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_ANY_CONST ((1 << PROGRAM_LOCAL_PARAM) | \
74 (1 << PROGRAM_ENV_PARAM) | \
75 (1 << PROGRAM_STATE_VAR) | \
76 (1 << PROGRAM_NAMED_PARAM) | \
77 (1 << PROGRAM_CONSTANT) | \
78 (1 << PROGRAM_UNIFORM))
83 static int swizzle_for_size(int size
);
86 * This struct is a corresponding struct to TGSI ureg_src.
90 st_src_reg(gl_register_file file
, int index
, const glsl_type
*type
)
94 if (type
&& (type
->is_scalar() || type
->is_vector() || type
->is_matrix()))
95 this->swizzle
= swizzle_for_size(type
->vector_elements
);
97 this->swizzle
= SWIZZLE_XYZW
;
99 this->type
= type
? type
->base_type
: GLSL_TYPE_ERROR
;
100 this->reladdr
= NULL
;
103 st_src_reg(gl_register_file file
, int index
, int type
)
108 this->swizzle
= SWIZZLE_XYZW
;
110 this->reladdr
= NULL
;
115 this->type
= GLSL_TYPE_ERROR
;
116 this->file
= PROGRAM_UNDEFINED
;
120 this->reladdr
= NULL
;
123 explicit st_src_reg(st_dst_reg reg
);
125 gl_register_file file
; /**< PROGRAM_* from Mesa */
126 int index
; /**< temporary index, VERT_ATTRIB_*, FRAG_ATTRIB_*, etc. */
127 GLuint swizzle
; /**< SWIZZLE_XYZWONEZERO swizzles from Mesa. */
128 int negate
; /**< NEGATE_XYZW mask from mesa */
129 int type
; /** GLSL_TYPE_* from GLSL IR (enum glsl_base_type) */
130 /** Register index should be offset by the integer in this reg. */
136 st_dst_reg(gl_register_file file
, int writemask
, int type
)
140 this->writemask
= writemask
;
141 this->cond_mask
= COND_TR
;
142 this->reladdr
= NULL
;
148 this->type
= GLSL_TYPE_ERROR
;
149 this->file
= PROGRAM_UNDEFINED
;
152 this->cond_mask
= COND_TR
;
153 this->reladdr
= NULL
;
156 explicit st_dst_reg(st_src_reg reg
);
158 gl_register_file file
; /**< PROGRAM_* from Mesa */
159 int index
; /**< temporary index, VERT_ATTRIB_*, FRAG_ATTRIB_*, etc. */
160 int writemask
; /**< Bitfield of WRITEMASK_[XYZW] */
162 int type
; /** GLSL_TYPE_* from GLSL IR (enum glsl_base_type) */
163 /** Register index should be offset by the integer in this reg. */
167 st_src_reg::st_src_reg(st_dst_reg reg
)
169 this->type
= reg
.type
;
170 this->file
= reg
.file
;
171 this->index
= reg
.index
;
172 this->swizzle
= SWIZZLE_XYZW
;
174 this->reladdr
= NULL
;
177 st_dst_reg::st_dst_reg(st_src_reg reg
)
179 this->type
= reg
.type
;
180 this->file
= reg
.file
;
181 this->index
= reg
.index
;
182 this->writemask
= WRITEMASK_XYZW
;
183 this->cond_mask
= COND_TR
;
184 this->reladdr
= reg
.reladdr
;
187 class glsl_to_tgsi_instruction
: public exec_node
{
189 /* Callers of this ralloc-based new need not call delete. It's
190 * easier to just ralloc_free 'ctx' (or any of its ancestors). */
191 static void* operator new(size_t size
, void *ctx
)
195 node
= rzalloc_size(ctx
, size
);
196 assert(node
!= NULL
);
204 /** Pointer to the ir source this tree came from for debugging */
206 GLboolean cond_update
;
208 int sampler
; /**< sampler index */
209 int tex_target
; /**< One of TEXTURE_*_INDEX */
210 GLboolean tex_shadow
;
212 class function_entry
*function
; /* Set on TGSI_OPCODE_CAL or TGSI_OPCODE_BGNSUB */
215 class variable_storage
: public exec_node
{
217 variable_storage(ir_variable
*var
, gl_register_file file
, int index
)
218 : file(file
), index(index
), var(var
)
223 gl_register_file file
;
225 ir_variable
*var
; /* variable that maps to this, if any */
228 class function_entry
: public exec_node
{
230 ir_function_signature
*sig
;
233 * identifier of this function signature used by the program.
235 * At the point that Mesa instructions for function calls are
236 * generated, we don't know the address of the first instruction of
237 * the function body. So we make the BranchTarget that is called a
238 * small integer and rewrite them during set_branchtargets().
243 * Pointer to first instruction of the function body.
245 * Set during function body emits after main() is processed.
247 glsl_to_tgsi_instruction
*bgn_inst
;
250 * Index of the first instruction of the function body in actual
253 * Set after convertion from glsl_to_tgsi_instruction to prog_instruction.
