this->result
= get_temp(ir
->type
);
1296 st_dst_reg result_dst
= st_dst_reg(this->result
);
1297 result_dst
.writemask
= (1 << ir
->type
->vector_elements
) - 1;
1298 glsl_to_tgsi_instruction
*inst
;
1299 inst
= emit(ir
, TGSI_OPCODE_MOV
, result_dst
, src
);
1300 inst
->saturate
= true;
1307 glsl_to_tgsi_visitor::reladdr_to_temp(ir_instruction
*ir
,
1308 st_src_reg
*reg
, int *num_reladdr
)
1313 emit_arl(ir
, address_reg
, *reg
->reladdr
);
1315 if (*num_reladdr
!= 1) {
1316 st_src_reg temp
= get_temp(glsl_type::vec4_type
);
1318 emit(ir
, TGSI_OPCODE_MOV
, st_dst_reg(temp
), *reg
);
1326 glsl_to_tgsi_visitor::visit(ir_expression
*ir
)
1328 unsigned int operand
;
1329 st_src_reg op
[Elements(ir
->operands
)];
1330 st_src_reg result_src
;
1331 st_dst_reg result_dst
;
1333 /* Quick peephole: Emit MAD(a, b, c) instead of ADD(MUL(a, b), c)
1335 if (ir
->operation
== ir_binop_add
) {
1336 if (try_emit_mad(ir
, 1))
1338 if (try_emit_mad(ir
, 0))
1342 /* Quick peephole: Emit OPCODE_MAD(-a, -b, a) instead of AND(a, NOT(b))
1344 if (ir
->operation
== ir_binop_logic_and
) {
1345 if (try_emit_mad_for_and_not(ir
, 1))
1347 if (try_emit_mad_for_and_not(ir
, 0))
1351 if (try_emit_sat(ir
))
1354 if (ir
->operation
== ir_quadop_vector
)
1355 assert(!"ir_quadop_vector should have been lowered");
1357 for (operand
= 0; operand
< ir
->get_num_operands(); operand
++) {
1358 this->result
.file
= PROGRAM_UNDEFINED
;
1359 ir
->operands
[operand
]->accept(this);
1360 if (this->result
.file
== PROGRAM_UNDEFINED
) {
1362 printf("Failed to get tree for expression operand:\n");
1363 ir
->operands
[operand
]->accept(&v
);
1366 op
[operand
] = this->result
;
1368 /* Matrix expression operands should have been broken down to vector
1369 * operations already.
1371 assert(!ir
->operands
[operand
]->type
->is_matrix());
1374 int vector_elements
= ir
->operands
[0]->type
->vector_elements
;
1375 if (ir
->operands
[1]) {
1376 vector_elements
= MAX2(vector_elements
,
1377 ir
->operands
[1]->type
->vector_elements
);
1380 this->result
.file
= PROGRAM_UNDEFINED
;
1382 /* Storage for our result. Ideally for an assignment we'd be using
1383 * the actual storage for the result here, instead.
1385 result_src
= get_temp(ir
->type
);
1386 /* convenience for the emit functions below. */
1387 result_dst
= st_dst_reg(result_src
);
1388 /* Limit writes to the channels that will be used by result_src later.
1389 * This does limit this temp's use as a temporary for multi-instruction
1392 result_dst
.writemask
= (1 << ir
->type
->vector_elements
) - 1;
1394 switch (ir
->operation
) {
1395 case ir_unop_logic_not
:
1396 if (result_dst
.type
!= GLSL_TYPE_FLOAT
)
1397 emit(ir
, TGSI_OPCODE_SEQ
, result_dst
, op
[0], st_src_reg_for_type(result_dst
.type
, 0));
1399 /* Previously 'SEQ dst, src, 0.0' was used for this. However, many
1400 * older GPUs implement SEQ using multiple instructions (i915 uses two
1401 * SGE instructions and a MUL instruction). Since our logic values are
1402 * 0.0 and 1.0, 1-x also implements !x.
1404 op
[0].negate
= ~op
[0].negate
;
1405 emit(ir
, TGSI_OPCODE_ADD
, result_dst
, op
[0], st_src_reg_for_float(1.0));
1409 assert(result_dst
.type
== GLSL_TYPE_FLOAT
|| result_dst
.type
== GLSL_TYPE_INT
);
1410 if (result_dst
.type
== GLSL_TYPE_INT
)
1411 emit(ir
, TGSI_OPCODE_INEG
, result_dst
, op
[0]);
1413 op
[0].negate
= ~op
[0].negate
;
1418 assert(result_dst
.type
== GLSL_TYPE_FLOAT
);
1419 emit(ir
, TGSI_OPCODE_ABS
, result_dst
, op
[0]);
1422 emit(ir
, TGSI_OPCODE_SSG
, result_dst
, op
[0]);
1425 emit_scalar(ir
, TGSI_OPCODE_RCP
, result_dst
, op
[0]);
1429 emit_scalar(ir
, TGSI_OPCODE_EX2
, result_dst
, op
[0]);
1433 assert(!"not reached: should be handled by ir_explog_to_explog2");
1436 emit_scalar(ir
, TGSI_OPCODE_LG2
, result_dst
, op
[0]);
1439 emit_scalar(ir
, TGSI_OPCODE_SIN
, result_dst
, op
[0]);
1442 emit_scalar(ir
, TGSI_OPCODE_COS
, result_dst
, op
[0]);
1444 case ir_unop_sin_reduced
:
1445 emit_scs(ir
, TGSI_OPCODE_SIN
, result_dst
, op
[0]);
1447 case ir_unop_cos_reduced
:
1448 emit_scs(ir
, TGSI_OPCODE_COS
, result_dst
, op
[0]);
1452 emit(ir
, TGSI_OPCODE_DDX
, result_dst
, op
[0]);
1455 op
[0].negate
= ~op
[0].negate
;
1456 emit(ir
, TGSI_OPCODE_DDY
, result_dst
, op
[0]);
1459 case ir_unop_noise
: {
1460 /* At some point, a motivated person could add a better
1461 * implementation of noise. Currently not even the nvidia
1462 * binary drivers do anything more than this. In any case, the
1463 * place to do this is in the GL state tracker, not the poor
1466 emit(ir
, TGSI_OPCODE_MOV
, result_dst
, st_src_reg_for_float(0.5));
1471 emit(ir
, TGSI_OPCODE_ADD
, result_dst
, op
[0], op
[1]);
1474 emit(ir
, TGSI_OPCODE_SUB
, result_dst
, op
[0], op
[1]);
1478 emit(ir
, TGSI_OPCODE_MUL
, result_dst
, op
[0], op
[1]);
1481 if (result_dst
.type
== GLSL_TYPE_FLOAT
)
1482 assert(!"not reached: should be handled by ir_div_to_mul_rcp");
1484 emit(ir
, TGSI_OPCODE_DIV
, result_dst
, op
[0], op
[1]);
1487 if (result_dst
.type
== GLSL_TYPE_FLOAT
)
1488 assert(!"ir_binop_mod should have been converted to b * fract(a/b)");
1490 emit(ir
, TGSI_OPCODE_MOD
, result_dst
, op
[0], op
[1]);
1494 emit(ir
, TGSI_OPCODE_SLT
, result_dst
, op
[0], op
[1]);
1496 case ir_binop_greater
:
1497 emit(ir
, TGSI_OPCODE_SGT
, result_dst
, op
[0], op
[1]);
1499 case ir_binop_lequal
:
1500 emit(ir
, TGSI_OPCODE_SLE
, result_dst
, op
[0], op
[1]);
1502 case ir_binop_gequal
:
1503 emit(ir
, TGSI_OPCODE_SGE
, result_dst
, op
[0], op
[1]);
1505 case ir_binop_equal
:
1506 emit(ir
, TGSI_OPCODE_SEQ
, result_dst
, op
[0], op
[1]);
1508 case ir_binop_nequal
:
1509 emit(ir
, TGSI_OPCODE_SNE
, result_dst
, op
[0], op
[1]);
1511 case ir_binop_all_equal
:
1512 /* "==" operator producing a scalar boolean. */
1513 if (ir
->operands
[0]->type
->is_vector() ||
1514 ir
->operands
[1]->type
->is_vector()) {
1515 st_src_reg temp
= get_temp(native_integers
?
1516 glsl_type::get_instance(ir
->operands
[0]->type
->base_type
, 4, 1) :
1517 glsl_type::vec4_type
);
1518 assert(ir
->operands
[0]->type
->base_type
== GLSL_TYPE_FLOAT
);
1519 emit(ir
, TGSI_OPCODE_SNE
, st_dst_reg(temp
), op
[0], op
[1]);
1521 /* After the dot-product, the value will be an integer on the
1522 * range [0,4]. Zero becomes 1.0, and positive values become zero.
1524 emit_dp(ir
, result_dst
, temp
, temp
, vector_elements
);
1526 if (result_dst
.type
== GLSL_TYPE_FLOAT
) {
1527 /* Negating the result of the dot-product gives values on the range
1528 * [-4, 0]. Zero becomes 1.0, and negative values become zero.
1529 * This is achieved using SGE.
1531 st_src_reg sge_src
= result_src
;
1532 sge_src
.negate
= ~sge_src
.negate
;
1533 emit(ir
, TGSI_OPCODE_SGE
, result_dst
, sge_src
, st_src_reg_for_float(0.0));
1535 /* The TGSI negate flag doesn't work for integers, so use SEQ 0
1538 emit(ir
, TGSI_OPCODE_SEQ
, result_dst
, result_src
, st_src_reg_for_int(0));
1541 emit(ir
, TGSI_OPCODE_SEQ
, result_dst
, op
[0], op
[1]);
1544 case ir_binop_any_nequal
:
1545 /* "!=" operator producing a scalar boolean. */
1546 if (ir
->operands
[0]->type
->is_vector() ||
1547 ir
->operands
[1]->type
->is_vector()) {
1548 st_src_reg temp
= get_temp(native_integers
?
1549 glsl_type::get_instance(ir
->operands
[0]->type
->base_type
, 4, 1) :
1550 glsl_type::vec4_type
);
1551 assert(ir
->operands
[0]->type
->base_type
== GLSL_TYPE_FLOAT
);
1552 emit(ir
, TGSI_OPCODE_SNE
, st_dst_reg(temp
), op
[0], op
[1]);
1554 /* After the dot-product, the value will be an integer on the
1555 * range [0,4]. Zero stays zero, and positive values become 1.0.
1557 glsl_to_tgsi_instruction
*const dp
=
1558 emit_dp(ir
, result_dst
, temp
, temp
, vector_elements
);
1559 if (this->prog
->Target
== GL_FRAGMENT_PROGRAM_ARB
&&
1560 result_dst
.type
== GLSL_TYPE_FLOAT
) {
1561 /* The clamping to [0,1] can be done for free in the fragment
1562 * shader with a saturate.
1564 dp
->saturate
= true;
1565 } else if (result_dst
.type
== GLSL_TYPE_FLOAT
) {
1566 /* Negating the result of the dot-product gives values on the range
1567 * [-4, 0]. Zero stays zero, and negative values become 1.0. This
1568 * achieved using SLT.
1570 st_src_reg slt_src
= result_src
;
1571 slt_src
.negate
= ~slt_src
.negate
;
1572 emit(ir
, TGSI_OPCODE_SLT
, result_dst
, slt_src
, st_src_reg_for_float(0.0));
1574 emit(ir
, TGSI_OPCODE_SNE
, result_dst
, result_src
, st_src_reg_for_int(0));
1577 emit(ir
, TGSI_OPCODE_SNE
, result_dst
, op
[0], op
[1]);
1582 assert(ir
->operands
[0]->type
->is_vector());
1584 /* After the dot-product, the value will be an integer on the
1585 * range [0,4]. Zero stays zero, and positive values become 1.0.
1587 glsl_to_tgsi_instruction
*const dp
=
1588 emit_dp(ir
, result_dst
, op
[0], op
[0],
1589 ir
->operands
[0]->type
->vector_elements
);
1590 if (this->prog
->Target
== GL_FRAGMENT_PROGRAM_ARB
&&
1591 result_dst
.type
== GLSL_TYPE_FLOAT
) {
1592 /* The clamping to [0,1] can be done for free in the fragment
1593 * shader with a saturate.
1595 dp
->saturate
= true;
1596 } else if (result_dst
.type
== GLSL_TYPE_FLOAT
) {
1597 /* Negating the result of the dot-product gives values on the range
1598 * [-4, 0]. Zero stays zero, and negative values become 1.0. This
1599 * is achieved using SLT.
1601 st_src_reg slt_src
= result_src
;
1602 slt_src
.negate
= ~slt_src
.negate
;
1603 emit(ir
, TGSI_OPCODE_SLT
, result_dst
, slt_src
, st_src_reg_for_float(0.0));
1606 /* Use SNE 0 if integers are being used as boolean values. */
1607 emit(ir
, TGSI_OPCODE_SNE
, result_dst
, result_src
, st_src_reg_for_int(0));
1612 case ir_binop_logic_xor
:
1613 emit(ir
, TGSI_OPCODE_SNE
, result_dst
, op
[0], op
[1]);
1616 case ir_binop_logic_or
: {
1617 /* After the addition, the value will be an integer on the
1618 * range [0,2]. Zero stays zero, and positive values become 1.0.
1620 glsl_to_tgsi_instruction
*add
=
1621 emit(ir
, TGSI_OPCODE_ADD
, result_dst
, op
[0], op
[1]);
1622 if (this->prog
->Target
== GL_FRAGMENT_PROGRAM_ARB
&&
1623 result_dst
.type
== GLSL_TYPE_FLOAT
) {
1624 /* The clamping to [0,1] can be done for free in the fragment
1625 * shader with a saturate if floats are being used as boolean values.
1627 add
->saturate
= true;
1628 } else if (result_dst
.type
== GLSL_TYPE_FLOAT
) {
1629 /* Negating the result of the addition gives values on the range
1630 * [-2, 0]. Zero stays zero, and negative values become 1.0. This
1631 * is achieved using SLT.
1633 st_src_reg slt_src
= result_src
;
1634 slt_src
.negate
= ~slt_src
.negate
;
1635 emit(ir
, TGSI_OPCODE_SLT
, result_dst
, slt_src
, st_src_reg_for_float(0.0));
1637 /* Use an SNE on the result of the addition. Zero stays zero,
1638 * 1 stays 1, and 2 becomes 1.
1640 emit(ir
, TGSI_OPCODE_SNE
, result_dst
, result_src
, st_src_reg_for_int(0));
1645 case ir_binop_logic_and
:
1646 /* the bool args are stored as float 0.0 or 1.0, so "mul" gives us "and". */
1647 emit(ir
, TGSI_OPCODE_MUL
, result_dst
, op
[0], op
[1]);
1651 assert(ir
->operands
[0]->type
->is_vector());
1652 assert(ir
->operands
[0]->type
== ir
->operands
[1]->type
);
1653 emit_dp(ir
, result_dst
, op
[0], op
[1],
1654 ir
->operands
[0]->type
->vector_elements
);
1658 /* sqrt(x) = x * rsq(x). */
1659 emit_scalar(ir
, TGSI_OPCODE_RSQ
, result_dst
, op
[0]);
1660 emit(ir
, TGSI_OPCODE_MUL
, result_dst
, result_src
, op
[0]);
1661 /* For incoming channels <= 0, set the result to 0. */
1662 op
[0].negate
= ~op
[0].negate
;
1663 emit(ir
, TGSI_OPCODE_CMP
, result_dst
,
1664 op
[0], result_src
, st_src_reg_for_float(0.0));
1667 emit_scalar(ir
, TGSI_OPCODE_RSQ
, result_dst
, op
[0]);
1671 if (native_integers
) {
1672 emit(ir
, TGSI_OPCODE_I2F
, result_dst
, op
[0]);
1677 /* Converting between signed and unsigned integers is a no-op. */
1679 /* Booleans are stored as integers (or floats in GLSL 1.20 and lower). */
1683 if (native_integers
)
1684 emit(ir
, TGSI_OPCODE_F2I
, result_dst
, op
[0]);
1686 emit(ir
, TGSI_OPCODE_TRUNC
, result_dst
, op
[0]);
1690 emit(ir
, TGSI_OPCODE_SNE
, result_dst
, op
[0],
1691 st_src_reg_for_type(result_dst
.type
, 0));
1694 emit(ir
, TGSI_OPCODE_TRUNC
, result_dst
, op
[0]);
1697 op
[0].negate
= ~op
[0].negate
;
1698 emit(ir
, TGSI_OPCODE_FLR
, result_dst
, op
[0]);
1699 result_src
.negate
= ~result_src
.negate
;
1702 emit(ir
, TGSI_OPCODE_FLR
, result_dst
, op
[0]);
1705 emit(ir
, TGSI_OPCODE_FRC
, result_dst
, op
[0]);
1709 emit(ir
, TGSI_OPCODE_MIN
, result_dst
, op
[0], op
[1]);
1712 emit(ir
, TGSI_OPCODE_MAX
, result_dst
, op
[0], op
[1]);
1715 emit_scalar(ir
, TGSI_OPCODE_POW
, result_dst
, op
[0], op
[1]);
1718 case ir_unop_bit_not
:
1719 if (glsl_version
>= 130) {
1720 emit(ir
, TGSI_OPCODE_NOT
, result_dst
, op
[0]);
1724 if (native_integers
) {
1725 emit(ir
, TGSI_OPCODE_U2F
, result_dst
, op
[0]);
1728 case ir_binop_lshift
:
1729 if (glsl_version
>= 130) {
1730 emit(ir
, TGSI_OPCODE_SHL
, result_dst
, op
[0]);
1733 case ir_binop_rshift
:
1734 if (glsl_version
>= 130) {
1735 emit(ir
, TGSI_OPCODE_ISHR
, result_dst
, op
[0]);
1738 case ir_binop_bit_and
:
1739 if (glsl_version
>= 130) {
1740 emit(ir
, TGSI_OPCODE_AND
, result_dst
, op
[0]);
1743 case ir_binop_bit_xor
:
1744 if (glsl_version
>= 130) {
1745 emit(ir
, TGSI_OPCODE_XOR
, result_dst
, op
[0]);
1748 case ir_binop_bit_or
:
1749 if (glsl_version
>= 130) {
1750 emit(ir
, TGSI_OPCODE_OR
, result_dst
, op
[0]);
1753 case ir_unop_round_even
:
1754 assert(!"GLSL 1.30 features unsupported");
1757 case ir_quadop_vector
:
1758 /* This operation should have already been handled.
1760 assert(!"Should not get here.");
1764 this->result
= result_src
;
1769 glsl_to_tgsi_visitor::visit(ir_swizzle
*ir
)
1775 /* Note that this is only swizzles in expressions, not those on the left
1776 * hand side of an assignment, which do write masking. See ir_assignment
1780 ir
->val
->accept(this);
1782 assert(src
.file
!= PROGRAM_UNDEFINED
);
1784 for (i
= 0; i
< 4; i
++) {
1785 if (i
< ir
->type
->vector_elements
) {
1788 swizzle
[i
] = GET_SWZ(src
.swizzle
, ir
->mask
.x
);
1791 swizzle
[i
] = GET_SWZ(src
.swizzle
, ir
->mask
.y
);
1794 swizzle
[i
] = GET_SWZ(src
.swizzle
, ir
->mask
.z
);
1797 swizzle
[i
] = GET_SWZ(src
.swizzle
, ir
->mask
.w
);
1801 /* If the type is smaller than a vec4, replicate the last
1804 swizzle
[i
] = swizzle
[ir
->type
->vector_elements
- 1];
1808 src
.swizzle
= MAKE_SWIZZLE4(swizzle
[0], swizzle
[1], swizzle
[2], swizzle
[3]);
1814 glsl_to_tgsi_visitor::visit(ir_dereference_variable
*ir
)
1816 variable_storage
*entry
= find_variable_storage(ir
->var
);
1817 ir_variable
*var
= ir
->var
;
1820 switch (var
->mode
) {
1821 case ir_var_uniform
:
1822 entry
= new(mem_ctx
) variable_storage(var
, PROGRAM_UNIFORM
,
1824 this->variables
.push_tail(entry
);
1828 /* The linker assigns locations for varyings and attributes,
1829 * including deprecated builtins (like gl_Color), user-assign
1830 * generic attributes (glBindVertexLocation), and
1831 * user-defined varyings.
1833 * FINISHME: We would hit this path for function arguments. Fix!
1835 assert(var
->location
!= -1);
1836 entry
= new(mem_ctx
) variable_storage(var
,
1839 if (this->prog
->Target
== GL_VERTEX_PROGRAM_ARB
&&
1840 var
->location
>= VERT_ATTRIB_GENERIC0
) {
1841 _mesa_add_attribute(this->prog
->Attributes
,
1843 _mesa_sizeof_glsl_type(var
->type
->gl_type
),
1845 var
->location
- VERT_ATTRIB_GENERIC0
);
1849 assert(var
->location
!= -1);
1850 entry
= new(mem_ctx
) variable_storage(var
,
1854 case ir_var_system_value
:
1855 entry
= new(mem_ctx
) variable_storage(var
,
1856 PROGRAM_SYSTEM_VALUE
,
1860 case ir_var_temporary
:
1861 entry
= new(mem_ctx
) variable_storage(var
, PROGRAM_TEMPORARY
,
1863 this->variables
.push_tail(entry
);
1865 next_temp
+= type_size(var
->type
);
1870 printf("Failed to make storage for %s\n", var
->name
);
1875 this->result
= st_src_reg(entry
->file
, entry
->index
, var
->type
);
1876 if (!native_integers
)
1877 this->result
.type
= GLSL_TYPE_FLOAT
;
1881 glsl_to_tgsi_visitor::visit(ir_dereference_array
*ir
)
1885 int element_size
= type_size(ir
->type
);
1887 index
= ir
->array_index
->constant_expression_value();
1889 ir
->array
->accept(this);
1893 src
.index
+= index
->value
.i
[0] * element_size
;
1895 /* Variable index array dereference. It eats the "vec4" of the
1896 * base of the array and an index that offsets the TGSI register
1899 ir
->array_index
->accept(this);
1901 st_src_reg index_reg
;
1903 if (element_size
== 1) {
1904 index_reg
= this->result
;
1906 index_reg
= get_temp(glsl_type::float_type
);
1908 emit(ir
, TGSI_OPCODE_MUL
, st_dst_reg(index_reg
),
1909 this->result
, st_src_reg_for_float(element_size
));
1912 /* If there was already a relative address register involved, add the
1913 * new and the old together to get the new offset.
1915 if (src
.reladdr
!= NULL
) {
1916 st_src_reg accum_reg
= get_temp(glsl_type::float_type
);
1918 emit(ir
, TGSI_OPCODE_ADD
, st_dst_reg(accum_reg
),
1919 index_reg
, *src
.reladdr
);
1921 index_reg
= accum_reg
;
1924 src
.reladdr
= ralloc(mem_ctx
, st_src_reg
);
1925 memcpy(src
.reladdr
, &index_reg
, sizeof(index_reg
));
1928 /* If the type is smaller than a vec4, replicate the last channel out. */
1929 if (ir
->type
->is_scalar() || ir
->type
->is_vector())
1930 src
.swizzle
= swizzle_for_size(ir
->type
->vector_elements
);
1932 src
.swizzle
= SWIZZLE_NOOP
;
1938 glsl_to_tgsi_visitor::visit(ir_dereference_record
*ir
)
1941 const glsl_type
*struct_type
= ir
->record
->type
;
1944 ir
->record
->accept(this);
1946 for (i
= 0; i
< struct_type
->length
; i
++) {
1947 if (strcmp(struct_type
->fields
.structure
[i
].name
, ir
->field
) == 0)
1949 offset
+= type_size(struct_type
->fields
.structure
[i
].type
);
1952 /* If the type is smaller than a vec4, replicate the last channel out. */
1953 if (ir
->type
->is_scalar() || ir
->type
->is_vector())
1954 this->result
.swizzle
= swizzle_for_size(ir
->type
->vector_elements
);
1956 this->result
.swizzle
= SWIZZLE_NOOP
;
1958 this->result
.index
+= offset
;
1962 * We want to be careful in assignment setup to hit the actual storage
1963 * instead of potentially using a temporary like we might with the
1964 * ir_dereference handler.
1967 get_assignment_lhs(ir_dereference
*ir
, glsl_to_tgsi_visitor
*v
)
1969 /* The LHS must be a dereference. If the LHS is a variable indexed array
1970 * access of a vector, it must be separated into a series conditional moves
1971 * before reaching this point (see ir_vec_index_to_cond_assign).
1973 assert(ir
->as_dereference());
1974 ir_dereference_array
*deref_array
= ir
->as_dereference_array();
1976 assert(!deref_array
->array
->type
->is_vector());
1979 /* Use the rvalue deref handler for the most part. We'll ignore
1980 * swizzles in it and write swizzles using writemask, though.
1983 return st_dst_reg(v
->result
);
1987 * Process the condition of a conditional assignment
1989 * Examines the condition of a conditional assignment to generate the optimal
1990 * first operand of a \c CMP instruction. If the condition is a relational
1991 * operator with 0 (e.g., \c ir_binop_less), the value being compared will be
1992 * used as the source for the \c CMP instruction. Otherwise the comparison
1993 * is processed to a boolean result, and the boolean result is used as the
1994 * operand to the CMP instruction.
1997 glsl_to_tgsi_visitor::process_move_condition(ir_rvalue
*ir
)
1999 ir_rvalue
*src_ir
= ir
;
2001 bool switch_order
= false;
2003 ir_expression
*const expr
= ir
->as_expression();
2004 if ((expr
!= NULL
) && (expr
->get_num_operands() == 2)) {
2005 bool zero_on_left
= false;
2007 if (expr
->operands
[0]->is_zero()) {
2008 src_ir
= expr
->operands
[1];
2009 zero_on_left
= true;
2010 } else if (expr
->operands
[1]->is_zero()) {
2011 src_ir
= expr
->operands
[0];
2012 zero_on_left
= false;
2016 * (a < 0) T F F ( a < 0) T F F
2017 * (0 < a) F F T (-a < 0) F F T
2018 * (a <= 0) T T F (-a < 0) F F T (swap order of other operands)
2019 * (0 <= a) F T T ( a < 0) T F F (swap order of other operands)
2020 * (a > 0) F F T (-a < 0) F F T
2021 * (0 > a) T F F ( a < 0) T F F
2022 * (a >= 0) F T T ( a < 0) T F F (swap order of other operands)
2023 * (0 >= a) T T F (-a < 0) F F T (swap order of other operands)
2025 * Note that exchanging the order of 0 and 'a' in the comparison simply
2026 * means that the value of 'a' should be negated.
2029 switch (expr
->operation
) {
2031 switch_order
= false;
2032 negate
= zero_on_left
;
2035 case ir_binop_greater
:
2036 switch_order
= false;
2037 negate
= !zero_on_left
;
2040 case ir_binop_lequal
:
2041 switch_order
= true;
2042 negate
= !zero_on_left
;
2045 case ir_binop_gequal
:
2046 switch_order
= true;
2047 negate
= zero_on_left
;
2051 /* This isn't the right kind of comparison afterall, so make sure
2052 * the whole condition is visited.
2060 src_ir
->accept(this);
2062 /* We use the TGSI_OPCODE_CMP (a < 0 ? b : c) for conditional moves, and the
2063 * condition we produced is 0.0 or 1.0. By flipping the sign, we can
2064 * choose which value TGSI_OPCODE_CMP produces without an extra instruction
2065 * computing the condition.
2068 this->result
.negate
= ~this->result
.negate
;
2070 return switch_order
;
2074 glsl_to_tgsi_visitor::visit(ir_assignment
*ir
)
2080 ir
->rhs
->accept(this);
2083 l
= get_assignment_lhs(ir
->lhs
, this);
2085 /* FINISHME: This should really set to the correct maximal writemask for each
2086 * FINISHME: component written (in the loops below). This case can only
2087 * FINISHME: occur for matrices, arrays, and structures.
2089 if (ir
->write_mask
== 0) {
2090 assert(!ir
->lhs
->type
->is_scalar() && !ir
->lhs
->type
->is_vector());
2091 l
.writemask
= WRITEMASK_XYZW
;
2092 } else if (ir
->lhs
->type
->is_scalar() &&
2093 ir
->lhs
->variable_referenced()->mode
== ir_var_out
) {
2094 /* FINISHME: This hack makes writing to gl_FragDepth, which lives in the
2095 * FINISHME: W component of fragment shader output zero, work correctly.
2097 l
.writemask
= WRITEMASK_XYZW
;
2100 int first_enabled_chan
= 0;
2103 l
.writemask
= ir
->write_mask
;
2105 for (int i
= 0; i
< 4; i
++) {
2106 if (l
.writemask
& (1 << i
)) {
2107 first_enabled_chan
= GET_SWZ(r
.swizzle
, i
);
2112 /* Swizzle a small RHS vector into the channels being written.
2114 * glsl ir treats write_mask as dictating how many channels are
2115 * present on the RHS while TGSI treats write_mask as just
2116 * showing which channels of the vec4 RHS get written.
2118 for (int i
= 0; i
< 4; i
++) {
2119 if (l
.writemask
& (1 << i
))
2120 swizzles
[i
] = GET_SWZ(r
.swizzle
, rhs_chan
++);
2122 swizzles
[i
] = first_enabled_chan
;
2124 r
.swizzle
= MAKE_SWIZZLE4(swizzles
[0], swizzles
[1],
2125 swizzles
[2], swizzles
[3]);
2128 assert(l
.file
!= PROGRAM_UNDEFINED
);
2129 assert(r
.file
!= PROGRAM_UNDEFINED
);
2131 if (ir
->condition
) {
2132 const bool switch_order
= this->process_move_condition(ir
->condition
);
2133 st_src_reg condition
= this->result
;
2135 for (i
= 0; i
< type_size(ir
->lhs
->type
); i
++) {
2136 st_src_reg l_src
= st_src_reg(l
);
2137 l_src
.swizzle
= swizzle_for_size(ir
->lhs
->type
->vector_elements
);
2140 emit(ir
, TGSI_OPCODE_CMP
, l
, condition
, l_src
, r
);
2142 emit(ir
, TGSI_OPCODE_CMP
, l
, condition
, r
, l_src
);
2148 } else if (ir
->rhs
->as_expression() &&
2149 this->instructions
.get_tail() &&
2150 ir
->rhs
== ((glsl_to_tgsi_instruction
*)this->instructions
.get_tail())->ir
&&
2151 type_size(ir
->lhs
->type
) == 1 &&
2152 l
.writemask
== ((glsl_to_tgsi_instruction
*)this->instructions
.get_tail())->dst
.writemask
) {
2153 /* To avoid emitting an extra MOV when assigning an expression to a
2154 * variable, emit the last instruction of the expression again, but
2155 * replace the destination register with the target of the assignment.
2156 * Dead code elimination will remove the original instruction.
2158 glsl_to_tgsi_instruction
*inst
, *new_inst
;
2159 inst
= (glsl_to_tgsi_instruction
*)this->instructions
.get_tail();
2160 new_inst
= emit(ir
, inst
->op
, l
, inst
->src
[0], inst
->src
[1], inst
->src
[2]);
2161 new_inst
->saturate
= inst
->saturate
;
2163 for (i
= 0; i
< type_size(ir
->lhs
->type
); i
++) {
2164 emit(ir
, TGSI_OPCODE_MOV
, l
, r
);
2173 glsl_to_tgsi_visitor::visit(ir_constant
*ir
)
2176 GLfloat stack_vals
[4] = { 0 };
2177 gl_constant_value
*values
= (gl_constant_value
*) stack_vals
;
2178 GLenum gl_type
= GL_NONE
;
2180 static int in_array
= 0;
2181 gl_register_file file
= in_array
? PROGRAM_CONSTANT
: PROGRAM_IMMEDIATE
;
2183 /* Unfortunately, 4 floats is all we can get into
2184 * _mesa_add_typed_unnamed_constant. So, make a temp to store an
2185 * aggregate constant and move each constant value into it. If we
2186 * get lucky, copy propagation will eliminate the extra moves.