257 /** Storage for the return value. */
258 st_src_reg return_reg
;
261 class glsl_to_tgsi_visitor
: public ir_visitor
{
263 glsl_to_tgsi_visitor();
264 ~glsl_to_tgsi_visitor();
266 function_entry
*current_function
;
268 struct gl_context
*ctx
;
269 struct gl_program
*prog
;
270 struct gl_shader_program
*shader_program
;
271 struct gl_shader_compiler_options
*options
;
275 int num_address_regs
;
277 bool indirect_addr_temps
;
278 bool indirect_addr_consts
;
282 variable_storage
*find_variable_storage(ir_variable
*var
);
284 function_entry
*get_function_signature(ir_function_signature
*sig
);
286 st_src_reg
get_temp(const glsl_type
*type
);
287 void reladdr_to_temp(ir_instruction
*ir
, st_src_reg
*reg
, int *num_reladdr
);
289 st_src_reg
st_src_reg_for_float(float val
);
290 st_src_reg
st_src_reg_for_int(int val
);
291 st_src_reg
st_src_reg_for_type(int type
, int val
);
294 * \name Visit methods
296 * As typical for the visitor pattern, there must be one \c visit method for
297 * each concrete subclass of \c ir_instruction. Virtual base classes within
298 * the hierarchy should not have \c visit methods.
301 virtual void visit(ir_variable
*);
302 virtual void visit(ir_loop
*);
303 virtual void visit(ir_loop_jump
*);
304 virtual void visit(ir_function_signature
*);
305 virtual void visit(ir_function
*);
306 virtual void visit(ir_expression
*);
307 virtual void visit(ir_swizzle
*);
308 virtual void visit(ir_dereference_variable
*);
309 virtual void visit(ir_dereference_array
*);
310 virtual void visit(ir_dereference_record
*);
311 virtual void visit(ir_assignment
*);
312 virtual void visit(ir_constant
*);
313 virtual void visit(ir_call
*);
314 virtual void visit(ir_return
*);
315 virtual void visit(ir_discard
*);
316 virtual void visit(ir_texture
*);
317 virtual void visit(ir_if
*);
322 /** List of variable_storage */
325 /** List of function_entry */
326 exec_list function_signatures
;
327 int next_signature_id
;
329 /** List of glsl_to_tgsi_instruction */
330 exec_list instructions
;
332 glsl_to_tgsi_instruction
*emit(ir_instruction
*ir
, unsigned op
);
334 glsl_to_tgsi_instruction
*emit(ir_instruction
*ir
, unsigned op
,
335 st_dst_reg dst
, st_src_reg src0
);
337 glsl_to_tgsi_instruction
*emit(ir_instruction
*ir
, unsigned op
,
338 st_dst_reg dst
, st_src_reg src0
, st_src_reg src1
);
340 glsl_to_tgsi_instruction
*emit(ir_instruction
*ir
, unsigned op
,
342 st_src_reg src0
, st_src_reg src1
, st_src_reg src2
);
344 unsigned get_opcode(ir_instruction
*ir
, unsigned op
,
346 st_src_reg src0
, st_src_reg src1
);
349 * Emit the correct dot-product instruction for the type of arguments
351 void emit_dp(ir_instruction
*ir
,
357 void emit_scalar(ir_instruction
*ir
, unsigned op
,
358 st_dst_reg dst
, st_src_reg src0
);
360 void emit_scalar(ir_instruction
*ir
, unsigned op
,
361 st_dst_reg dst
, st_src_reg src0
, st_src_reg src1
);
363 void emit_arl(ir_instruction
*ir
, st_dst_reg dst
, st_src_reg src0
);
365 void emit_scs(ir_instruction
*ir
, unsigned op
,
366 st_dst_reg dst
, const st_src_reg
&src
);
368 GLboolean
try_emit_mad(ir_expression
*ir
,
370 GLboolean
try_emit_sat(ir_expression
*ir
);
372 void emit_swz(ir_expression
*ir
);
374 bool process_move_condition(ir_rvalue
*ir
);
376 void remove_output_reads(gl_register_file type
);
377 void simplify_cmp(void);
379 void rename_temp_register(int index
, int new_index
);
380 int get_first_temp_read(int index
);
381 int get_first_temp_write(int index
);
382 int get_last_temp_read(int index
);
383 int get_last_temp_write(int index
);
385 void copy_propagate(void);
386 void eliminate_dead_code(void);
387 void merge_registers(void);
388 void renumber_registers(void);
393 static st_src_reg undef_src
= st_src_reg(PROGRAM_UNDEFINED
, 0, GLSL_TYPE_ERROR
);
395 static st_dst_reg undef_dst
= st_dst_reg(PROGRAM_UNDEFINED
, SWIZZLE_NOOP
, GLSL_TYPE_ERROR
);
397 static st_dst_reg address_reg
= st_dst_reg(PROGRAM_ADDRESS
, WRITEMASK_X
, GLSL_TYPE_FLOAT
);
400 fail_link(struct gl_shader_program
*prog
, const char *fmt
, ...) PRINTFLIKE(2, 3);
403 fail_link(struct gl_shader_program
*prog
, const char *fmt
, ...)