2188 if (ir
->type
->base_type
== GLSL_TYPE_STRUCT
) {
2189 st_src_reg temp_base
= get_temp(ir
->type
);
2190 st_dst_reg temp
= st_dst_reg(temp_base
);
2192 foreach_iter(exec_list_iterator
, iter
, ir
->components
) {
2193 ir_constant
*field_value
= (ir_constant
*)iter
.get();
2194 int size
= type_size(field_value
->type
);
2198 field_value
->accept(this);
2201 for (i
= 0; i
< (unsigned int)size
; i
++) {
2202 emit(ir
, TGSI_OPCODE_MOV
, temp
, src
);
2208 this->result
= temp_base
;
2212 if (ir
->type
->is_array()) {
2213 st_src_reg temp_base
= get_temp(ir
->type
);
2214 st_dst_reg temp
= st_dst_reg(temp_base
);
2215 int size
= type_size(ir
->type
->fields
.array
);
2220 for (i
= 0; i
< ir
->type
->length
; i
++) {
2221 ir
->array_elements
[i
]->accept(this);
2223 for (int j
= 0; j
< size
; j
++) {
2224 emit(ir
, TGSI_OPCODE_MOV
, temp
, src
);
2230 this->result
= temp_base
;
2235 if (ir
->type
->is_matrix()) {
2236 st_src_reg mat
= get_temp(ir
->type
);
2237 st_dst_reg mat_column
= st_dst_reg(mat
);
2239 for (i
= 0; i
< ir
->type
->matrix_columns
; i
++) {
2240 assert(ir
->type
->base_type
== GLSL_TYPE_FLOAT
);
2241 values
= (gl_constant_value
*) &ir
->value
.f
[i
* ir
->type
->vector_elements
];
2243 src
= st_src_reg(file
, -1, ir
->type
->base_type
);
2244 src
.index
= add_constant(file
,
2246 ir
->type
->vector_elements
,
2249 emit(ir
, TGSI_OPCODE_MOV
, mat_column
, src
);
2258 switch (ir
->type
->base_type
) {
2259 case GLSL_TYPE_FLOAT
:
2261 for (i
= 0; i
< ir
->type
->vector_elements
; i
++) {
2262 values
[i
].f
= ir
->value
.f
[i
];
2265 case GLSL_TYPE_UINT
:
2266 gl_type
= native_integers
? GL_UNSIGNED_INT
: GL_FLOAT
;
2267 for (i
= 0; i
< ir
->type
->vector_elements
; i
++) {
2268 if (native_integers
)
2269 values
[i
].u
= ir
->value
.u
[i
];
2271 values
[i
].f
= ir
->value
.u
[i
];
2275 gl_type
= native_integers
? GL_INT
: GL_FLOAT
;
2276 for (i
= 0; i
< ir
->type
->vector_elements
; i
++) {
2277 if (native_integers
)
2278 values
[i
].i
= ir
->value
.i
[i
];
2280 values
[i
].f
= ir
->value
.i
[i
];
2283 case GLSL_TYPE_BOOL
:
2284 gl_type
= native_integers
? GL_BOOL
: GL_FLOAT
;
2285 for (i
= 0; i
< ir
->type
->vector_elements
; i
++) {
2286 if (native_integers
)
2287 values
[i
].b
= ir
->value
.b
[i
];
2289 values
[i
].f
= ir
->value
.b
[i
];
2293 assert(!"Non-float/uint/int/bool constant");
2296 this->result
= st_src_reg(file
, -1, ir
->type
);
2297 this->result
.index
= add_constant(file
,
2299 ir
->type
->vector_elements
,
2301 &this->result
.swizzle
);
2305 glsl_to_tgsi_visitor::get_function_signature(ir_function_signature
*sig
)
2307 function_entry
*entry
;
2309 foreach_iter(exec_list_iterator
, iter
, this->function_signatures
) {
2310 entry
= (function_entry
*)iter
.get();
2312 if (entry
->sig
== sig
)
2316 entry
= ralloc(mem_ctx
, function_entry
);
2318 entry
->sig_id
= this->next_signature_id
++;
2319 entry
->bgn_inst
= NULL
;
2321 /* Allocate storage for all the parameters. */
2322 foreach_iter(exec_list_iterator
, iter
, sig
->parameters
) {
2323 ir_variable
*param
= (ir_variable
*)iter
.get();
2324 variable_storage
*storage
;
2326 storage
= find_variable_storage(param
);
2329 storage
= new(mem_ctx
) variable_storage(param
, PROGRAM_TEMPORARY
,
2331 this->variables
.push_tail(storage
);
2333 this->next_temp
+= type_size(param
->type
);
2336 if (!sig
->return_type
->is_void()) {
2337 entry
->return_reg
= get_temp(sig
->return_type
);
2339 entry
->return_reg
= undef_src
;
2342 this->function_signatures
.push_tail(entry
);
2347 glsl_to_tgsi_visitor::visit(ir_call
*ir
)
2349 glsl_to_tgsi_instruction
*call_inst
;
2350 ir_function_signature
*sig
= ir
->get_callee();
2351 function_entry
*entry
= get_function_signature(sig
);
2354 /* Process in parameters. */
2355 exec_list_iterator sig_iter
= sig
->parameters
.iterator();
2356 foreach_iter(exec_list_iterator
, iter
, *ir
) {
2357 ir_rvalue
*param_rval
= (ir_rvalue
*)iter
.get();
2358 ir_variable
*param
= (ir_variable
*)sig_iter
.get();
2360 if (param
->mode
== ir_var_in
||
2361 param
->mode
== ir_var_inout
) {
2362 variable_storage
*storage
= find_variable_storage(param
);
2365 param_rval
->accept(this);
2366 st_src_reg r
= this->result
;
2369 l
.file
= storage
->file
;
2370 l
.index
= storage
->index
;
2372 l
.writemask
= WRITEMASK_XYZW
;
2373 l
.cond_mask
= COND_TR
;
2375 for (i
= 0; i
< type_size(param
->type
); i
++) {
2376 emit(ir
, TGSI_OPCODE_MOV
, l
, r
);
2384 assert(!sig_iter
.has_next());
2386 /* Emit call instruction */
2387 call_inst
= emit(ir
, TGSI_OPCODE_CAL
);
2388 call_inst
->function
= entry
;
2390 /* Process out parameters. */
2391 sig_iter
= sig
->parameters
.iterator();
2392 foreach_iter(exec_list_iterator
, iter
, *ir
) {
2393 ir_rvalue
*param_rval
= (ir_rvalue
*)iter
.get();
2394 ir_variable
*param
= (ir_variable
*)sig_iter
.get();
2396 if (param
->mode
== ir_var_out
||
2397 param
->mode
== ir_var_inout
) {
2398 variable_storage
*storage
= find_variable_storage(param
);
2402 r
.file
= storage
->file
;
2403 r
.index
= storage
->index
;
2405 r
.swizzle
= SWIZZLE_NOOP
;
2408 param_rval
->accept(this);
2409 st_dst_reg l
= st_dst_reg(this->result
);
2411 for (i
= 0; i
< type_size(param
->type
); i
++) {
2412 emit(ir
, TGSI_OPCODE_MOV
, l
, r
);
2420 assert(!sig_iter
.has_next());
2422 /* Process return value. */
2423 this->result
= entry
->return_reg
;
2427 glsl_to_tgsi_visitor::visit(ir_texture
*ir
)
2429 st_src_reg result_src
, coord
, lod_info
, projector
, dx
, dy
, offset
;
2430 st_dst_reg result_dst
, coord_dst
;
2431 glsl_to_tgsi_instruction
*inst
= NULL
;
2432 unsigned opcode
= TGSI_OPCODE_NOP
;
2434 if (ir
->coordinate
) {
2435 ir
->coordinate
->accept(this);
2437 /* Put our coords in a temp. We'll need to modify them for shadow,
2438 * projection, or LOD, so the only case we'd use it as is is if
2439 * we're doing plain old texturing. The optimization passes on
2440 * glsl_to_tgsi_visitor should handle cleaning up our mess in that case.
2442 coord
= get_temp(glsl_type::vec4_type
);
2443 coord_dst
= st_dst_reg(coord
);
2444 emit(ir
, TGSI_OPCODE_MOV
, coord_dst
, this->result
);
2447 if (ir
->projector
) {
2448 ir
->projector
->accept(this);
2449 projector
= this->result
;
2452 /* Storage for our result. Ideally for an assignment we'd be using
2453 * the actual storage for the result here, instead.
2455 result_src
= get_temp(glsl_type::vec4_type
);
2456 result_dst
= st_dst_reg(result_src
);
2460 opcode
= TGSI_OPCODE_TEX
;
2463 opcode
= TGSI_OPCODE_TXB
;
2464 ir
->lod_info
.bias
->accept(this);
2465 lod_info
= this->result
;
2468 opcode
= TGSI_OPCODE_TXL
;
2469 ir
->lod_info
.lod
->accept(this);
2470 lod_info
= this->result
;
2473 opcode
= TGSI_OPCODE_TXD
;
2474 ir
->lod_info
.grad
.dPdx
->accept(this);
2476 ir
->lod_info
.grad
.dPdy
->accept(this);
2480 opcode
= TGSI_OPCODE_TXQ
;
2481 ir
->lod_info
.lod
->accept(this);
2482 lod_info
= this->result
;
2485 opcode
= TGSI_OPCODE_TXF
;
2486 ir
->lod_info
.lod
->accept(this);
2487 lod_info
= this->result
;
2489 ir
->offset
->accept(this);
2490 offset
= this->result
;
2495 if (ir
->projector
) {
2496 if (opcode
== TGSI_OPCODE_TEX
) {
2497 /* Slot the projector in as the last component of the coord. */
2498 coord_dst
.writemask
= WRITEMASK_W
;
2499 emit(ir
, TGSI_OPCODE_MOV
, coord_dst
, projector
);
2500 coord_dst
.writemask
= WRITEMASK_XYZW
;
2501 opcode
= TGSI_OPCODE_TXP
;
2503 st_src_reg coord_w
= coord
;
2504 coord_w
.swizzle
= SWIZZLE_WWWW
;
2506 /* For the other TEX opcodes there's no projective version
2507 * since the last slot is taken up by LOD info. Do the
2508 * projective divide now.
2510 coord_dst
.writemask
= WRITEMASK_W
;
2511 emit(ir
, TGSI_OPCODE_RCP
, coord_dst
, projector
);
2513 /* In the case where we have to project the coordinates "by hand,"
2514 * the shadow comparator value must also be projected.
2516 st_src_reg tmp_src
= coord
;
2517 if (ir
->shadow_comparitor
) {
2518 /* Slot the shadow value in as the second to last component of the
2521 ir
->shadow_comparitor
->accept(this);
2523 tmp_src
= get_temp(glsl_type::vec4_type
);
2524 st_dst_reg tmp_dst
= st_dst_reg(tmp_src
);
2526 tmp_dst
.writemask
= WRITEMASK_Z
;
2527 emit(ir
, TGSI_OPCODE_MOV
, tmp_dst
, this->result
);
2529 tmp_dst
.writemask
= WRITEMASK_XY
;
2530 emit(ir
, TGSI_OPCODE_MOV
, tmp_dst
, coord
);
2533 coord_dst
.writemask
= WRITEMASK_XYZ
;
2534 emit(ir
, TGSI_OPCODE_MUL
, coord_dst
, tmp_src
, coord_w
);
2536 coord_dst
.writemask
= WRITEMASK_XYZW
;
2537 coord
.swizzle
= SWIZZLE_XYZW
;
2541 /* If projection is done and the opcode is not TGSI_OPCODE_TXP, then the shadow
2542 * comparator was put in the correct place (and projected) by the code,
2543 * above, that handles by-hand projection.
2545 if (ir
->shadow_comparitor
&& (!ir
->projector
|| opcode
== TGSI_OPCODE_TXP
)) {
2546 /* Slot the shadow value in as the second to last component of the
2549 ir
->shadow_comparitor
->accept(this);
2550 coord_dst
.writemask
= WRITEMASK_Z
;
2551 emit(ir
, TGSI_OPCODE_MOV
, coord_dst
, this->result
);
2552 coord_dst
.writemask
= WRITEMASK_XYZW
;
2555 if (opcode
== TGSI_OPCODE_TXL
|| opcode
== TGSI_OPCODE_TXB
||
2556 opcode
== TGSI_OPCODE_TXF
) {
2557 /* TGSI stores LOD or LOD bias in the last channel of the coords. */
2558 coord_dst
.writemask
= WRITEMASK_W
;
2559 emit(ir
, TGSI_OPCODE_MOV
, coord_dst
, lod_info
);
2560 coord_dst
.writemask
= WRITEMASK_XYZW
;
2563 if (opcode
== TGSI_OPCODE_TXD
)
2564 inst
= emit(ir
, opcode
, result_dst
, coord
, dx
, dy
);
2565 else if (opcode
== TGSI_OPCODE_TXQ
)
2566 inst
= emit(ir
, opcode
, result_dst
, lod_info
);
2567 else if (opcode
== TGSI_OPCODE_TXF
) {
2568 inst
= emit(ir
, opcode
, result_dst
, coord
);
2570 inst
= emit(ir
, opcode
, result_dst
, coord
);
2572 if (ir
->shadow_comparitor
)
2573 inst
->tex_shadow
= GL_TRUE
;
2575 inst
->sampler
= _mesa_get_sampler_uniform_value(ir
->sampler
,
2576 this->shader_program
,
2580 inst
->tex_offset_num_offset
= 1;
2581 inst
->tex_offsets
[0].Index
= offset
.index
;
2582 inst
->tex_offsets
[0].File
= offset
.file
;
2583 inst
->tex_offsets
[0].SwizzleX
= GET_SWZ(offset
.swizzle
, 0);
2584 inst
->tex_offsets
[0].SwizzleY
= GET_SWZ(offset
.swizzle
, 1);
2585 inst
->tex_offsets
[0].SwizzleZ
= GET_SWZ(offset
.swizzle
, 2);
2588 const glsl_type
*sampler_type
= ir
->sampler
->type
;
2590 switch (sampler_type
->sampler_dimensionality
) {
2591 case GLSL_SAMPLER_DIM_1D
:
2592 inst
->tex_target
= (sampler_type
->sampler_array
)
2593 ? TEXTURE_1D_ARRAY_INDEX
: TEXTURE_1D_INDEX
;
2595 case GLSL_SAMPLER_DIM_2D
:
2596 inst
->tex_target
= (sampler_type
->sampler_array
)
2597 ? TEXTURE_2D_ARRAY_INDEX
: TEXTURE_2D_INDEX
;
2599 case GLSL_SAMPLER_DIM_3D
:
2600 inst
->tex_target
= TEXTURE_3D_INDEX
;
2602 case GLSL_SAMPLER_DIM_CUBE
:
2603 inst
->tex_target
= TEXTURE_CUBE_INDEX
;
2605 case GLSL_SAMPLER_DIM_RECT
:
2606 inst
->tex_target
= TEXTURE_RECT_INDEX
;
2608 case GLSL_SAMPLER_DIM_BUF
:
2609 assert(!"FINISHME: Implement ARB_texture_buffer_object");
2612 assert(!"Should not get here.");
2615 this->result
= result_src
;
2619 glsl_to_tgsi_visitor::visit(ir_return
*ir
)
2621 if (ir
->get_value()) {
2625 assert(current_function
);
2627 ir
->get_value()->accept(this);
2628 st_src_reg r
= this->result
;
2630 l
= st_dst_reg(current_function
->return_reg
);
2632 for (i
= 0; i
< type_size(current_function
->sig
->return_type
); i
++) {
2633 emit(ir
, TGSI_OPCODE_MOV
, l
, r
);
2639 emit(ir
, TGSI_OPCODE_RET
);
2643 glsl_to_tgsi_visitor::visit(ir_discard
*ir
)
2645 struct gl_fragment_program
*fp
= (struct gl_fragment_program
*)this->prog
;
2647 if (ir
->condition
) {
2648 ir
->condition
->accept(this);
2649 this->result
.negate
= ~this->result
.negate
;
2650 emit(ir
, TGSI_OPCODE_KIL
, undef_dst
, this->result
);
2652 emit(ir
, TGSI_OPCODE_KILP
);
2655 fp
->UsesKill
= GL_TRUE
;
2659 glsl_to_tgsi_visitor::visit(ir_if
*ir
)
2661 glsl_to_tgsi_instruction
*cond_inst
, *if_inst
;
2662 glsl_to_tgsi_instruction
*prev_inst
;
2664 prev_inst
= (glsl_to_tgsi_instruction
*)this->instructions
.get_tail();
2666 ir
->condition
->accept(this);
2667 assert(this->result
.file
!= PROGRAM_UNDEFINED
);
2669 if (this->options
->EmitCondCodes
) {
2670 cond_inst
= (glsl_to_tgsi_instruction
*)this->instructions
.get_tail();
2672 /* See if we actually generated any instruction for generating
2673 * the condition. If not, then cook up a move to a temp so we
2674 * have something to set cond_update on.
2676 if (cond_inst
== prev_inst
) {
2677 st_src_reg temp
= get_temp(glsl_type::bool_type
);
2678 cond_inst
= emit(ir
->condition
, TGSI_OPCODE_MOV
, st_dst_reg(temp
), result
);
2680 cond_inst
->cond_update
= GL_TRUE
;
2682 if_inst
= emit(ir
->condition
, TGSI_OPCODE_IF
);
2683 if_inst
->dst
.cond_mask
= COND_NE
;
2685 if_inst
= emit(ir
->condition
, TGSI_OPCODE_IF
, undef_dst
, this->result
);
2688 this->instructions
.push_tail(if_inst
);
2690 visit_exec_list(&ir
->then_instructions
, this);
2692 if (!ir
->else_instructions
.is_empty()) {
2693 emit(ir
->condition
, TGSI_OPCODE_ELSE
);
2694 visit_exec_list(&ir
->else_instructions
, this);
2697 if_inst
= emit(ir
->condition
, TGSI_OPCODE_ENDIF
);
2700 glsl_to_tgsi_visitor::glsl_to_tgsi_visitor()
2702 result
.file
= PROGRAM_UNDEFINED
;
2704 next_signature_id
= 1;
2706 current_function
= NULL
;
2707 num_address_regs
= 0;
2708 indirect_addr_temps
= false;
2709 indirect_addr_consts
= false;
2710 mem_ctx
= ralloc_context(NULL
);
2713 glsl_to_tgsi_visitor::~glsl_to_tgsi_visitor()
2715 ralloc_free(mem_ctx
);
2718 extern "C" void free_glsl_to_tgsi_visitor(glsl_to_tgsi_visitor
*v
)
2725 * Count resources used by the given gpu program (number of texture
2729 count_resources(glsl_to_tgsi_visitor
*v
, gl_program
*prog
)
2731 v
->samplers_used
= 0;
2733 foreach_iter(exec_list_iterator
, iter
, v
->instructions
) {
2734 glsl_to_tgsi_instruction
*inst
= (glsl_to_tgsi_instruction
*)iter
.get();
2736 if (is_tex_instruction(inst
->op
)) {
2737 v
->samplers_used
|= 1 << inst
->sampler
;
2739 prog
->SamplerTargets
[inst
->sampler
] =
2740 (gl_texture_index
)inst
->tex_target
;
2741 if (inst
->tex_shadow
) {
2742 prog
->ShadowSamplers
|= 1 << inst
->sampler
;
2747 prog
->SamplersUsed
= v
->samplers_used
;
2748 _mesa_update_shader_textures_used(prog
);
2753 * Check if the given vertex/fragment/shader program is within the
2754 * resource limits of the context (number of texture units, etc).
2755 * If any of those checks fail, record a linker error.
2757 * XXX more checks are needed...
2760 check_resources(const struct gl_context
*ctx
,
2761 struct gl_shader_program
*shader_program
,
2762 glsl_to_tgsi_visitor
*prog
,
2763 struct gl_program
*proginfo
)
2765 switch (proginfo
->Target
) {
2766 case GL_VERTEX_PROGRAM_ARB
:
2767 if (_mesa_bitcount(prog
->samplers_used
) >
2768 ctx
->Const
.MaxVertexTextureImageUnits
) {
2769 fail_link(shader_program
, "Too many vertex shader texture samplers");
2771 if (proginfo
->Parameters
->NumParameters
> MAX_UNIFORMS
) {
2772 fail_link(shader_program
, "Too many vertex shader constants");
2775 case MESA_GEOMETRY_PROGRAM
:
2776 if (_mesa_bitcount(prog
->samplers_used
) >
2777 ctx
->Const
.MaxGeometryTextureImageUnits
) {
2778 fail_link(shader_program
, "Too many geometry shader texture samplers");
2780 if (proginfo
->Parameters
->NumParameters
>
2781 MAX_GEOMETRY_UNIFORM_COMPONENTS
/ 4) {
2782 fail_link(shader_program
, "Too many geometry shader constants");
2785 case GL_FRAGMENT_PROGRAM_ARB
:
2786 if (_mesa_bitcount(prog
->samplers_used
) >
2787 ctx
->Const
.MaxTextureImageUnits
) {
2788 fail_link(shader_program
, "Too many fragment shader texture samplers");
2790 if (proginfo
->Parameters
->NumParameters
> MAX_UNIFORMS
) {
2791 fail_link(shader_program
, "Too many fragment shader constants");
2795 _mesa_problem(ctx
, "unexpected program type in check_resources()");
2801 struct uniform_sort
{
2802 struct gl_uniform
*u
;
2806 /* The shader_program->Uniforms list is almost sorted in increasing
2807 * uniform->{Frag,Vert}Pos locations, but not quite when there are
2808 * uniforms shared between targets. We need to add parameters in
2809 * increasing order for the targets.
2812 sort_uniforms(const void *a
, const void *b
)
2814 struct uniform_sort
*u1
= (struct uniform_sort
*)a
;
2815 struct uniform_sort
*u2
= (struct uniform_sort
*)b
;
2817 return u1
->pos
- u2
->pos
;
2820 /* Add the uniforms to the parameters. The linker chose locations
2821 * in our parameters lists (which weren't created yet), which the
2822 * uniforms code will use to poke values into our parameters list
2823 * when uniforms are updated.
2826 add_uniforms_to_parameters_list(struct gl_shader_program
*shader_program
,
2827 struct gl_shader
*shader
,
2828 struct gl_program
*prog
)
2831 unsigned int next_sampler
= 0, num_uniforms
= 0;
2832 struct uniform_sort
*sorted_uniforms
;
2834 sorted_uniforms
= ralloc_array(NULL
, struct uniform_sort
,
2835 shader_program
->Uniforms
->NumUniforms
);
2837 for (i
= 0; i
< shader_program
->Uniforms
->NumUniforms
; i
++) {
2838 struct gl_uniform
*uniform
= shader_program
->Uniforms
->Uniforms
+ i
;
2839 int parameter_index
= -1;
2841 switch (shader
->Type
) {
2842 case GL_VERTEX_SHADER
:
2843 parameter_index
= uniform
->VertPos
;
2845 case GL_FRAGMENT_SHADER
:
2846 parameter_index
= uniform
->FragPos
;
2848 case GL_GEOMETRY_SHADER
:
2849 parameter_index
= uniform
->GeomPos
;
2853 /* Only add uniforms used in our target. */
2854 if (parameter_index
!= -1) {
2855 sorted_uniforms
[num_uniforms
].pos
= parameter_index
;
2856 sorted_uniforms
[num_uniforms
].u
= uniform
;
2861 qsort(sorted_uniforms
, num_uniforms
, sizeof(struct uniform_sort
),
2864 for (i
= 0; i
< num_uniforms
; i
++) {
2865 struct gl_uniform
*uniform
= sorted_uniforms
[i
].u
;
2866 int parameter_index
= sorted_uniforms
[i
].pos
;
2867 const glsl_type
*type
= uniform
->Type
;
2870 if (type
->is_vector() ||
2871 type
->is_scalar()) {
2872 size
= type
->vector_elements
;
2874 size
= type_size(type
) * 4;
2877 gl_register_file file
;
2878 if (type
->is_sampler() ||
2879 (type
->is_array() && type
->fields
.array
->is_sampler())) {
2880 file
= PROGRAM_SAMPLER
;
2882 file
= PROGRAM_UNIFORM
;
2885 GLint index
= _mesa_lookup_parameter_index(prog
->Parameters
, -1,
2889 index
= _mesa_add_parameter(prog
->Parameters
, file
,
2890 uniform
->Name
, size
, type
->gl_type
,
2893 /* Sampler uniform values are stored in prog->SamplerUnits,
2894 * and the entry in that array is selected by this index we
2895 * store in ParameterValues[].
2897 if (file
== PROGRAM_SAMPLER
) {
2898 for (unsigned int j
= 0; j
< size
/ 4; j
++)
2899 prog
->Parameters
->ParameterValues
[index
+ j
][0].f
= next_sampler
++;
2902 /* The location chosen in the Parameters list here (returned
2903 * from _mesa_add_uniform) has to match what the linker chose.
2905 if (index
!= parameter_index
) {
2906 fail_link(shader_program
, "Allocation of uniform `%s' to target "
2907 "failed (%d vs %d)\n",
2908 uniform
->Name
, index
, parameter_index
);
2913 ralloc_free(sorted_uniforms
);
2917 set_uniform_initializer(struct gl_context
*ctx
, void *mem_ctx
,
2918 struct gl_shader_program
*shader_program
,
2919 const char *name
, const glsl_type
*type
,
2922 if (type
->is_record()) {
2923 ir_constant
*field_constant
;
2925 field_constant
= (ir_constant
*)val
->components
.get_head();
2927 for (unsigned int i
= 0; i
< type
->length
; i
++) {
2928 const glsl_type
*field_type
= type
->fields
.structure
[i
].type
;
2929 const char *field_name
= ralloc_asprintf(mem_ctx
, "%s.%s", name
,
2930 type
->fields
.structure
[i
].name
);
2931 set_uniform_initializer(ctx
, mem_ctx
, shader_program
, field_name
,
2932 field_type
, field_constant
);
2933 field_constant
= (ir_constant
*)field_constant
->next
;
2938 int loc
= _mesa_get_uniform_location(ctx
, shader_program
, name
);
2941 fail_link(shader_program
,
2942 "Couldn't find uniform for initializer %s\n", name
);
2946 for (unsigned int i
= 0; i
< (type
->is_array() ? type
->length
: 1); i
++) {
2947 ir_constant
*element
;
2948 const glsl_type
*element_type
;
2949 if (type
->is_array()) {
2950 element
= val
->array_elements
[i
];
2951 element_type
= type
->fields
.array
;
2954 element_type
= type
;
2959 if (element_type
->base_type
== GLSL_TYPE_BOOL
) {
2960 int *conv
= ralloc_array(mem_ctx
, int, element_type
->components());
2961 for (unsigned int j
= 0; j
< element_type
->components(); j
++) {
2962 conv
[j
] = element
->value
.b
[j
];
2964 values
= (void *)conv
;
2965 element_type
= glsl_type::get_instance(GLSL_TYPE_INT
,
2966 element_type
->vector_elements
,
2969 values
= &element
->value
;
2972 if (element_type
->is_matrix()) {
2973 _mesa_uniform_matrix(ctx
, shader_program
,
2974 element_type
->matrix_columns
,
2975 element_type
->vector_elements
,
2976 loc
, 1, GL_FALSE
, (GLfloat
*)values
);
2977 loc
+= element_type
->matrix_columns
;
2979 _mesa_uniform(ctx
, shader_program
, loc
, element_type
->matrix_columns
,
2980 values
, element_type
->gl_type
);
2981 loc
+= type_size(element_type
);
2987 * Scan/rewrite program to remove reads of custom (output) registers.
2988 * The passed type has to be either PROGRAM_OUTPUT or PROGRAM_VARYING
2989 * (for vertex shaders).
2990 * In GLSL shaders, varying vars can be read and written.
2991 * On some hardware, trying to read an output register causes trouble.
2992 * So, rewrite the program to use a temporary register in this case.
2994 * Based on _mesa_remove_output_reads from programopt.c.
2997 glsl_to_tgsi_visitor::remove_output_reads(gl_register_file type
)
3000 GLint outputMap
[VERT_RESULT_MAX
];
3001 GLint outputTypes
[VERT_RESULT_MAX
];
3002 GLuint numVaryingReads
= 0;
3003 GLboolean usedTemps
[MAX_TEMPS
];
3004 GLuint firstTemp
= 0;
3006 _mesa_find_used_registers(prog
, PROGRAM_TEMPORARY
,
3007 usedTemps
, MAX_TEMPS
);
3009 assert(type
== PROGRAM_VARYING
|| type
== PROGRAM_OUTPUT
);
3010 assert(prog
->Target
== GL_VERTEX_PROGRAM_ARB
|| type
!= PROGRAM_VARYING
);
3012 for (i
= 0; i
< VERT_RESULT_MAX
; i
++)
3015 /* look for instructions which read from varying vars */
3016 foreach_iter(exec_list_iterator
, iter
, this->instructions
) {
3017 glsl_to_tgsi_instruction
*inst
= (glsl_to_tgsi_instruction
*)iter
.get();
3018 const GLuint numSrc
= num_inst_src_regs(inst
->op
);
3020 for (j
= 0; j
< numSrc
; j
++) {
3021 if (inst
->src
[j
].file
== type
) {
3022 /* replace the read with a temp reg */
3023 const GLuint var
= inst
->src
[j
].index
;
3024 if (outputMap
[var
] == -1) {
3026 outputMap
[var
] = _mesa_find_free_register(usedTemps
,
3029 outputTypes
[var
] = inst
->src
[j
].type
;
3030 firstTemp
= outputMap
[var
] + 1;
3032 inst
->src
[j
].file
= PROGRAM_TEMPORARY
;
3033 inst
->src
[j
].index
= outputMap
[var
];
3038 if (numVaryingReads
== 0)
3039 return; /* nothing to be done */
3041 /* look for instructions which write to the varying vars identified above */
3042 foreach_iter(exec_list_iterator
, iter
, this->instructions
) {
3043 glsl_to_tgsi_instruction
*inst
= (glsl_to_tgsi_instruction
*)iter
.get();
3044 if (inst
->dst
.file
== type
&& outputMap
[inst
->dst
.index
] >= 0) {
3045 /* change inst to write to the temp reg, instead of the varying */
3046 inst
->dst
.file
= PROGRAM_TEMPORARY
;
3047 inst
->dst
.index
= outputMap
[inst
->dst
.index
];
3051 /* insert new MOV instructions at the end */
3052 for (i
= 0; i
< VERT_RESULT_MAX
; i
++) {
3053 if (outputMap
[i
] >= 0) {
3054 /* MOV VAR[i], TEMP[tmp]; */
3055 st_src_reg src
= st_src_reg(PROGRAM_TEMPORARY
, outputMap
[i
], outputTypes
[i
]);
3056 st_dst_reg dst
= st_dst_reg(type
, WRITEMASK_XYZW
, outputTypes
[i
]);
3058 this->emit(NULL
, TGSI_OPCODE_MOV
, dst
, src
);
3064 * Returns the mask of channels (bitmask of WRITEMASK_X,Y,Z,W) which
3065 * are read from the given src in this instruction
3068 get_src_arg_mask(st_dst_reg dst
, st_src_reg src
)
3070 int read_mask
= 0, comp
;
3072 /* Now, given the src swizzle and the written channels, find which
3073 * components are actually read
3075 for (comp
= 0; comp
< 4; ++comp
) {
3076 const unsigned coord
= GET_SWZ(src
.swizzle
, comp
);
3078 if (dst
.writemask
& (1 << comp
) && coord
<= SWIZZLE_W
)
3079 read_mask
|= 1 << coord
;
3086 * This pass replaces CMP T0, T1 T2 T0 with MOV T0, T2 when the CMP
3087 * instruction is the first instruction to write to register T0. There are
3088 * several lowering passes done in GLSL IR (e.g. branches and
3089 * relative addressing) that create a large number of conditional assignments
3090 * that ir_to_mesa converts to CMP instructions like the one mentioned above.
3092 * Here is why this conversion is safe:
3093 * CMP T0, T1 T2 T0 can be expanded to:
3099 * If (T1 < 0.0) evaluates to true then our replacement MOV T0, T2 is the same
3100 * as the original program. If (T1 < 0.0) evaluates to false, executing
3101 * MOV T0, T0 will store a garbage value in T0 since T0 is uninitialized.
3102 * Therefore, it doesn't matter that we are replacing MOV T0, T0 with MOV T0, T2
3103 * because any instruction that was going to read from T0 after this was going
3104 * to read a garbage value anyway.
3107 glsl_to_tgsi_visitor::simplify_cmp(void)
3109 unsigned tempWrites
[MAX_TEMPS
];
3110 unsigned outputWrites
[MAX_PROGRAM_OUTPUTS
];
3112 memset(tempWrites
, 0, sizeof(tempWrites
));
3113 memset(outputWrites
, 0, sizeof(outputWrites
));
3115 foreach_iter(exec_list_iterator
, iter
, this->instructions
) {
3116 glsl_to_tgsi_instruction
*inst
= (glsl_to_tgsi_instruction
*)iter
.get();
3117 unsigned prevWriteMask
= 0;
3119 /* Give up if we encounter relative addressing or flow control. */
3120 if (inst
->dst
.reladdr
||
3121 tgsi_get_opcode_info(inst
->op
)->is_branch
||
3122 inst
->op
== TGSI_OPCODE_BGNSUB
||
3123 inst
->op
== TGSI_OPCODE_CONT
||
3124 inst
->op
== TGSI_OPCODE_END
||
3125 inst
->op
== TGSI_OPCODE_ENDSUB
||
3126 inst
->op
== TGSI_OPCODE_RET
) {
3130 if (inst
->dst
.file
== PROGRAM_OUTPUT
) {
3131 assert(inst
->dst
.index
< MAX_PROGRAM_OUTPUTS
);
3132 prevWriteMask
= outputWrites
[inst
->dst
.index
];
3133 outputWrites
[inst
->dst
.index
] |= inst
->dst
.writemask
;
3134 } else if (inst
->dst
.file
== PROGRAM_TEMPORARY
) {
3135 assert(inst
->dst
.index
< MAX_TEMPS
);
3136 prevWriteMask
= tempWrites
[inst
->dst
.index
];
3137 tempWrites
[inst
->dst
.index
] |= inst
->dst
.writemask
;
3140 /* For a CMP to be considered a conditional write, the destination
3141 * register and source register two must be the same. */
3142 if (inst
->op
== TGSI_OPCODE_CMP
3143 && !(inst
->dst
.writemask
& prevWriteMask
)
3144 && inst
->src
[2].file
== inst
->dst
.file
3145 && inst
->src
[2].index
== inst
->dst
.index
3146 && inst
->dst
.writemask
== get_src_arg_mask(inst
->dst
, inst
->src
[2])) {
3148 inst
->op
= TGSI_OPCODE_MOV
;
3149 inst
->src
[0] = inst
->src
[1];
3154 /* Replaces all references to a temporary register index with another index. */
3156 glsl_to_tgsi_visitor::rename_temp_register(int index
, int new_index
)
3158 foreach_iter(exec_list_iterator
, iter
, this->instructions
) {
3159 glsl_to_tgsi_instruction
*inst
= (glsl_to_tgsi_instruction
*)iter
.get();
3162 for (j
=0; j
< num_inst_src_regs(inst
->op
); j
++) {
3163 if (inst
->src
[j
].file
== PROGRAM_TEMPORARY
&&
3164 inst
->src
[j
].index
== index
) {
3165 inst
->src
[j
].index
= new_index
;
3169 if (inst
->dst
.file
== PROGRAM_TEMPORARY
&& inst
->dst
.index
== index
) {
3170 inst
->dst
.index
= new_index
;
3176 glsl_to_tgsi_visitor::get_first_temp_read(int index
)
3178 int depth
= 0; /* loop depth */
3179 int loop_start
= -1; /* index of the first active BGNLOOP (if any) */
3182 foreach_iter(exec_list_iterator
, iter
, this->instructions
) {
3183 glsl_to_tgsi_instruction
*inst
= (glsl_to_tgsi_instruction
*)iter
.get();
3185 for (j
=0; j
< num_inst_src_regs(inst
->op
); j
++) {
3186 if (inst
->src
[j
].file
== PROGRAM_TEMPORARY
&&
3187 inst
->src
[j
].index
== index
) {
3188 return (depth
== 0) ? i
: loop_start
;
3192 if (inst
->op
== TGSI_OPCODE_BGNLOOP
) {
3195 } else if (inst
->op
== TGSI_OPCODE_ENDLOOP
) {
3208 glsl_to_tgsi_visitor::get_first_temp_write(int index
)
3210 int depth
= 0; /* loop depth */
3211 int loop_start
= -1; /* index of the first active BGNLOOP (if any) */
3214 foreach_iter(exec_list_iterator
, iter
, this->instructions
) {
3215 glsl_to_tgsi_instruction
*inst
= (glsl_to_tgsi_instruction
*)iter
.get();
3217 if (inst
->dst
.file
== PROGRAM_TEMPORARY
&& inst
->dst
.index
== index
) {
3218 return (depth
== 0) ? i
: loop_start
;
3221 if (inst
->op
== TGSI_OPCODE_BGNLOOP
) {
3224 } else if (inst
->op
== TGSI_OPCODE_ENDLOOP
) {
3237 glsl_to_tgsi_visitor::get_last_temp_read(int index
)
3239 int depth
= 0; /* loop depth */
3240 int last
= -1; /* index of last instruction that reads the temporary */
3243 foreach_iter(exec_list_iterator
, iter
, this->instructions
) {
3244 glsl_to_tgsi_instruction
*inst
= (glsl_to_tgsi_instruction
*)iter
.get();
3246 for (j
=0; j
< num_inst_src_regs(inst
->op
); j
++) {
3247 if (inst
->src
[j
].file
== PROGRAM_TEMPORARY
&&
3248 inst
->src
[j
].index
== index
) {
3249 last
= (depth
== 0) ? i
: -2;
3253 if (inst
->op
== TGSI_OPCODE_BGNLOOP
)
3255 else if (inst
->op
== TGSI_OPCODE_ENDLOOP
)
3256 if (--depth
== 0 && last
== -2)
3268 glsl_to_tgsi_visitor::get_last_temp_write(int index
)
3270 int depth
= 0; /* loop depth */
3271 int last
= -1; /* index of last instruction that writes to the temporary */
3274 foreach_iter(exec_list_iterator
, iter
, this->instructions
) {
3275 glsl_to_tgsi_instruction
*inst
= (glsl_to_tgsi_instruction
*)iter
.get();
3277 if (inst
->dst
.file
== PROGRAM_TEMPORARY
&& inst
->dst
.index
== index
)
3278 last
= (depth
== 0) ? i
: -2;
3280 if (inst
->op
== TGSI_OPCODE_BGNLOOP
)
3282 else if (inst
->op
== TGSI_OPCODE_ENDLOOP
)
3283 if (--depth
== 0 && last
== -2)
3295 * On a basic block basis, tracks available PROGRAM_TEMPORARY register
3296 * channels for copy propagation and updates following instructions to
3297 * use the original versions.