407 ralloc_vasprintf_append(&prog
->InfoLog
, fmt
, args
);
410 prog
->LinkStatus
= GL_FALSE
;
414 swizzle_for_size(int size
)
416 int size_swizzles
[4] = {
417 MAKE_SWIZZLE4(SWIZZLE_X
, SWIZZLE_X
, SWIZZLE_X
, SWIZZLE_X
),
418 MAKE_SWIZZLE4(SWIZZLE_X
, SWIZZLE_Y
, SWIZZLE_Y
, SWIZZLE_Y
),
419 MAKE_SWIZZLE4(SWIZZLE_X
, SWIZZLE_Y
, SWIZZLE_Z
, SWIZZLE_Z
),
420 MAKE_SWIZZLE4(SWIZZLE_X
, SWIZZLE_Y
, SWIZZLE_Z
, SWIZZLE_W
),
423 assert((size
>= 1) && (size
<= 4));
424 return size_swizzles
[size
- 1];
428 is_tex_instruction(unsigned opcode
)
430 const tgsi_opcode_info
* info
= tgsi_get_opcode_info(opcode
);
435 num_inst_dst_regs(unsigned opcode
)
437 const tgsi_opcode_info
* info
= tgsi_get_opcode_info(opcode
);
438 return info
->num_dst
;
442 num_inst_src_regs(unsigned opcode
)
444 const tgsi_opcode_info
* info
= tgsi_get_opcode_info(opcode
);
445 return info
->is_tex
? info
->num_src
- 1 : info
->num_src
;
448 glsl_to_tgsi_instruction
*
449 glsl_to_tgsi_visitor::emit(ir_instruction
*ir
, unsigned op
,
451 st_src_reg src0
, st_src_reg src1
, st_src_reg src2
)
453 glsl_to_tgsi_instruction
*inst
= new(mem_ctx
) glsl_to_tgsi_instruction();
454 int num_reladdr
= 0, i
;
456 op
= get_opcode(ir
, op
, dst
, src0
, src1
);
458 /* If we have to do relative addressing, we want to load the ARL
459 * reg directly for one of the regs, and preload the other reladdr
460 * sources into temps.
462 num_reladdr
+= dst
.reladdr
!= NULL
;
463 num_reladdr
+= src0
.reladdr
!= NULL
;
464 num_reladdr
+= src1
.reladdr
!= NULL
;
465 num_reladdr
+= src2
.reladdr
!= NULL
;
467 reladdr_to_temp(ir
, &src2
, &num_reladdr
);
468 reladdr_to_temp(ir
, &src1
, &num_reladdr
);
469 reladdr_to_temp(ir
, &src0
, &num_reladdr
);
472 emit_arl(ir
, address_reg
, *dst
.reladdr
);
475 assert(num_reladdr
== 0);
484 inst
->function
= NULL
;
486 if (op
== TGSI_OPCODE_ARL
)
487 this->num_address_regs
= 1;
489 /* Update indirect addressing status used by TGSI */
492 case PROGRAM_TEMPORARY
:
493 this->indirect_addr_temps
= true;
495 case PROGRAM_LOCAL_PARAM
:
496 case PROGRAM_ENV_PARAM
:
497 case PROGRAM_STATE_VAR
:
498 case PROGRAM_NAMED_PARAM
:
499 case PROGRAM_CONSTANT
:
500 case PROGRAM_UNIFORM
:
501 this->indirect_addr_consts
= true;
508 for (i
=0; i
<3; i
++) {
509 if(inst
->src
[i
].reladdr
) {
510 switch(inst
->src
[i
].file
) {
511 case PROGRAM_TEMPORARY
:
512 this->indirect_addr_temps
= true;
514 case PROGRAM_LOCAL_PARAM
:
515 case PROGRAM_ENV_PARAM
:
516 case PROGRAM_STATE_VAR
:
517 case PROGRAM_NAMED_PARAM
:
518 case PROGRAM_CONSTANT
:
519 case PROGRAM_UNIFORM
:
520 this->indirect_addr_consts
= true;
529 this->instructions
.push_tail(inst
);
535 glsl_to_tgsi_instruction
*
536 glsl_to_tgsi_visitor::emit(ir_instruction
*ir
, unsigned op
,
537 st_dst_reg dst
, st_src_reg src0
, st_src_reg src1
)
539 return emit(ir
, op
, dst
, src0
, src1
, undef_src
);
542 glsl_to_tgsi_instruction
*
543 glsl_to_tgsi_visitor::emit(ir_instruction
*ir
, unsigned op
,
544 st_dst_reg dst
, st_src_reg src0
)
546 assert(dst
.writemask
!= 0);
547 return emit(ir
, op
, dst
, src0
, undef_src
, undef_src
);
550 glsl_to_tgsi_instruction
*
551 glsl_to_tgsi_visitor::emit(ir_instruction
*ir
, unsigned op
)
553 return emit(ir
, op
, undef_dst
, undef_src
, undef_src
, undef_src
);
557 * Determines whether to use an integer, unsigned integer, or float opcode
558 * based on the operands and input opcode, then emits the result.