3299 * The glsl_to_tgsi_visitor lazily produces code assuming that this pass
3300 * will occur. As an example, a TXP production before this pass:
3302 * 0: MOV TEMP[1], INPUT[4].xyyy;
3303 * 1: MOV TEMP[1].w, INPUT[4].wwww;
3304 * 2: TXP TEMP[2], TEMP[1], texture[0], 2D;
3308 * 0: MOV TEMP[1], INPUT[4].xyyy;
3309 * 1: MOV TEMP[1].w, INPUT[4].wwww;
3310 * 2: TXP TEMP[2], INPUT[4].xyyw, texture[0], 2D;
3312 * which allows for dead code elimination on TEMP[1]'s writes.
3315 glsl_to_tgsi_visitor::copy_propagate(void)
3317 glsl_to_tgsi_instruction
**acp
= rzalloc_array(mem_ctx
,
3318 glsl_to_tgsi_instruction
*,
3319 this->next_temp
* 4);
3320 int *acp_level
= rzalloc_array(mem_ctx
, int, this->next_temp
* 4);
3323 foreach_iter(exec_list_iterator
, iter
, this->instructions
) {
3324 glsl_to_tgsi_instruction
*inst
= (glsl_to_tgsi_instruction
*)iter
.get();
3326 assert(inst
->dst
.file
!= PROGRAM_TEMPORARY
3327 || inst
->dst
.index
< this->next_temp
);
3329 /* First, do any copy propagation possible into the src regs. */
3330 for (int r
= 0; r
< 3; r
++) {
3331 glsl_to_tgsi_instruction
*first
= NULL
;
3333 int acp_base
= inst
->src
[r
].index
* 4;
3335 if (inst
->src
[r
].file
!= PROGRAM_TEMPORARY
||
3336 inst
->src
[r
].reladdr
)
3339 /* See if we can find entries in the ACP consisting of MOVs
3340 * from the same src register for all the swizzled channels
3341 * of this src register reference.
3343 for (int i
= 0; i
< 4; i
++) {
3344 int src_chan
= GET_SWZ(inst
->src
[r
].swizzle
, i
);
3345 glsl_to_tgsi_instruction
*copy_chan
= acp
[acp_base
+ src_chan
];
3352 assert(acp_level
[acp_base
+ src_chan
] <= level
);
3357 if (first
->src
[0].file
!= copy_chan
->src
[0].file
||
3358 first
->src
[0].index
!= copy_chan
->src
[0].index
) {
3366 /* We've now validated that we can copy-propagate to
3367 * replace this src register reference. Do it.
3369 inst
->src
[r
].file
= first
->src
[0].file
;
3370 inst
->src
[r
].index
= first
->src
[0].index
;
3373 for (int i
= 0; i
< 4; i
++) {
3374 int src_chan
= GET_SWZ(inst
->src
[r
].swizzle
, i
);
3375 glsl_to_tgsi_instruction
*copy_inst
= acp
[acp_base
+ src_chan
];
3376 swizzle
|= (GET_SWZ(copy_inst
->src
[0].swizzle
, src_chan
) <<
3379 inst
->src
[r
].swizzle
= swizzle
;
3384 case TGSI_OPCODE_BGNLOOP
:
3385 case TGSI_OPCODE_ENDLOOP
:
3386 /* End of a basic block, clear the ACP entirely. */
3387 memset(acp
, 0, sizeof(*acp
) * this->next_temp
* 4);
3390 case TGSI_OPCODE_IF
:
3394 case TGSI_OPCODE_ENDIF
:
3395 case TGSI_OPCODE_ELSE
:
3396 /* Clear all channels written inside the block from the ACP, but
3397 * leaving those that were not touched.
3399 for (int r
= 0; r
< this->next_temp
; r
++) {
3400 for (int c
= 0; c
< 4; c
++) {
3401 if (!acp
[4 * r
+ c
])
3404 if (acp_level
[4 * r
+ c
] >= level
)
3405 acp
[4 * r
+ c
] = NULL
;
3408 if (inst
->op
== TGSI_OPCODE_ENDIF
)
3413 /* Continuing the block, clear any written channels from
3416 if (inst
->dst
.file
== PROGRAM_TEMPORARY
&& inst
->dst
.reladdr
) {
3417 /* Any temporary might be written, so no copy propagation
3418 * across this instruction.
3420 memset(acp
, 0, sizeof(*acp
) * this->next_temp
* 4);
3421 } else if (inst
->dst
.file
== PROGRAM_OUTPUT
&&
3422 inst
->dst
.reladdr
) {
3423 /* Any output might be written, so no copy propagation
3424 * from outputs across this instruction.
3426 for (int r
= 0; r
< this->next_temp
; r
++) {
3427 for (int c
= 0; c
< 4; c
++) {
3428 if (!acp
[4 * r
+ c
])
3431 if (acp
[4 * r
+ c
]->src
[0].file
== PROGRAM_OUTPUT
)
3432 acp
[4 * r
+ c
] = NULL
;
3435 } else if (inst
->dst
.file
== PROGRAM_TEMPORARY
||
3436 inst
->dst
.file
== PROGRAM_OUTPUT
) {
3437 /* Clear where it's used as dst. */
3438 if (inst
->dst
.file
== PROGRAM_TEMPORARY
) {
3439 for (int c
= 0; c
< 4; c
++) {
3440 if (inst
->dst
.writemask
& (1 << c
)) {
3441 acp
[4 * inst
->dst
.index
+ c
] = NULL
;
3446 /* Clear where it's used as src. */
3447 for (int r
= 0; r
< this->next_temp
; r
++) {
3448 for (int c
= 0; c
< 4; c
++) {
3449 if (!acp
[4 * r
+ c
])
3452 int src_chan
= GET_SWZ(acp
[4 * r
+ c
]->src
[0].swizzle
, c
);
3454 if (acp
[4 * r
+ c
]->src
[0].file
== inst
->dst
.file
&&
3455 acp
[4 * r
+ c
]->src
[0].index
== inst
->dst
.index
&&
3456 inst
->dst
.writemask
& (1 << src_chan
))
3458 acp
[4 * r
+ c
] = NULL
;
3466 /* If this is a copy, add it to the ACP. */
3467 if (inst
->op
== TGSI_OPCODE_MOV
&&
3468 inst
->dst
.file
== PROGRAM_TEMPORARY
&&
3469 !inst
->dst
.reladdr
&&
3471 !inst
->src
[0].reladdr
&&
3472 !inst
->src
[0].negate
) {
3473 for (int i
= 0; i
< 4; i
++) {
3474 if (inst
->dst
.writemask
& (1 << i
)) {
3475 acp
[4 * inst
->dst
.index
+ i
] = inst
;
3476 acp_level
[4 * inst
->dst
.index
+ i
] = level
;
3482 ralloc_free(acp_level
);
3487 * Tracks available PROGRAM_TEMPORARY registers for dead code elimination.
3489 * The glsl_to_tgsi_visitor lazily produces code assuming that this pass
3490 * will occur. As an example, a TXP production after copy propagation but
3493 * 0: MOV TEMP[1], INPUT[4].xyyy;
3494 * 1: MOV TEMP[1].w, INPUT[4].wwww;
3495 * 2: TXP TEMP[2], INPUT[4].xyyw, texture[0], 2D;
3497 * and after this pass:
3499 * 0: TXP TEMP[2], INPUT[4].xyyw, texture[0], 2D;
3501 * FIXME: assumes that all functions are inlined (no support for BGNSUB/ENDSUB)
3502 * FIXME: doesn't eliminate all dead code inside of loops; it steps around them
3505 glsl_to_tgsi_visitor::eliminate_dead_code(void)
3509 for (i
=0; i
< this->next_temp
; i
++) {
3510 int last_read
= get_last_temp_read(i
);
3513 foreach_iter(exec_list_iterator
, iter
, this->instructions
) {
3514 glsl_to_tgsi_instruction
*inst
= (glsl_to_tgsi_instruction
*)iter
.get();
3516 if (inst
->dst
.file
== PROGRAM_TEMPORARY
&& inst
->dst
.index
== i
&&
3529 * On a basic block basis, tracks available PROGRAM_TEMPORARY registers for dead
3530 * code elimination. This is less primitive than eliminate_dead_code(), as it
3531 * is per-channel and can detect consecutive writes without a read between them
3532 * as dead code. However, there is some dead code that can be eliminated by
3533 * eliminate_dead_code() but not this function - for example, this function
3534 * cannot eliminate an instruction writing to a register that is never read and
3535 * is the only instruction writing to that register.
3537 * The glsl_to_tgsi_visitor lazily produces code assuming that this pass
3541 glsl_to_tgsi_visitor::eliminate_dead_code_advanced(void)
3543 glsl_to_tgsi_instruction
**writes
= rzalloc_array(mem_ctx
,
3544 glsl_to_tgsi_instruction
*,
3545 this->next_temp
* 4);
3546 int *write_level
= rzalloc_array(mem_ctx
, int, this->next_temp
* 4);
3550 foreach_iter(exec_list_iterator
, iter
, this->instructions
) {
3551 glsl_to_tgsi_instruction
*inst
= (glsl_to_tgsi_instruction
*)iter
.get();
3553 assert(inst
->dst
.file
!= PROGRAM_TEMPORARY
3554 || inst
->dst
.index
< this->next_temp
);
3557 case TGSI_OPCODE_BGNLOOP
:
3558 case TGSI_OPCODE_ENDLOOP
:
3559 /* End of a basic block, clear the write array entirely.
3560 * FIXME: This keeps us from killing dead code when the writes are
3561 * on either side of a loop, even when the register isn't touched
3564 memset(writes
, 0, sizeof(*writes
) * this->next_temp
* 4);
3567 case TGSI_OPCODE_ENDIF
:
3571 case TGSI_OPCODE_ELSE
:
3572 /* Clear all channels written inside the preceding if block from the
3573 * write array, but leave those that were not touched.
3575 * FIXME: This destroys opportunities to remove dead code inside of
3576 * IF blocks that are followed by an ELSE block.
3578 for (int r
= 0; r
< this->next_temp
; r
++) {
3579 for (int c
= 0; c
< 4; c
++) {
3580 if (!writes
[4 * r
+ c
])
3583 if (write_level
[4 * r
+ c
] >= level
)
3584 writes
[4 * r
+ c
] = NULL
;
3589 case TGSI_OPCODE_IF
:
3591 /* fallthrough to default case to mark the condition as read */
3594 /* Continuing the block, clear any channels from the write array that
3595 * are read by this instruction.
3597 for (unsigned i
= 0; i
< Elements(inst
->src
); i
++) {
3598 if (inst
->src
[i
].file
== PROGRAM_TEMPORARY
&& inst
->src
[i
].reladdr
){
3599 /* Any temporary might be read, so no dead code elimination
3600 * across this instruction.
3602 memset(writes
, 0, sizeof(*writes
) * this->next_temp
* 4);
3603 } else if (inst
->src
[i
].file
== PROGRAM_TEMPORARY
) {
3604 /* Clear where it's used as src. */
3605 int src_chans
= 1 << GET_SWZ(inst
->src
[i
].swizzle
, 0);
3606 src_chans
|= 1 << GET_SWZ(inst
->src
[i
].swizzle
, 1);
3607 src_chans
|= 1 << GET_SWZ(inst
->src
[i
].swizzle
, 2);
3608 src_chans
|= 1 << GET_SWZ(inst
->src
[i
].swizzle
, 3);
3610 for (int c
= 0; c
< 4; c
++) {
3611 if (src_chans
& (1 << c
)) {
3612 writes
[4 * inst
->src
[i
].index
+ c
] = NULL
;
3620 /* If this instruction writes to a temporary, add it to the write array.
3621 * If there is already an instruction in the write array for one or more
3622 * of the channels, flag that channel write as dead.
3624 if (inst
->dst
.file
== PROGRAM_TEMPORARY
&&
3625 !inst
->dst
.reladdr
&&
3627 for (int c
= 0; c
< 4; c
++) {
3628 if (inst
->dst
.writemask
& (1 << c
)) {
3629 if (writes
[4 * inst
->dst
.index
+ c
]) {
3630 if (write_level
[4 * inst
->dst
.index
+ c
] < level
)
3633 writes
[4 * inst
->dst
.index
+ c
]->dead_mask
|= (1 << c
);
3635 writes
[4 * inst
->dst
.index
+ c
] = inst
;
3636 write_level
[4 * inst
->dst
.index
+ c
] = level
;
3642 /* Anything still in the write array at this point is dead code. */
3643 for (int r
= 0; r
< this->next_temp
; r
++) {
3644 for (int c
= 0; c
< 4; c
++) {
3645 glsl_to_tgsi_instruction
*inst
= writes
[4 * r
+ c
];
3647 inst
->dead_mask
|= (1 << c
);
3651 /* Now actually remove the instructions that are completely dead and update
3652 * the writemask of other instructions with dead channels.
3654 foreach_iter(exec_list_iterator
, iter
, this->instructions
) {
3655 glsl_to_tgsi_instruction
*inst
= (glsl_to_tgsi_instruction
*)iter
.get();
3657 if (!inst
->dead_mask
|| !inst
->dst
.writemask
)
3659 else if (inst
->dead_mask
== inst
->dst
.writemask
) {
3664 inst
->dst
.writemask
&= ~(inst
->dead_mask
);
3667 ralloc_free(write_level
);
3668 ralloc_free(writes
);
3673 /* Merges temporary registers together where possible to reduce the number of
3674 * registers needed to run a program.
3676 * Produces optimal code only after copy propagation and dead code elimination
3679 glsl_to_tgsi_visitor::merge_registers(void)
3681 int *last_reads
= rzalloc_array(mem_ctx
, int, this->next_temp
);
3682 int *first_writes
= rzalloc_array(mem_ctx
, int, this->next_temp
);
3685 /* Read the indices of the last read and first write to each temp register
3686 * into an array so that we don't have to traverse the instruction list as
3688 for (i
=0; i
< this->next_temp
; i
++) {
3689 last_reads
[i
] = get_last_temp_read(i
);
3690 first_writes
[i
] = get_first_temp_write(i
);
3693 /* Start looking for registers with non-overlapping usages that can be
3694 * merged together. */
3695 for (i
=0; i
< this->next_temp
; i
++) {
3696 /* Don't touch unused registers. */
3697 if (last_reads
[i
] < 0 || first_writes
[i
] < 0) continue;
3699 for (j
=0; j
< this->next_temp
; j
++) {
3700 /* Don't touch unused registers. */
3701 if (last_reads
[j
] < 0 || first_writes
[j
] < 0) continue;
3703 /* We can merge the two registers if the first write to j is after or
3704 * in the same instruction as the last read from i. Note that the
3705 * register at index i will always be used earlier or at the same time
3706 * as the register at index j. */
3707 if (first_writes
[i
] <= first_writes
[j
] &&
3708 last_reads
[i
] <= first_writes
[j
])
3710 rename_temp_register(j
, i
); /* Replace all references to j with i.*/
3712 /* Update the first_writes and last_reads arrays with the new
3713 * values for the merged register index, and mark the newly unused
3714 * register index as such. */
3715 last_reads
[i
] = last_reads
[j
];
3716 first_writes
[j
] = -1;
3722 ralloc_free(last_reads
);
3723 ralloc_free(first_writes
);
3726 /* Reassign indices to temporary registers by reusing unused indices created
3727 * by optimization passes. */
3729 glsl_to_tgsi_visitor::renumber_registers(void)
3734 for (i
=0; i
< this->next_temp
; i
++) {
3735 if (get_first_temp_read(i
) < 0) continue;
3737 rename_temp_register(i
, new_index
);
3741 this->next_temp
= new_index
;
3745 * Returns a fragment program which implements the current pixel transfer ops.
3746 * Based on get_pixel_transfer_program in st_atom_pixeltransfer.c.
3749 get_pixel_transfer_visitor(struct st_fragment_program
*fp
,
3750 glsl_to_tgsi_visitor
*original
,
3751 int scale_and_bias
, int pixel_maps
)
3753 glsl_to_tgsi_visitor
*v
= new glsl_to_tgsi_visitor();
3754 struct st_context
*st
= st_context(original
->ctx
);
3755 struct gl_program
*prog
= &fp
->Base
.Base
;
3756 struct gl_program_parameter_list
*params
= _mesa_new_parameter_list();
3757 st_src_reg coord
, src0
;
3759 glsl_to_tgsi_instruction
*inst
;
3761 /* Copy attributes of the glsl_to_tgsi_visitor in the original shader. */
3762 v
->ctx
= original
->ctx
;
3764 v
->glsl_version
= original
->glsl_version
;
3765 v
->native_integers
= original
->native_integers
;
3766 v
->options
= original
->options
;
3767 v
->next_temp
= original
->next_temp
;
3768 v
->num_address_regs
= original
->num_address_regs
;
3769 v
->samplers_used
= prog
->SamplersUsed
= original
->samplers_used
;
3770 v
->indirect_addr_temps
= original
->indirect_addr_temps
;
3771 v
->indirect_addr_consts
= original
->indirect_addr_consts
;
3772 memcpy(&v
->immediates
, &original
->immediates
, sizeof(v
->immediates
));
3775 * Get initial pixel color from the texture.
3776 * TEX colorTemp, fragment.texcoord[0], texture[0], 2D;
3778 coord
= st_src_reg(PROGRAM_INPUT
, FRAG_ATTRIB_TEX0
, glsl_type::vec2_type
);
3779 src0
= v
->get_temp(glsl_type::vec4_type
);
3780 dst0
= st_dst_reg(src0
);
3781 inst
= v
->emit(NULL
, TGSI_OPCODE_TEX
, dst0
, coord
);
3783 inst
->tex_target
= TEXTURE_2D_INDEX
;
3785 prog
->InputsRead
|= (1 << FRAG_ATTRIB_TEX0
);
3786 prog
->SamplersUsed
|= (1 << 0); /* mark sampler 0 as used */
3787 v
->samplers_used
|= (1 << 0);
3789 if (scale_and_bias
) {
3790 static const gl_state_index scale_state
[STATE_LENGTH
] =
3791 { STATE_INTERNAL
, STATE_PT_SCALE
,
3792 (gl_state_index
) 0, (gl_state_index
) 0, (gl_state_index
) 0 };
3793 static const gl_state_index bias_state
[STATE_LENGTH
] =
3794 { STATE_INTERNAL
, STATE_PT_BIAS
,
3795 (gl_state_index
) 0, (gl_state_index
) 0, (gl_state_index
) 0 };
3796 GLint scale_p
, bias_p
;
3797 st_src_reg scale
, bias
;
3799 scale_p
= _mesa_add_state_reference(params
, scale_state
);
3800 bias_p
= _mesa_add_state_reference(params
, bias_state
);
3802 /* MAD colorTemp, colorTemp, scale, bias; */
3803 scale
= st_src_reg(PROGRAM_STATE_VAR
, scale_p
, GLSL_TYPE_FLOAT
);
3804 bias
= st_src_reg(PROGRAM_STATE_VAR
, bias_p
, GLSL_TYPE_FLOAT
);
3805 inst
= v
->emit(NULL
, TGSI_OPCODE_MAD
, dst0
, src0
, scale
, bias
);
3809 st_src_reg temp
= v
->get_temp(glsl_type::vec4_type
);
3810 st_dst_reg temp_dst
= st_dst_reg(temp
);
3812 assert(st
->pixel_xfer
.pixelmap_texture
);
3814 /* With a little effort, we can do four pixel map look-ups with
3815 * two TEX instructions:
3818 /* TEX temp.rg, colorTemp.rgba, texture[1], 2D; */
3819 temp_dst
.writemask
= WRITEMASK_XY
; /* write R,G */
3820 inst
= v
->emit(NULL
, TGSI_OPCODE_TEX
, temp_dst
, src0
);
3822 inst
->tex_target
= TEXTURE_2D_INDEX
;
3824 /* TEX temp.ba, colorTemp.baba, texture[1], 2D; */
3825 src0
.swizzle
= MAKE_SWIZZLE4(SWIZZLE_Z
, SWIZZLE_W
, SWIZZLE_Z
, SWIZZLE_W
);
3826 temp_dst
.writemask
= WRITEMASK_ZW
; /* write B,A */
3827 inst
= v
->emit(NULL
, TGSI_OPCODE_TEX
, temp_dst
, src0
);
3829 inst
->tex_target
= TEXTURE_2D_INDEX
;
3831 prog
->SamplersUsed
|= (1 << 1); /* mark sampler 1 as used */
3832 v
->samplers_used
|= (1 << 1);
3834 /* MOV colorTemp, temp; */
3835 inst
= v
->emit(NULL
, TGSI_OPCODE_MOV
, dst0
, temp
);
3838 /* Now copy the instructions from the original glsl_to_tgsi_visitor into the
3840 foreach_iter(exec_list_iterator
, iter
, original
->instructions
) {
3841 glsl_to_tgsi_instruction
*inst
= (glsl_to_tgsi_instruction
*)iter
.get();
3842 st_src_reg src_regs
[3];
3844 if (inst
->dst
.file
== PROGRAM_OUTPUT
)
3845 prog
->OutputsWritten
|= BITFIELD64_BIT(inst
->dst
.index
);
3847 for (int i
=0; i
<3; i
++) {
3848 src_regs
[i
] = inst
->src
[i
];
3849 if (src_regs
[i
].file
== PROGRAM_INPUT
&&
3850 src_regs
[i
].index
== FRAG_ATTRIB_COL0
)
3852 src_regs
[i
].file
= PROGRAM_TEMPORARY
;
3853 src_regs
[i
].index
= src0
.index
;
3855 else if (src_regs
[i
].file
== PROGRAM_INPUT
)
3856 prog
->InputsRead
|= (1 << src_regs
[i
].index
);
3859 v
->emit(NULL
, inst
->op
, inst
->dst
, src_regs
[0], src_regs
[1], src_regs
[2]);
3862 /* Make modifications to fragment program info. */
3863 prog
->Parameters
= _mesa_combine_parameter_lists(params
,
3864 original
->prog
->Parameters
);
3865 prog
->Attributes
= _mesa_clone_parameter_list(original
->prog
->Attributes
);
3866 prog
->Varying
= _mesa_clone_parameter_list(original
->prog
->Varying
);
3867 _mesa_free_parameter_list(params
);
3868 count_resources(v
, prog
);
3869 fp
->glsl_to_tgsi
= v
;
3873 * Make fragment program for glBitmap:
3874 * Sample the texture and kill the fragment if the bit is 0.
3875 * This program will be combined with the user's fragment program.
3877 * Based on make_bitmap_fragment_program in st_cb_bitmap.c.
3880 get_bitmap_visitor(struct st_fragment_program
*fp
,
3881 glsl_to_tgsi_visitor
*original
, int samplerIndex
)
3883 glsl_to_tgsi_visitor
*v
= new glsl_to_tgsi_visitor();
3884 struct st_context
*st
= st_context(original
->ctx
);
3885 struct gl_program
*prog
= &fp
->Base
.Base
;
3886 st_src_reg coord
, src0
;
3888 glsl_to_tgsi_instruction
*inst
;
3890 /* Copy attributes of the glsl_to_tgsi_visitor in the original shader. */
3891 v
->ctx
= original
->ctx
;
3893 v
->glsl_version
= original
->glsl_version
;
3894 v
->native_integers
= original
->native_integers
;
3895 v
->options
= original
->options
;
3896 v
->next_temp
= original
->next_temp
;
3897 v
->num_address_regs
= original
->num_address_regs
;
3898 v
->samplers_used
= prog
->SamplersUsed
= original
->samplers_used
;
3899 v
->indirect_addr_temps
= original
->indirect_addr_temps
;
3900 v
->indirect_addr_consts
= original
->indirect_addr_consts
;
3901 memcpy(&v
->immediates
, &original
->immediates
, sizeof(v
->immediates
));
3903 /* TEX tmp0, fragment.texcoord[0], texture[0], 2D; */
3904 coord
= st_src_reg(PROGRAM_INPUT
, FRAG_ATTRIB_TEX0
, glsl_type::vec2_type
);
3905 src0
= v
->get_temp(glsl_type::vec4_type
);
3906 dst0
= st_dst_reg(src0
);
3907 inst
= v
->emit(NULL
, TGSI_OPCODE_TEX
, dst0
, coord
);
3908 inst
->sampler
= samplerIndex
;
3909 inst
->tex_target
= TEXTURE_2D_INDEX
;
3911 prog
->InputsRead
|= (1 << FRAG_ATTRIB_TEX0
);
3912 prog
->SamplersUsed
|= (1 << samplerIndex
); /* mark sampler as used */
3913 v
->samplers_used
|= (1 << samplerIndex
);
3915 /* KIL if -tmp0 < 0 # texel=0 -> keep / texel=0 -> discard */
3916 src0
.negate
= NEGATE_XYZW
;
3917 if (st
->bitmap
.tex_format
== PIPE_FORMAT_L8_UNORM
)
3918 src0
.swizzle
= SWIZZLE_XXXX
;
3919 inst
= v
->emit(NULL
, TGSI_OPCODE_KIL
, undef_dst
, src0
);
3921 /* Now copy the instructions from the original glsl_to_tgsi_visitor into the
3923 foreach_iter(exec_list_iterator
, iter
, original
->instructions
) {
3924 glsl_to_tgsi_instruction
*inst
= (glsl_to_tgsi_instruction
*)iter
.get();
3925 st_src_reg src_regs
[3];
3927 if (inst
->dst
.file
== PROGRAM_OUTPUT
)
3928 prog
->OutputsWritten
|= BITFIELD64_BIT(inst
->dst
.index
);
3930 for (int i
=0; i
<3; i
++) {
3931 src_regs
[i
] = inst
->src
[i
];
3932 if (src_regs
[i
].file
== PROGRAM_INPUT
)
3933 prog
->InputsRead
|= (1 << src_regs
[i
].index
);
3936 v
->emit(NULL
, inst
->op
, inst
->dst
, src_regs
[0], src_regs
[1], src_regs
[2]);
3939 /* Make modifications to fragment program info. */
3940 prog
->Parameters
= _mesa_clone_parameter_list(original
->prog
->Parameters
);
3941 prog
->Attributes
= _mesa_clone_parameter_list(original
->prog
->Attributes
);
3942 prog
->Varying
= _mesa_clone_parameter_list(original
->prog
->Varying
);
3943 count_resources(v
, prog
);
3944 fp
->glsl_to_tgsi
= v
;
3947 /* ------------------------- TGSI conversion stuff -------------------------- */
3949 unsigned branch_target
;
3954 * Intermediate state used during shader translation.
3956 struct st_translate
{
3957 struct ureg_program
*ureg
;
3959 struct ureg_dst temps
[MAX_TEMPS
];
3960 struct ureg_src
*constants
;
3961 struct ureg_src
*immediates
;
3962 struct ureg_dst outputs
[PIPE_MAX_SHADER_OUTPUTS
];
3963 struct ureg_src inputs
[PIPE_MAX_SHADER_INPUTS
];
3964 struct ureg_dst address
[1];
3965 struct ureg_src samplers
[PIPE_MAX_SAMPLERS
];
3966 struct ureg_src systemValues
[SYSTEM_VALUE_MAX
];
3968 /* Extra info for handling point size clamping in vertex shader */
3969 struct ureg_dst pointSizeResult
; /**< Actual point size output register */
3970 struct ureg_src pointSizeConst
; /**< Point size range constant register */
3971 GLint pointSizeOutIndex
; /**< Temp point size output register */
3972 GLboolean prevInstWrotePointSize
;
3974 const GLuint
*inputMapping
;
3975 const GLuint
*outputMapping
;
3977 /* For every instruction that contains a label (eg CALL), keep
3978 * details so that we can go back afterwards and emit the correct
3979 * tgsi instruction number for each label.
3981 struct label
*labels
;
3982 unsigned labels_size
;
3983 unsigned labels_count
;
3985 /* Keep a record of the tgsi instruction number that each mesa
3986 * instruction starts at, will be used to fix up labels after
3991 unsigned insn_count
;
3993 unsigned procType
; /**< TGSI_PROCESSOR_VERTEX/FRAGMENT */
3998 /** Map Mesa's SYSTEM_VALUE_x to TGSI_SEMANTIC_x */
3999 static unsigned mesa_sysval_to_semantic
[SYSTEM_VALUE_MAX
] = {
4001 TGSI_SEMANTIC_INSTANCEID
4005 * Make note of a branch to a label in the TGSI code.
4006 * After we've emitted all instructions, we'll go over the list
4007 * of labels built here and patch the TGSI code with the actual
4008 * location of each label.
4010 static unsigned *get_label(struct st_translate
*t
, unsigned branch_target
)
4014 if (t
->labels_count
+ 1 >= t
->labels_size
) {
4015 t
->labels_size
= 1 << (util_logbase2(t
->labels_size
) + 1);
4016 t
->labels
= (struct label
*)realloc(t
->labels
,
4017 t
->labels_size
* sizeof(struct label
));
4018 if (t
->labels
== NULL
) {
4019 static unsigned dummy
;
4025 i
= t
->labels_count
++;
4026 t
->labels
[i
].branch_target
= branch_target
;
4027 return &t
->labels
[i
].token
;
4031 * Called prior to emitting the TGSI code for each instruction.
4032 * Allocate additional space for instructions if needed.
4033 * Update the insn[] array so the next glsl_to_tgsi_instruction points to
4034 * the next TGSI instruction.
4036 static void set_insn_start(struct st_translate
*t
, unsigned start
)
4038 if (t
->insn_count
+ 1 >= t
->insn_size
) {
4039 t
->insn_size
= 1 << (util_logbase2(t
->insn_size
) + 1);
4040 t
->insn
= (unsigned *)realloc(t
->insn
, t
->insn_size
* sizeof(t
->insn
[0]));
4041 if (t
->insn
== NULL
) {
4047 t
->insn
[t
->insn_count
++] = start
;
4051 * Map a glsl_to_tgsi constant/immediate to a TGSI immediate.
4053 static struct ureg_src
4054 emit_immediate(struct st_translate
*t
,
4055 gl_constant_value values
[4],
4058 struct ureg_program
*ureg
= t
->ureg
;
4063 return ureg_DECL_immediate(ureg
, &values
[0].f
, size
);
4065 return ureg_DECL_immediate_int(ureg
, &values
[0].i
, size
);
4066 case GL_UNSIGNED_INT
:
4068 return ureg_DECL_immediate_uint(ureg
, &values
[0].u
, size
);
4070 assert(!"should not get here - type must be float, int, uint, or bool");
4071 return ureg_src_undef();
4076 * Map a glsl_to_tgsi dst register to a TGSI ureg_dst register.
4078 static struct ureg_dst
4079 dst_register(struct st_translate
*t
,
4080 gl_register_file file
,
4084 case PROGRAM_UNDEFINED
:
4085 return ureg_dst_undef();
4087 case PROGRAM_TEMPORARY
:
4088 if (ureg_dst_is_undef(t
->temps
[index
]))
4089 t
->temps
[index
] = ureg_DECL_temporary(t
->ureg
);
4091 return t
->temps
[index
];
4093 case PROGRAM_OUTPUT
:
4094 if (t
->procType
== TGSI_PROCESSOR_VERTEX
&& index
== VERT_RESULT_PSIZ
)
4095 t
->prevInstWrotePointSize
= GL_TRUE
;
4097 if (t
->procType
== TGSI_PROCESSOR_VERTEX
)
4098 assert(index
< VERT_RESULT_MAX
);
4099 else if (t
->procType
== TGSI_PROCESSOR_FRAGMENT
)
4100 assert(index
< FRAG_RESULT_MAX
);
4102 assert(index
< GEOM_RESULT_MAX
);
4104 assert(t
->outputMapping
[index
] < Elements(t
->outputs
));
4106 return t
->outputs
[t
->outputMapping
[index
]];
4108 case PROGRAM_ADDRESS
:
4109 return t
->address
[index
];
4112 assert(!"unknown dst register file");
4113 return ureg_dst_undef();
4118 * Map a glsl_to_tgsi src register to a TGSI ureg_src register.
4120 static struct ureg_src
4121 src_register(struct st_translate
*t
,
4122 gl_register_file file
,
4126 case PROGRAM_UNDEFINED
:
4127 return ureg_src_undef();
4129 case PROGRAM_TEMPORARY
:
4131 assert(index
< Elements(t
->temps
));
4132 if (ureg_dst_is_undef(t
->temps
[index
]))
4133 t
->temps
[index
] = ureg_DECL_temporary(t
->ureg
);
4134 return ureg_src(t
->temps
[index
]);
4136 case PROGRAM_NAMED_PARAM
:
4137 case PROGRAM_ENV_PARAM
:
4138 case PROGRAM_LOCAL_PARAM
:
4139 case PROGRAM_UNIFORM
:
4141 return t
->constants
[index
];
4142 case PROGRAM_STATE_VAR
:
4143 case PROGRAM_CONSTANT
: /* ie, immediate */
4145 return ureg_DECL_constant(t
->ureg
, 0);
4147 return t
->constants
[index
];
4149 case PROGRAM_IMMEDIATE
:
4150 return t
->immediates
[index
];
4153 assert(t
->inputMapping
[index
] < Elements(t
->inputs
));
4154 return t
->inputs
[t
->inputMapping
[index
]];
4156 case PROGRAM_OUTPUT
:
4157 assert(t
->outputMapping
[index
] < Elements(t
->outputs
));
4158 return ureg_src(t
->outputs
[t
->outputMapping
[index
]]); /* not needed? */
4160 case PROGRAM_ADDRESS
:
4161 return ureg_src(t
->address
[index
]);
4163 case PROGRAM_SYSTEM_VALUE
:
4164 assert(index
< Elements(t
->systemValues
));
4165 return t
->systemValues
[index
];
4168 assert(!"unknown src register file");
4169 return ureg_src_undef();
4174 * Create a TGSI ureg_dst register from an st_dst_reg.
4176 static struct ureg_dst
4177 translate_dst(struct st_translate
*t
,
4178 const st_dst_reg
*dst_reg
,
4181 struct ureg_dst dst
= dst_register(t
,
4185 dst
= ureg_writemask(dst
, dst_reg
->writemask
);
4188 dst
= ureg_saturate(dst
);
4190 if (dst_reg
->reladdr
!= NULL
)
4191 dst
= ureg_dst_indirect(dst
, ureg_src(t
->address
[0]));
4197 * Create a TGSI ureg_src register from an st_src_reg.