560 * TODO: type checking for remaining TGSI opcodes
563 glsl_to_tgsi_visitor::get_opcode(ir_instruction
*ir
, unsigned op
,
565 st_src_reg src0
, st_src_reg src1
)
567 int type
= GLSL_TYPE_FLOAT
;
569 if (src0
.type
== GLSL_TYPE_FLOAT
|| src1
.type
== GLSL_TYPE_FLOAT
)
570 type
= GLSL_TYPE_FLOAT
;
571 else if (glsl_version
>= 130)
574 #define case4(c, f, i, u) \
575 case TGSI_OPCODE_##c: \
576 if (type == GLSL_TYPE_INT) op = TGSI_OPCODE_##i; \
577 else if (type == GLSL_TYPE_UINT) op = TGSI_OPCODE_##u; \
578 else op = TGSI_OPCODE_##f; \
580 #define case3(f, i, u) case4(f, f, i, u)
581 #define case2fi(f, i) case4(f, f, i, i)
582 #define case2iu(i, u) case4(i, LAST, i, u)
588 case3(DIV
, IDIV
, UDIV
);
589 case3(MAX
, IMAX
, UMAX
);
590 case3(MIN
, IMIN
, UMIN
);
595 case3(SGE
, ISGE
, USGE
);
596 case3(SLT
, ISLT
, USLT
);
608 assert(op
!= TGSI_OPCODE_LAST
);
613 glsl_to_tgsi_visitor::emit_dp(ir_instruction
*ir
,
614 st_dst_reg dst
, st_src_reg src0
, st_src_reg src1
,
617 static const unsigned dot_opcodes
[] = {
618 TGSI_OPCODE_DP2
, TGSI_OPCODE_DP3
, TGSI_OPCODE_DP4
621 emit(ir
, dot_opcodes
[elements
- 2], dst
, src0
, src1
);
625 * Emits TGSI scalar opcodes to produce unique answers across channels.
627 * Some TGSI opcodes are scalar-only, like ARB_fp/vp. The src X
628 * channel determines the result across all channels. So to do a vec4
629 * of this operation, we want to emit a scalar per source channel used
630 * to produce dest channels.
633 glsl_to_tgsi_visitor::emit_scalar(ir_instruction
*ir
, unsigned op
,
635 st_src_reg orig_src0
, st_src_reg orig_src1
)
638 int done_mask
= ~dst
.writemask
;
640 /* TGSI RCP is a scalar operation splatting results to all channels,
641 * like ARB_fp/vp. So emit as many RCPs as necessary to cover our
644 for (i
= 0; i
< 4; i
++) {
645 GLuint this_mask
= (1 << i
);
646 glsl_to_tgsi_instruction
*inst
;
647 st_src_reg src0
= orig_src0
;
648 st_src_reg src1
= orig_src1
;
650 if (done_mask
& this_mask
)
653 GLuint src0_swiz
= GET_SWZ(src0
.swizzle
, i
);
654 GLuint src1_swiz
= GET_SWZ(src1
.swizzle
, i
);
655 for (j
= i
+ 1; j
< 4; j
++) {
656 /* If there is another enabled component in the destination that is
657 * derived from the same inputs, generate its value on this pass as
660 if (!(done_mask
& (1 << j
)) &&
661 GET_SWZ(src0
.swizzle
, j
) == src0_swiz
&&
662 GET_SWZ(src1
.swizzle
, j
) == src1_swiz
) {
663 this_mask
|= (1 << j
);
666 src0
.swizzle
= MAKE_SWIZZLE4(src0_swiz
, src0_swiz
,
667 src0_swiz
, src0_swiz
);
668 src1
.swizzle
= MAKE_SWIZZLE4(src1_swiz
, src1_swiz
,
669 src1_swiz
, src1_swiz
);
671 inst
= emit(ir
, op
, dst
, src0
, src1
);
672 inst
->dst
.writemask
= this_mask
;
673 done_mask
|= this_mask
;
678 glsl_to_tgsi_visitor::emit_scalar(ir_instruction
*ir
, unsigned op
,
679 st_dst_reg dst
, st_src_reg src0
)
681 st_src_reg undef
= undef_src
;
683 undef
.swizzle
= SWIZZLE_XXXX
;
685 emit_scalar(ir
, op
, dst
, src0
, undef
);
689 glsl_to_tgsi_visitor::emit_arl(ir_instruction
*ir
,
690 st_dst_reg dst
, st_src_reg src0
)
692 st_src_reg tmp
= get_temp(glsl_type::float_type
);
694 if (src0
.type
== GLSL_TYPE_INT
)
695 emit(ir
, TGSI_OPCODE_I2F
, st_dst_reg(tmp
), src0
);
696 else if (src0
.type
== GLSL_TYPE_UINT
)
697 emit(ir
, TGSI_OPCODE_U2F
, st_dst_reg(tmp
), src0
);
701 emit(ir
, TGSI_OPCODE_ARL
, dst
, tmp
);
705 * Emit an TGSI_OPCODE_SCS instruction
707 * The \c SCS opcode functions a bit differently than the other TGSI opcodes.