4199 static struct ureg_src
4200 translate_src(struct st_translate
*t
, const st_src_reg
*src_reg
)
4202 struct ureg_src src
= src_register(t
, src_reg
->file
, src_reg
->index
);
4204 src
= ureg_swizzle(src
,
4205 GET_SWZ(src_reg
->swizzle
, 0) & 0x3,
4206 GET_SWZ(src_reg
->swizzle
, 1) & 0x3,
4207 GET_SWZ(src_reg
->swizzle
, 2) & 0x3,
4208 GET_SWZ(src_reg
->swizzle
, 3) & 0x3);
4210 if ((src_reg
->negate
& 0xf) == NEGATE_XYZW
)
4211 src
= ureg_negate(src
);
4213 if (src_reg
->reladdr
!= NULL
) {
4214 /* Normally ureg_src_indirect() would be used here, but a stupid compiler
4215 * bug in g++ makes ureg_src_indirect (an inline C function) erroneously
4216 * set the bit for src.Negate. So we have to do the operation manually
4217 * here to work around the compiler's problems. */
4218 /*src = ureg_src_indirect(src, ureg_src(t->address[0]));*/
4219 struct ureg_src addr
= ureg_src(t
->address
[0]);
4221 src
.IndirectFile
= addr
.File
;
4222 src
.IndirectIndex
= addr
.Index
;
4223 src
.IndirectSwizzle
= addr
.SwizzleX
;
4225 if (src_reg
->file
!= PROGRAM_INPUT
&&
4226 src_reg
->file
!= PROGRAM_OUTPUT
) {
4227 /* If src_reg->index was negative, it was set to zero in
4228 * src_register(). Reassign it now. But don't do this
4229 * for input/output regs since they get remapped while
4230 * const buffers don't.
4232 src
.Index
= src_reg
->index
;
4239 static struct tgsi_texture_offset
4240 translate_tex_offset(struct st_translate
*t
,
4241 const struct tgsi_texture_offset
*in_offset
)
4243 struct tgsi_texture_offset offset
;
4245 assert(in_offset
->File
== PROGRAM_IMMEDIATE
);
4247 offset
.File
= TGSI_FILE_IMMEDIATE
;
4248 offset
.Index
= in_offset
->Index
;
4249 offset
.SwizzleX
= in_offset
->SwizzleX
;
4250 offset
.SwizzleY
= in_offset
->SwizzleY
;
4251 offset
.SwizzleZ
= in_offset
->SwizzleZ
;
4257 compile_tgsi_instruction(struct st_translate
*t
,
4258 const glsl_to_tgsi_instruction
*inst
)
4260 struct ureg_program
*ureg
= t
->ureg
;
4262 struct ureg_dst dst
[1];
4263 struct ureg_src src
[4];
4264 struct tgsi_texture_offset texoffsets
[MAX_GLSL_TEXTURE_OFFSET
];
4269 num_dst
= num_inst_dst_regs(inst
->op
);
4270 num_src
= num_inst_src_regs(inst
->op
);
4273 dst
[0] = translate_dst(t
,
4277 for (i
= 0; i
< num_src
; i
++)
4278 src
[i
] = translate_src(t
, &inst
->src
[i
]);
4281 case TGSI_OPCODE_BGNLOOP
:
4282 case TGSI_OPCODE_CAL
:
4283 case TGSI_OPCODE_ELSE
:
4284 case TGSI_OPCODE_ENDLOOP
:
4285 case TGSI_OPCODE_IF
:
4286 assert(num_dst
== 0);
4287 ureg_label_insn(ureg
,
4291 inst
->op
== TGSI_OPCODE_CAL
? inst
->function
->sig_id
: 0));
4294 case TGSI_OPCODE_TEX
:
4295 case TGSI_OPCODE_TXB
:
4296 case TGSI_OPCODE_TXD
:
4297 case TGSI_OPCODE_TXL
:
4298 case TGSI_OPCODE_TXP
:
4299 case TGSI_OPCODE_TXQ
:
4300 case TGSI_OPCODE_TXF
:
4301 src
[num_src
++] = t
->samplers
[inst
->sampler
];
4302 for (i
= 0; i
< inst
->tex_offset_num_offset
; i
++) {
4303 texoffsets
[i
] = translate_tex_offset(t
, &inst
->tex_offsets
[i
]);
4308 translate_texture_target(inst
->tex_target
, inst
->tex_shadow
),
4309 texoffsets
, inst
->tex_offset_num_offset
,
4313 case TGSI_OPCODE_SCS
:
4314 dst
[0] = ureg_writemask(dst
[0], TGSI_WRITEMASK_XY
);
4315 ureg_insn(ureg
, inst
->op
, dst
, num_dst
, src
, num_src
);
4328 * Emit the TGSI instructions to adjust the WPOS pixel center convention
4329 * Basically, add (adjX, adjY) to the fragment position.
4332 emit_adjusted_wpos(struct st_translate
*t
,
4333 const struct gl_program
*program
,
4334 float adjX
, float adjY
)
4336 struct ureg_program
*ureg
= t
->ureg
;
4337 struct ureg_dst wpos_temp
= ureg_DECL_temporary(ureg
);
4338 struct ureg_src wpos_input
= t
->inputs
[t
->inputMapping
[FRAG_ATTRIB_WPOS
]];
4340 /* Note that we bias X and Y and pass Z and W through unchanged.
4341 * The shader might also use gl_FragCoord.w and .z.
4343 ureg_ADD(ureg
, wpos_temp
, wpos_input
,
4344 ureg_imm4f(ureg
, adjX
, adjY
, 0.0f
, 0.0f
));
4346 t
->inputs
[t
->inputMapping
[FRAG_ATTRIB_WPOS
]] = ureg_src(wpos_temp
);
4351 * Emit the TGSI instructions for inverting the WPOS y coordinate.
4352 * This code is unavoidable because it also depends on whether
4353 * a FBO is bound (STATE_FB_WPOS_Y_TRANSFORM).
4356 emit_wpos_inversion(struct st_translate
*t
,
4357 const struct gl_program
*program
,
4360 struct ureg_program
*ureg
= t
->ureg
;
4362 /* Fragment program uses fragment position input.
4363 * Need to replace instances of INPUT[WPOS] with temp T
4364 * where T = INPUT[WPOS] by y is inverted.
4366 static const gl_state_index wposTransformState
[STATE_LENGTH
]
4367 = { STATE_INTERNAL
, STATE_FB_WPOS_Y_TRANSFORM
,
4368 (gl_state_index
)0, (gl_state_index
)0, (gl_state_index
)0 };
4370 /* XXX: note we are modifying the incoming shader here! Need to
4371 * do this before emitting the constant decls below, or this
4374 unsigned wposTransConst
= _mesa_add_state_reference(program
->Parameters
,
4375 wposTransformState
);
4377 struct ureg_src wpostrans
= ureg_DECL_constant(ureg
, wposTransConst
);
4378 struct ureg_dst wpos_temp
;
4379 struct ureg_src wpos_input
= t
->inputs
[t
->inputMapping
[FRAG_ATTRIB_WPOS
]];
4381 /* MOV wpos_temp, input[wpos]
4383 if (wpos_input
.File
== TGSI_FILE_TEMPORARY
)
4384 wpos_temp
= ureg_dst(wpos_input
);
4386 wpos_temp
= ureg_DECL_temporary(ureg
);
4387 ureg_MOV(ureg
, wpos_temp
, wpos_input
);
4391 /* MAD wpos_temp.y, wpos_input, wpostrans.xxxx, wpostrans.yyyy
4394 ureg_writemask(wpos_temp
, TGSI_WRITEMASK_Y
),
4396 ureg_scalar(wpostrans
, 0),
4397 ureg_scalar(wpostrans
, 1));
4399 /* MAD wpos_temp.y, wpos_input, wpostrans.zzzz, wpostrans.wwww
4402 ureg_writemask(wpos_temp
, TGSI_WRITEMASK_Y
),
4404 ureg_scalar(wpostrans
, 2),
4405 ureg_scalar(wpostrans
, 3));
4408 /* Use wpos_temp as position input from here on:
4410 t
->inputs
[t
->inputMapping
[FRAG_ATTRIB_WPOS
]] = ureg_src(wpos_temp
);
4415 * Emit fragment position/ooordinate code.
4418 emit_wpos(struct st_context
*st
,
4419 struct st_translate
*t
,
4420 const struct gl_program
*program
,
4421 struct ureg_program
*ureg
)
4423 const struct gl_fragment_program
*fp
=
4424 (const struct gl_fragment_program
*) program
;
4425 struct pipe_screen
*pscreen
= st
->pipe
->screen
;
4426 boolean invert
= FALSE
;
4428 if (fp
->OriginUpperLeft
) {
4429 /* Fragment shader wants origin in upper-left */
4430 if (pscreen
->get_param(pscreen
, PIPE_CAP_TGSI_FS_COORD_ORIGIN_UPPER_LEFT
)) {
4431 /* the driver supports upper-left origin */
4433 else if (pscreen
->get_param(pscreen
, PIPE_CAP_TGSI_FS_COORD_ORIGIN_LOWER_LEFT
)) {
4434 /* the driver supports lower-left origin, need to invert Y */
4435 ureg_property_fs_coord_origin(ureg
, TGSI_FS_COORD_ORIGIN_LOWER_LEFT
);
4442 /* Fragment shader wants origin in lower-left */
4443 if (pscreen
->get_param(pscreen
, PIPE_CAP_TGSI_FS_COORD_ORIGIN_LOWER_LEFT
))
4444 /* the driver supports lower-left origin */
4445 ureg_property_fs_coord_origin(ureg
, TGSI_FS_COORD_ORIGIN_LOWER_LEFT
);
4446 else if (pscreen
->get_param(pscreen
, PIPE_CAP_TGSI_FS_COORD_ORIGIN_UPPER_LEFT
))
4447 /* the driver supports upper-left origin, need to invert Y */
4453 if (fp
->PixelCenterInteger
) {
4454 /* Fragment shader wants pixel center integer */
4455 if (pscreen
->get_param(pscreen
, PIPE_CAP_TGSI_FS_COORD_PIXEL_CENTER_INTEGER
))
4456 /* the driver supports pixel center integer */
4457 ureg_property_fs_coord_pixel_center(ureg
, TGSI_FS_COORD_PIXEL_CENTER_INTEGER
);
4458 else if (pscreen
->get_param(pscreen
, PIPE_CAP_TGSI_FS_COORD_PIXEL_CENTER_HALF_INTEGER
))
4459 /* the driver supports pixel center half integer, need to bias X,Y */
4460 emit_adjusted_wpos(t
, program
, 0.5f
, invert
? 0.5f
: -0.5f
);
4465 /* Fragment shader wants pixel center half integer */
4466 if (pscreen
->get_param(pscreen
, PIPE_CAP_TGSI_FS_COORD_PIXEL_CENTER_HALF_INTEGER
)) {
4467 /* the driver supports pixel center half integer */
4469 else if (pscreen
->get_param(pscreen
, PIPE_CAP_TGSI_FS_COORD_PIXEL_CENTER_INTEGER
)) {
4470 /* the driver supports pixel center integer, need to bias X,Y */
4471 ureg_property_fs_coord_pixel_center(ureg
, TGSI_FS_COORD_PIXEL_CENTER_INTEGER
);
4472 emit_adjusted_wpos(t
, program
, 0.5f
, invert
? -0.5f
: 0.5f
);
4478 /* we invert after adjustment so that we avoid the MOV to temporary,
4479 * and reuse the adjustment ADD instead */
4480 emit_wpos_inversion(t
, program
, invert
);
4484 * OpenGL's fragment gl_FrontFace input is 1 for front-facing, 0 for back.
4485 * TGSI uses +1 for front, -1 for back.
4486 * This function converts the TGSI value to the GL value. Simply clamping/
4487 * saturating the value to [0,1] does the job.
4490 emit_face_var(struct st_translate
*t
)
4492 struct ureg_program
*ureg
= t
->ureg
;
4493 struct ureg_dst face_temp
= ureg_DECL_temporary(ureg
);
4494 struct ureg_src face_input
= t
->inputs
[t
->inputMapping
[FRAG_ATTRIB_FACE
]];
4496 /* MOV_SAT face_temp, input[face] */
4497 face_temp
= ureg_saturate(face_temp
);
4498 ureg_MOV(ureg
, face_temp
, face_input
);
4500 /* Use face_temp as face input from here on: */
4501 t
->inputs
[t
->inputMapping
[FRAG_ATTRIB_FACE
]] = ureg_src(face_temp
);
4505 emit_edgeflags(struct st_translate
*t
)
4507 struct ureg_program
*ureg
= t
->ureg
;
4508 struct ureg_dst edge_dst
= t
->outputs
[t
->outputMapping
[VERT_RESULT_EDGE
]];
4509 struct ureg_src edge_src
= t
->inputs
[t
->inputMapping
[VERT_ATTRIB_EDGEFLAG
]];
4511 ureg_MOV(ureg
, edge_dst
, edge_src
);
4515 * Translate intermediate IR (glsl_to_tgsi_instruction) to TGSI format.
4516 * \param program the program to translate
4517 * \param numInputs number of input registers used
4518 * \param inputMapping maps Mesa fragment program inputs to TGSI generic
4520 * \param inputSemanticName the TGSI_SEMANTIC flag for each input
4521 * \param inputSemanticIndex the semantic index (ex: which texcoord) for
4523 * \param interpMode the TGSI_INTERPOLATE_LINEAR/PERSP mode for each input
4524 * \param numOutputs number of output registers used
4525 * \param outputMapping maps Mesa fragment program outputs to TGSI
4527 * \param outputSemanticName the TGSI_SEMANTIC flag for each output
4528 * \param outputSemanticIndex the semantic index (ex: which texcoord) for
4531 * \return PIPE_OK or PIPE_ERROR_OUT_OF_MEMORY
4533 extern "C" enum pipe_error
4534 st_translate_program(
4535 struct gl_context
*ctx
,
4537 struct ureg_program
*ureg
,
4538 glsl_to_tgsi_visitor
*program
,
4539 const struct gl_program
*proginfo
,
4541 const GLuint inputMapping
[],
4542 const ubyte inputSemanticName
[],
4543 const ubyte inputSemanticIndex
[],
4544 const GLuint interpMode
[],
4546 const GLuint outputMapping
[],
4547 const ubyte outputSemanticName
[],
4548 const ubyte outputSemanticIndex
[],
4549 boolean passthrough_edgeflags
)
4551 struct st_translate translate
, *t
;
4553 enum pipe_error ret
= PIPE_OK
;
4555 assert(numInputs
<= Elements(t
->inputs
));
4556 assert(numOutputs
<= Elements(t
->outputs
));
4559 memset(t
, 0, sizeof *t
);
4561 t
->procType
= procType
;
4562 t
->inputMapping
= inputMapping
;
4563 t
->outputMapping
= outputMapping
;
4565 t
->pointSizeOutIndex
= -1;
4566 t
->prevInstWrotePointSize
= GL_FALSE
;
4569 * Declare input attributes.
4571 if (procType
== TGSI_PROCESSOR_FRAGMENT
) {
4572 for (i
= 0; i
< numInputs
; i
++) {
4573 t
->inputs
[i
] = ureg_DECL_fs_input(ureg
,
4574 inputSemanticName
[i
],
4575 inputSemanticIndex
[i
],
4579 if (proginfo
->InputsRead
& FRAG_BIT_WPOS
) {
4580 /* Must do this after setting up t->inputs, and before
4581 * emitting constant references, below:
4583 emit_wpos(st_context(ctx
), t
, proginfo
, ureg
);
4586 if (proginfo
->InputsRead
& FRAG_BIT_FACE
)
4590 * Declare output attributes.
4592 for (i
= 0; i
< numOutputs
; i
++) {
4593 switch (outputSemanticName
[i
]) {
4594 case TGSI_SEMANTIC_POSITION
:
4595 t
->outputs
[i
] = ureg_DECL_output(ureg
,
4596 TGSI_SEMANTIC_POSITION
, /* Z/Depth */
4597 outputSemanticIndex
[i
]);
4598 t
->outputs
[i
] = ureg_writemask(t
->outputs
[i
], TGSI_WRITEMASK_Z
);
4600 case TGSI_SEMANTIC_STENCIL
:
4601 t
->outputs
[i
] = ureg_DECL_output(ureg
,
4602 TGSI_SEMANTIC_STENCIL
, /* Stencil */
4603 outputSemanticIndex
[i
]);
4604 t
->outputs
[i
] = ureg_writemask(t
->outputs
[i
], TGSI_WRITEMASK_Y
);
4606 case TGSI_SEMANTIC_COLOR
:
4607 t
->outputs
[i
] = ureg_DECL_output(ureg
,
4608 TGSI_SEMANTIC_COLOR
,
4609 outputSemanticIndex
[i
]);
4612 assert(!"fragment shader outputs must be POSITION/STENCIL/COLOR");
4613 return PIPE_ERROR_BAD_INPUT
;
4617 else if (procType
== TGSI_PROCESSOR_GEOMETRY
) {
4618 for (i
= 0; i
< numInputs
; i
++) {
4619 t
->inputs
[i
] = ureg_DECL_gs_input(ureg
,
4621 inputSemanticName
[i
],
4622 inputSemanticIndex
[i
]);
4625 for (i
= 0; i
< numOutputs
; i
++) {
4626 t
->outputs
[i
] = ureg_DECL_output(ureg
,
4627 outputSemanticName
[i
],
4628 outputSemanticIndex
[i
]);
4632 assert(procType
== TGSI_PROCESSOR_VERTEX
);
4634 for (i
= 0; i
< numInputs
; i
++) {
4635 t
->inputs
[i
] = ureg_DECL_vs_input(ureg
, i
);
4638 for (i
= 0; i
< numOutputs
; i
++) {
4639 t
->outputs
[i
] = ureg_DECL_output(ureg
,
4640 outputSemanticName
[i
],
4641 outputSemanticIndex
[i
]);
4642 if ((outputSemanticName
[i
] == TGSI_SEMANTIC_PSIZE
) && proginfo
->Id
) {
4643 /* Writing to the point size result register requires special
4644 * handling to implement clamping.