708 * Instead of splatting its result across all four components of the
709 * destination, it writes one value to the \c x component and another value to
710 * the \c y component.
712 * \param ir IR instruction being processed
713 * \param op Either \c TGSI_OPCODE_SIN or \c TGSI_OPCODE_COS depending
714 * on which value is desired.
715 * \param dst Destination register
716 * \param src Source register
719 glsl_to_tgsi_visitor::emit_scs(ir_instruction
*ir
, unsigned op
,
721 const st_src_reg
&src
)
723 /* Vertex programs cannot use the SCS opcode.
725 if (this->prog
->Target
== GL_VERTEX_PROGRAM_ARB
) {
726 emit_scalar(ir
, op
, dst
, src
);
730 const unsigned component
= (op
== TGSI_OPCODE_SIN
) ? 0 : 1;
731 const unsigned scs_mask
= (1U << component
);
732 int done_mask
= ~dst
.writemask
;
735 assert(op
== TGSI_OPCODE_SIN
|| op
== TGSI_OPCODE_COS
);
737 /* If there are compnents in the destination that differ from the component
738 * that will be written by the SCS instrution, we'll need a temporary.
740 if (scs_mask
!= unsigned(dst
.writemask
)) {
741 tmp
= get_temp(glsl_type::vec4_type
);
744 for (unsigned i
= 0; i
< 4; i
++) {
745 unsigned this_mask
= (1U << i
);
746 st_src_reg src0
= src
;
748 if ((done_mask
& this_mask
) != 0)
751 /* The source swizzle specified which component of the source generates
752 * sine / cosine for the current component in the destination. The SCS
753 * instruction requires that this value be swizzle to the X component.
754 * Replace the current swizzle with a swizzle that puts the source in
757 unsigned src0_swiz
= GET_SWZ(src
.swizzle
, i
);
759 src0
.swizzle
= MAKE_SWIZZLE4(src0_swiz
, src0_swiz
,
760 src0_swiz
, src0_swiz
);
761 for (unsigned j
= i
+ 1; j
< 4; j
++) {
762 /* If there is another enabled component in the destination that is
763 * derived from the same inputs, generate its value on this pass as
766 if (!(done_mask
& (1 << j
)) &&
767 GET_SWZ(src0
.swizzle
, j
) == src0_swiz
) {
768 this_mask
|= (1 << j
);
772 if (this_mask
!= scs_mask
) {
773 glsl_to_tgsi_instruction
*inst
;
774 st_dst_reg tmp_dst
= st_dst_reg(tmp
);
776 /* Emit the SCS instruction.
778 inst
= emit(ir
, TGSI_OPCODE_SCS
, tmp_dst
, src0
);
779 inst
->dst
.writemask
= scs_mask
;
781 /* Move the result of the SCS instruction to the desired location in
784 tmp
.swizzle
= MAKE_SWIZZLE4(component
, component
,
785 component
, component
);
786 inst
= emit(ir
, TGSI_OPCODE_SCS
, dst
, tmp
);
787 inst
->dst
.writemask
= this_mask
;
789 /* Emit the SCS instruction to write directly to the destination.
791 glsl_to_tgsi_instruction
*inst
= emit(ir
, TGSI_OPCODE_SCS
, dst
, src0
);
792 inst
->dst
.writemask
= scs_mask
;
795 done_mask
|= this_mask
;
800 glsl_to_tgsi_visitor::st_src_reg_for_float(float val
)
802 st_src_reg
src(PROGRAM_CONSTANT
, -1, GLSL_TYPE_FLOAT
);
803 union gl_constant_value uval
;
806 src
.index
= _mesa_add_typed_unnamed_constant(this->prog
->Parameters
,
807 &uval
, 1, GL_FLOAT
, &src
.swizzle
);
813 glsl_to_tgsi_visitor::st_src_reg_for_int(int val
)
815 st_src_reg
src(PROGRAM_CONSTANT
, -1, GLSL_TYPE_INT
);
816 union gl_constant_value uval
;
818 assert(glsl_version
>= 130);
821 src
.index
= _mesa_add_typed_unnamed_constant(this->prog
->Parameters
,
822 &uval
, 1, GL_INT
, &src
.swizzle
);
828 glsl_to_tgsi_visitor::st_src_reg_for_type(int type
, int val
)
830 if (glsl_version
>= 130)
831 return type
== GLSL_TYPE_FLOAT
? st_src_reg_for_float(val
) :
832 st_src_reg_for_int(val
);
834 return st_src_reg_for_float(val
);
838 type_size(const struct glsl_type
*type
)
843 switch (type
->base_type
) {
846 case GLSL_TYPE_FLOAT
:
848 if (type
->is_matrix()) {
849 return type
->matrix_columns
;
851 /* Regardless of size of vector, it gets a vec4. This is bad
852 * packing for things like floats, but otherwise arrays become a
853 * mess. Hopefully a later pass over the code can pack scalars
854 * down if appropriate.
858 case GLSL_TYPE_ARRAY
:
859 assert(type
->length
> 0);
860 return type_size(type
->fields
.array
) * type
->length
;
861 case GLSL_TYPE_STRUCT
:
863 for (i
= 0; i
< type
->length
; i
++) {
864 size
+= type_size(type
->fields
.structure
[i
].type
);
867 case GLSL_TYPE_SAMPLER
:
868 /* Samplers take up one slot in UNIFORMS[], but they're baked in
879 * In the initial pass of codegen, we assign temporary numbers to
880 * intermediate results. (not SSA -- variable assignments will reuse
884 glsl_to_tgsi_visitor::get_temp(const glsl_type
*type
)
890 src
.type
= glsl_version
>= 130 ? type
->base_type
: GLSL_TYPE_FLOAT
;
891 src
.file
= PROGRAM_TEMPORARY
;
892 src
.index
= next_temp
;
894 next_temp
+= type_size(type
);
896 if (type
->is_array() || type
->is_record()) {
897 src
.swizzle
= SWIZZLE_NOOP
;
899 for (i
= 0; i
< type
->vector_elements
; i
++)
902 swizzle
[i
] = type
->vector_elements
- 1;
903 src
.swizzle
= MAKE_SWIZZLE4(swizzle
[0], swizzle
[1],
904 swizzle
[2], swizzle
[3]);
912 glsl_to_tgsi_visitor::find_variable_storage(ir_variable
*var
)
915 variable_storage
*entry
;
917 foreach_iter(exec_list_iterator
, iter
, this->variables
) {
918 entry
= (variable_storage
*)iter
.get();
920 if (entry
->var
== var
)
928 glsl_to_tgsi_visitor::visit(ir_variable
*ir
)
930 if (strcmp(ir
->name
, "gl_FragCoord") == 0) {
931 struct gl_fragment_program
*fp
= (struct gl_fragment_program
*)this->prog
;
933 fp
->OriginUpperLeft
= ir
->origin_upper_left
;
934 fp
->PixelCenterInteger
= ir
->pixel_center_integer
;
936 } else if (strcmp(ir
->name
, "gl_FragDepth") == 0) {
937 struct gl_fragment_program
*fp
= (struct gl_fragment_program
*)this->prog
;
938 switch (ir
->depth_layout
) {
939 case ir_depth_layout_none
:
940 fp
->FragDepthLayout
= FRAG_DEPTH_LAYOUT_NONE
;
942 case ir_depth_layout_any
:
943 fp
->FragDepthLayout
= FRAG_DEPTH_LAYOUT_ANY
;
945 case ir_depth_layout_greater
:
946 fp
->FragDepthLayout
= FRAG_DEPTH_LAYOUT_GREATER
;
948 case ir_depth_layout_less
:
949 fp
->FragDepthLayout
= FRAG_DEPTH_LAYOUT_LESS
;
951 case ir_depth_layout_unchanged
:
952 fp
->FragDepthLayout
= FRAG_DEPTH_LAYOUT_UNCHANGED
;
960 if (ir
->mode
== ir_var_uniform
&& strncmp(ir
->name
, "gl_", 3) == 0) {
962 const ir_state_slot
*const slots
= ir
->state_slots
;
963 assert(ir
->state_slots
!= NULL
);
965 /* Check if this statevar's setup in the STATE file exactly
966 * matches how we'll want to reference it as a
967 * struct/array/whatever. If not, then we need to move it into
968 * temporary storage and hope that it'll get copy-propagated
971 for (i
= 0; i
< ir
->num_state_slots
; i
++) {
972 if (slots
[i
].swizzle
!= SWIZZLE_XYZW
) {
977 struct variable_storage
*storage
;
979 if (i
== ir
->num_state_slots
) {
980 /* We'll set the index later. */
981 storage
= new(mem_ctx
) variable_storage(ir
, PROGRAM_STATE_VAR
, -1);
982 this->variables
.push_tail(storage
);
986 /* The variable_storage constructor allocates slots based on the size
987 * of the type. However, this had better match the number of state
988 * elements that we're going to copy into the new temporary.