4646 static const gl_state_index pointSizeClampState
[STATE_LENGTH
]
4647 = { STATE_INTERNAL
, STATE_POINT_SIZE_IMPL_CLAMP
, (gl_state_index
)0, (gl_state_index
)0, (gl_state_index
)0 };
4648 /* XXX: note we are modifying the incoming shader here! Need to
4649 * do this before emitting the constant decls below, or this
4652 unsigned pointSizeClampConst
=
4653 _mesa_add_state_reference(proginfo
->Parameters
,
4654 pointSizeClampState
);
4655 struct ureg_dst psizregtemp
= ureg_DECL_temporary(ureg
);
4656 t
->pointSizeConst
= ureg_DECL_constant(ureg
, pointSizeClampConst
);
4657 t
->pointSizeResult
= t
->outputs
[i
];
4658 t
->pointSizeOutIndex
= i
;
4659 t
->outputs
[i
] = psizregtemp
;
4662 if (passthrough_edgeflags
)
4666 /* Declare address register.
4668 if (program
->num_address_regs
> 0) {
4669 assert(program
->num_address_regs
== 1);
4670 t
->address
[0] = ureg_DECL_address(ureg
);
4673 /* Declare misc input registers
4676 GLbitfield sysInputs
= proginfo
->SystemValuesRead
;
4677 unsigned numSys
= 0;
4678 for (i
= 0; sysInputs
; i
++) {
4679 if (sysInputs
& (1 << i
)) {
4680 unsigned semName
= mesa_sysval_to_semantic
[i
];
4681 t
->systemValues
[i
] = ureg_DECL_system_value(ureg
, numSys
, semName
, 0);
4683 sysInputs
&= ~(1 << i
);
4688 if (program
->indirect_addr_temps
) {
4689 /* If temps are accessed with indirect addressing, declare temporaries
4690 * in sequential order. Else, we declare them on demand elsewhere.
4691 * (Note: the number of temporaries is equal to program->next_temp)
4693 for (i
= 0; i
< (unsigned)program
->next_temp
; i
++) {
4694 /* XXX use TGSI_FILE_TEMPORARY_ARRAY when it's supported by ureg */
4695 t
->temps
[i
] = ureg_DECL_temporary(t
->ureg
);
4699 /* Emit constants and uniforms. TGSI uses a single index space for these,
4700 * so we put all the translated regs in t->constants.
4702 if (proginfo
->Parameters
) {
4703 t
->constants
= (struct ureg_src
*)CALLOC(proginfo
->Parameters
->NumParameters
* sizeof(t
->constants
[0]));
4704 if (t
->constants
== NULL
) {
4705 ret
= PIPE_ERROR_OUT_OF_MEMORY
;
4709 for (i
= 0; i
< proginfo
->Parameters
->NumParameters
; i
++) {
4710 switch (proginfo
->Parameters
->Parameters
[i
].Type
) {
4711 case PROGRAM_ENV_PARAM
:
4712 case PROGRAM_LOCAL_PARAM
:
4713 case PROGRAM_STATE_VAR
:
4714 case PROGRAM_NAMED_PARAM
:
4715 case PROGRAM_UNIFORM
:
4716 t
->constants
[i
] = ureg_DECL_constant(ureg
, i
);
4719 /* Emit immediates for PROGRAM_CONSTANT only when there's no indirect
4720 * addressing of the const buffer.
4721 * FIXME: Be smarter and recognize param arrays:
4722 * indirect addressing is only valid within the referenced
4725 case PROGRAM_CONSTANT
:
4726 if (program
->indirect_addr_consts
)
4727 t
->constants
[i
] = ureg_DECL_constant(ureg
, i
);
4729 t
->constants
[i
] = emit_immediate(t
,
4730 proginfo
->Parameters
->ParameterValues
[i
],
4731 proginfo
->Parameters
->Parameters
[i
].DataType
,
4740 /* Emit immediate values.
4742 t
->immediates
= (struct ureg_src
*)CALLOC(program
->num_immediates
* sizeof(struct ureg_src
));
4743 if (t
->immediates
== NULL
) {
4744 ret
= PIPE_ERROR_OUT_OF_MEMORY
;
4748 foreach_iter(exec_list_iterator
, iter
, program
->immediates
) {
4749 immediate_storage
*imm
= (immediate_storage
*)iter
.get();
4750 t
->immediates
[i
++] = emit_immediate(t
, imm
->values
, imm
->type
, imm
->size
);
4753 /* texture samplers */
4754 for (i
= 0; i
< ctx
->Const
.MaxTextureImageUnits
; i
++) {
4755 if (program
->samplers_used
& (1 << i
)) {
4756 t
->samplers
[i
] = ureg_DECL_sampler(ureg
, i
);
4760 /* Emit each instruction in turn:
4762 foreach_iter(exec_list_iterator
, iter
, program
->instructions
) {
4763 set_insn_start(t
, ureg_get_instruction_number(ureg
));
4764 compile_tgsi_instruction(t
, (glsl_to_tgsi_instruction
*)iter
.get());
4766 if (t
->prevInstWrotePointSize
&& proginfo
->Id
) {
4767 /* The previous instruction wrote to the (fake) vertex point size
4768 * result register. Now we need to clamp that value to the min/max
4769 * point size range, putting the result into the real point size
4771 * Note that we can't do this easily at the end of program due to
4772 * possible early return.
4774 set_insn_start(t
, ureg_get_instruction_number(ureg
));
4776 ureg_writemask(t
->outputs
[t
->pointSizeOutIndex
], WRITEMASK_X
),
4777 ureg_src(t
->outputs
[t
->pointSizeOutIndex
]),
4778 ureg_swizzle(t
->pointSizeConst
, 1,1,1,1));
4779 ureg_MIN(t
->ureg
, ureg_writemask(t
->pointSizeResult
, WRITEMASK_X
),
4780 ureg_src(t
->outputs
[t
->pointSizeOutIndex
]),
4781 ureg_swizzle(t
->pointSizeConst
, 2,2,2,2));
4783 t
->prevInstWrotePointSize
= GL_FALSE
;
4786 /* Fix up all emitted labels:
4788 for (i
= 0; i
< t
->labels_count
; i
++) {
4789 ureg_fixup_label(ureg
, t
->labels
[i
].token
,
4790 t
->insn
[t
->labels
[i
].branch_target
]);
4797 FREE(t
->immediates
);
4800 debug_printf("%s: translate error flag set\n", __FUNCTION__
);
4805 /* ----------------------------- End TGSI code ------------------------------ */
4808 * Convert a shader's GLSL IR into a Mesa gl_program, although without
4809 * generating Mesa IR.
4811 static struct gl_program
*
4812 get_mesa_program(struct gl_context
*ctx
,
4813 struct gl_shader_program
*shader_program
,
4814 struct gl_shader
*shader
)
4816 glsl_to_tgsi_visitor
* v
= new glsl_to_tgsi_visitor();
4817 struct gl_program
*prog
;
4819 const char *target_string
;
4821 struct gl_shader_compiler_options
*options
=
4822 &ctx
->ShaderCompilerOptions
[_mesa_shader_type_to_index(shader
->Type
)];
4824 switch (shader
->Type
) {
4825 case GL_VERTEX_SHADER
:
4826 target
= GL_VERTEX_PROGRAM_ARB
;
4827 target_string
= "vertex";
4829 case GL_FRAGMENT_SHADER
:
4830 target
= GL_FRAGMENT_PROGRAM_ARB
;
4831 target_string
= "fragment";
4833 case GL_GEOMETRY_SHADER
:
4834 target
= GL_GEOMETRY_PROGRAM_NV
;
4835 target_string
= "geometry";
4838 assert(!"should not be reached");
4842 validate_ir_tree(shader
->ir
);
4844 prog
= ctx
->Driver
.NewProgram(ctx
, target
, shader_program
->Name
);
4847 prog
->Parameters
= _mesa_new_parameter_list();
4848 prog
->Varying
= _mesa_new_parameter_list();
4849 prog
->Attributes
= _mesa_new_parameter_list();
4852 v
->shader_program
= shader_program
;
4853 v
->options
= options
;
4854 v
->glsl_version
= ctx
->Const
.GLSLVersion
;
4855 v
->native_integers
= ctx
->Const
.NativeIntegers
;
4857 add_uniforms_to_parameters_list(shader_program
, shader
, prog
);
4859 /* Emit intermediate IR for main(). */
4860 visit_exec_list(shader
->ir
, v
);
4862 /* Now emit bodies for any functions that were used. */
4864 progress
= GL_FALSE
;
4866 foreach_iter(exec_list_iterator
, iter
, v
->function_signatures
) {
4867 function_entry
*entry
= (function_entry
*)iter
.get();
4869 if (!entry
->bgn_inst
) {
4870 v
->current_function
= entry
;
4872 entry
->bgn_inst
= v
->emit(NULL
, TGSI_OPCODE_BGNSUB
);
4873 entry
->bgn_inst
->function
= entry
;
4875 visit_exec_list(&entry
->sig
->body
, v
);
4877 glsl_to_tgsi_instruction
*last
;
4878 last
= (glsl_to_tgsi_instruction
*)v
->instructions
.get_tail();
4879 if (last
->op
!= TGSI_OPCODE_RET
)
4880 v
->emit(NULL
, TGSI_OPCODE_RET
);
4882 glsl_to_tgsi_instruction
*end
;
4883 end
= v
->emit(NULL
, TGSI_OPCODE_ENDSUB
);
4884 end
->function
= entry
;
4892 /* Print out some information (for debugging purposes) used by the
4893 * optimization passes. */
4894 for (i
=0; i
< v
->next_temp
; i
++) {
4895 int fr
= v
->get_first_temp_read(i
);
4896 int fw
= v
->get_first_temp_write(i
);
4897 int lr
= v
->get_last_temp_read(i
);
4898 int lw
= v
->get_last_temp_write(i
);
4900 printf("Temp %d: FR=%3d FW=%3d LR=%3d LW=%3d\n", i
, fr
, fw
, lr
, lw
);
4905 /* Remove reads to output registers, and to varyings in vertex shaders. */
4906 v
->remove_output_reads(PROGRAM_OUTPUT
);
4907 if (target
== GL_VERTEX_PROGRAM_ARB
)
4908 v
->remove_output_reads(PROGRAM_VARYING
);
4910 /* Perform optimizations on the instructions in the glsl_to_tgsi_visitor. */
4912 v
->copy_propagate();
4913 while (v
->eliminate_dead_code_advanced());
4915 /* FIXME: These passes to optimize temporary registers don't work when there
4916 * is indirect addressing of the temporary register space. We need proper
4917 * array support so that we don't have to give up these passes in every
4918 * shader that uses arrays.
4920 if (!v
->indirect_addr_temps
) {
4921 v
->eliminate_dead_code();
4922 v
->merge_registers();
4923 v
->renumber_registers();
4926 /* Write the END instruction. */
4927 v
->emit(NULL
, TGSI_OPCODE_END
);
4929 if (ctx
->Shader
.Flags
& GLSL_DUMP
) {
4931 printf("GLSL IR for linked %s program %d:\n", target_string
,
4932 shader_program
->Name
);
4933 _mesa_print_ir(shader
->ir
, NULL
);
4938 prog
->Instructions
= NULL
;
4939 prog
->NumInstructions
= 0;
4941 do_set_program_inouts(shader
->ir
, prog
);
4942 count_resources(v
, prog
);
4944 check_resources(ctx
, shader_program
, v
, prog
);
4946 _mesa_reference_program(ctx
, &shader
->Program
, prog
);
4948 struct st_vertex_program
*stvp
;
4949 struct st_fragment_program
*stfp
;
4950 struct st_geometry_program
*stgp
;
4952 switch (shader
->Type
) {
4953 case GL_VERTEX_SHADER
:
4954 stvp
= (struct st_vertex_program
*)prog
;
4955 stvp
->glsl_to_tgsi
= v
;
4957 case GL_FRAGMENT_SHADER
:
4958 stfp
= (struct st_fragment_program
*)prog
;
4959 stfp
->glsl_to_tgsi
= v
;
4961 case GL_GEOMETRY_SHADER
:
4962 stgp
= (struct st_geometry_program
*)prog
;
4963 stgp
->glsl_to_tgsi
= v
;
4966 assert(!"should not be reached");
4976 st_new_shader(struct gl_context
*ctx
, GLuint name
, GLuint type
)
4978 struct gl_shader
*shader
;
4979 assert(type
== GL_FRAGMENT_SHADER
|| type
== GL_VERTEX_SHADER
||
4980 type
== GL_GEOMETRY_SHADER_ARB
);
4981 shader
= rzalloc(NULL
, struct gl_shader
);
4983 shader
->Type
= type
;
4984 shader
->Name
= name
;
4985 _mesa_init_shader(ctx
, shader
);
4990 struct gl_shader_program
*
4991 st_new_shader_program(struct gl_context
*ctx
, GLuint name
)
4993 struct gl_shader_program
*shProg
;
4994 shProg
= rzalloc(NULL
, struct gl_shader_program
);
4996 shProg
->Name
= name
;
4997 _mesa_init_shader_program(ctx
, shProg
);
5004 * Called via ctx->Driver.LinkShader()
5005 * This actually involves converting GLSL IR into an intermediate TGSI-like IR
5006 * with code lowering and other optimizations.
5009 st_link_shader(struct gl_context
*ctx
, struct gl_shader_program
*prog
)
5011 assert(prog
->LinkStatus
);
5013 for (unsigned i
= 0; i
< MESA_SHADER_TYPES
; i
++) {
5014 if (prog
->_LinkedShaders
[i
] == NULL
)
5018 exec_list
*ir
= prog
->_LinkedShaders
[i
]->ir
;
5019 const struct gl_shader_compiler_options
*options
=
5020 &ctx
->ShaderCompilerOptions
[_mesa_shader_type_to_index(prog
->_LinkedShaders
[i
]->Type
)];
5026 do_mat_op_to_vec(ir
);
5027 lower_instructions(ir
, (MOD_TO_FRACT
| DIV_TO_MUL_RCP
| EXP_TO_EXP2
5028 | LOG_TO_LOG2
| INT_DIV_TO_MUL_RCP
5029 | ((options
->EmitNoPow
) ? POW_TO_EXP2
: 0)));
5031 progress
= do_lower_jumps(ir
, true, true, options
->EmitNoMainReturn
, options
->EmitNoCont
, options
->EmitNoLoops
) || progress
;
5033 progress
= do_common_optimization(ir
, true, options
->MaxUnrollIterations
) || progress
;
5035 progress
= lower_quadop_vector(ir
, false) || progress
;
5037 if (options
->MaxIfDepth
== 0)
5038 progress
= lower_discard(ir
) || progress
;
5040 progress
= lower_if_to_cond_assign(ir
, options
->MaxIfDepth
) || progress
;
5042 if (options
->EmitNoNoise
)
5043 progress
= lower_noise(ir
) || progress
;
5045 /* If there are forms of indirect addressing that the driver
5046 * cannot handle, perform the lowering pass.
5048 if (options
->EmitNoIndirectInput
|| options
->EmitNoIndirectOutput
5049 || options
->EmitNoIndirectTemp
|| options
->EmitNoIndirectUniform
)
5051 lower_variable_index_to_cond_assign(ir
,
5052 options
->EmitNoIndirectInput
,
5053 options
->EmitNoIndirectOutput
,
5054 options
->EmitNoIndirectTemp
,
5055 options
->EmitNoIndirectUniform
)
5058 progress
= do_vec_index_to_cond_assign(ir
) || progress
;
5061 validate_ir_tree(ir
);
5064 for (unsigned i
= 0; i
< MESA_SHADER_TYPES
; i
++) {
5065 struct gl_program
*linked_prog
;
5067 if (prog
->_LinkedShaders
[i
] == NULL
)
5070 linked_prog
= get_mesa_program(ctx
, prog
, prog
->_LinkedShaders
[i
]);
5075 switch (prog
->_LinkedShaders
[i
]->Type
) {
5076 case GL_VERTEX_SHADER
:
5077 _mesa_reference_vertprog(ctx
, &prog
->VertexProgram
,
5078 (struct gl_vertex_program
*)linked_prog
);
5079 ok
= ctx
->Driver
.ProgramStringNotify(ctx
, GL_VERTEX_PROGRAM_ARB
,
5082 _mesa_reference_vertprog(ctx
, &prog
->VertexProgram
, NULL
);
5085 case GL_FRAGMENT_SHADER
:
5086 _mesa_reference_fragprog(ctx
, &prog
->FragmentProgram
,
5087 (struct gl_fragment_program
*)linked_prog
);
5088 ok
= ctx
->Driver
.ProgramStringNotify(ctx
, GL_FRAGMENT_PROGRAM_ARB
,
5091 _mesa_reference_fragprog(ctx
, &prog
->FragmentProgram
, NULL
);
5094 case GL_GEOMETRY_SHADER
:
5095 _mesa_reference_geomprog(ctx
, &prog
->GeometryProgram
,
5096 (struct gl_geometry_program
*)linked_prog
);
5097 ok
= ctx
->Driver
.ProgramStringNotify(ctx
, GL_GEOMETRY_PROGRAM_NV
,
5100 _mesa_reference_geomprog(ctx
, &prog
->GeometryProgram
, NULL
);
5105 _mesa_reference_program(ctx
, &prog
->_LinkedShaders
[i
]->Program
, NULL
);
5106 _mesa_reference_program(ctx
, &linked_prog
, NULL
);
5111 _mesa_reference_program(ctx
, &linked_prog
, NULL
);