990 assert((int) ir
->num_state_slots
== type_size(ir
->type
));
992 storage
= new(mem_ctx
) variable_storage(ir
, PROGRAM_TEMPORARY
,
994 this->variables
.push_tail(storage
);
995 this->next_temp
+= type_size(ir
->type
);
997 dst
= st_dst_reg(st_src_reg(PROGRAM_TEMPORARY
, storage
->index
,
998 glsl_version
>= 130 ? ir
->type
->base_type
: GLSL_TYPE_FLOAT
));
1002 for (unsigned int i
= 0; i
< ir
->num_state_slots
; i
++) {
1003 int index
= _mesa_add_state_reference(this->prog
->Parameters
,
1004 (gl_state_index
*)slots
[i
].tokens
);
1006 if (storage
->file
== PROGRAM_STATE_VAR
) {
1007 if (storage
->index
== -1) {
1008 storage
->index
= index
;
1010 assert(index
== storage
->index
+ (int)i
);
1013 st_src_reg
src(PROGRAM_STATE_VAR
, index
,
1014 glsl_version
>= 130 ? ir
->type
->base_type
: GLSL_TYPE_FLOAT
);
1015 src
.swizzle
= slots
[i
].swizzle
;
1016 emit(ir
, TGSI_OPCODE_MOV
, dst
, src
);
1017 /* even a float takes up a whole vec4 reg in a struct/array. */
1022 if (storage
->file
== PROGRAM_TEMPORARY
&&
1023 dst
.index
!= storage
->index
+ (int) ir
->num_state_slots
) {
1024 fail_link(this->shader_program
,
1025 "failed to load builtin uniform `%s' (%d/%d regs loaded)\n",
1026 ir
->name
, dst
.index
- storage
->index
,
1027 type_size(ir
->type
));
1033 glsl_to_tgsi_visitor::visit(ir_loop
*ir
)
1035 ir_dereference_variable
*counter
= NULL
;
1037 if (ir
->counter
!= NULL
)
1038 counter
= new(ir
) ir_dereference_variable(ir
->counter
);
1040 if (ir
->from
!= NULL
) {
1041 assert(ir
->counter
!= NULL
);
1043 ir_assignment
*a
= new(ir
) ir_assignment(counter
, ir
->from
, NULL
);
1049 emit(NULL
, TGSI_OPCODE_BGNLOOP
);
1053 new(ir
) ir_expression(ir
->cmp
, glsl_type::bool_type
,
1055 ir_if
*if_stmt
= new(ir
) ir_if(e
);
1057 ir_loop_jump
*brk
= new(ir
) ir_loop_jump(ir_loop_jump::jump_break
);
1059 if_stmt
->then_instructions
.push_tail(brk
);
1061 if_stmt
->accept(this);
1068 visit_exec_list(&ir
->body_instructions
, this);
1070 if (ir
->increment
) {
1072 new(ir
) ir_expression(ir_binop_add
, counter
->type
,
1073 counter
, ir
->increment
);
1075 ir_assignment
*a
= new(ir
) ir_assignment(counter
, e
, NULL
);
1082 emit(NULL
, TGSI_OPCODE_ENDLOOP
);
1086 glsl_to_tgsi_visitor::visit(ir_loop_jump
*ir
)
1089 case ir_loop_jump::jump_break
:
1090 emit(NULL
, TGSI_OPCODE_BRK
);
1092 case ir_loop_jump::jump_continue
:
1093 emit(NULL
, TGSI_OPCODE_CONT
);
1100 glsl_to_tgsi_visitor::visit(ir_function_signature
*ir
)
1107 glsl_to_tgsi_visitor::visit(ir_function
*ir
)
1109 /* Ignore function bodies other than main() -- we shouldn't see calls to
1110 * them since they should all be inlined before we get to glsl_to_tgsi.
1112 if (strcmp(ir
->name
, "main") == 0) {
1113 const ir_function_signature
*sig
;
1116 sig
= ir
->matching_signature(&empty
);
1120 foreach_iter(exec_list_iterator
, iter
, sig
->body
) {
1121 ir_instruction
*ir
= (ir_instruction
*)iter
.get();
1129 glsl_to_tgsi_visitor::try_emit_mad(ir_expression
*ir
, int mul_operand
)
1131 int nonmul_operand
= 1 - mul_operand
;
1134 ir_expression
*expr
= ir
->operands
[mul_operand
]->as_expression();
1135 if (!expr
|| expr
->operation
!= ir_binop_mul
)
1138 expr
->operands
[0]->accept(this);
1140 expr
->operands
[1]->accept(this);
1142 ir
->operands
[nonmul_operand
]->accept(this);
1145 this->result
= get_temp(ir
->type
);
1146 emit(ir
, TGSI_OPCODE_MAD
, st_dst_reg(this->result
), a
, b
, c
);
1152 glsl_to_tgsi_visitor::try_emit_sat(ir_expression
*ir
)
1154 /* Saturates were only introduced to vertex programs in
1155 * NV_vertex_program3, so don't give them to drivers in the VP.
1157 if (this->prog
->Target
== GL_VERTEX_PROGRAM_ARB
)
1160 ir_rvalue
*sat_src
= ir
->as_rvalue_to_saturate();
1164 sat_src
->accept(this);
1165 st_src_reg src
= this->result
;
1167 this->result
= get_temp(ir
->type
);
1168 glsl_to_tgsi_instruction
*inst
;
1169 inst
= emit(ir
, TGSI_OPCODE_MOV
, st_dst_reg(this->result
), src
);
1170 inst
->saturate
= true;
1176 glsl_to_tgsi_visitor::reladdr_to_temp(ir_instruction
*ir
,
1177 st_src_reg
*reg
, int *num_reladdr
)
1182 emit_arl(ir
, address_reg
, *reg
->reladdr
);
1184 if (*num_reladdr
!= 1) {
1185 st_src_reg temp
= get_temp(glsl_type::vec4_type
);
1187 emit(ir
, TGSI_OPCODE_MOV
, st_dst_reg(temp
), *reg
);
1195 glsl_to_tgsi_visitor::visit(ir_expression
*ir
)
1197 unsigned int operand
;
1198 st_src_reg op
[Elements(ir
->operands
)];
1199 st_src_reg result_src
;
1200 st_dst_reg result_dst
;
1202 /* Quick peephole: Emit MAD(a, b, c) instead of ADD(MUL(a, b), c)
1204 if (ir
->operation
== ir_binop_add
) {
1205 if (try_emit_mad(ir
, 1))
1207 if (try_emit_mad(ir
, 0))
1210 if (try_emit_sat(ir
))
1213 if (ir
->operation
== ir_quadop_vector
)
1214 assert(!"ir_quadop_vector should have been lowered");
1216 for (operand
= 0; operand
< ir
->get_num_operands(); operand
++) {
1217 this->result
.file
= PROGRAM_UNDEFINED
;
1218 ir
->operands
[operand
]->accept(this);
1219 if (this->result
.file
== PROGRAM_UNDEFINED
) {
1221 printf("Failed to get tree for expression operand:\n");
1222 ir
->operands
[operand
]->accept(&v
);
1225 op
[operand
] = this->result
;
1227 /* Matrix expression operands should have been broken down to vector
1228 * operations already.
1230 assert(!ir
->operands
[operand
]->type
->is_matrix());
1233 int vector_elements
= ir
->operands
[0]->type
->vector_elements
;
1234 if (ir
->operands
[1]) {
1235 vector_elements
= MAX2(vector_elements
,
1236 ir
->operands
[1]->type
->vector_elements
);
1239 this->result
.file
= PROGRAM_UNDEFINED
;
1241 /* Storage for our result. Ideally for an assignment we'd be using
1242 * the actual storage for the result here, instead.
1244 result_src
= get_temp(ir
->type
);
1245 /* convenience for the emit functions below. */
1246 result_dst
= st_dst_reg(result_src
);
1247 /* Limit writes to the channels that will be used by result_src later.
1248 * This does limit this temp's use as a temporary for multi-instruction
1251 result_dst
.writemask
= (1 << ir
->type
->vector_elements
) - 1;
1253 switch (ir
->operation
) {
1254 case ir_unop_logic_not
:
1255 emit(ir
, TGSI_OPCODE_SEQ
, result_dst
, op
[0], st_src_reg_for_type(result_dst
.type
, 0));
1258 assert(result_dst
.type
== GLSL_TYPE_FLOAT
|| result_dst
.type
== GLSL_TYPE_INT
);
1259 if (result_dst
.type
== GLSL_TYPE_INT
)
1260 emit(ir
, TGSI_OPCODE_INEG
, result_dst
, op
[0]);
1262 op
[0].negate
= ~op
[0].negate
;
1267 assert(result_dst
.type
== GLSL_TYPE_FLOAT
);
1268 emit(ir
, TGSI_OPCODE_ABS
, result_dst
, op
[0]);
1271 emit(ir
, TGSI_OPCODE_SSG
, result_dst
, op
[0]);
1274 emit_scalar(ir
, TGSI_OPCODE_RCP
, result_dst
, op
[0]);
1278 emit_scalar(ir
, TGSI_OPCODE_EX2
, result_dst
, op
[0]);
1282 assert(!"not reached: should be handled by ir_explog_to_explog2");
1285 emit_scalar(ir
, TGSI_OPCODE_LG2
, result_dst
, op
[0]);
1288 emit_scalar(ir
, TGSI_OPCODE_SIN
, result_dst
, op
[0]);
1291 emit_scalar(ir
, TGSI_OPCODE_COS
, result_dst
, op
[0]);
1293 case ir_unop_sin_reduced
:
1294 emit_scs(ir
, TGSI_OPCODE_SIN
, result_dst
, op
[0]);
1296 case ir_unop_cos_reduced
:
1297 emit_scs(ir
, TGSI_OPCODE_COS
, result_dst
, op
[0]);
1301 emit(ir
, TGSI_OPCODE_DDX
, result_dst
, op
[0]);
1304 op
[0].negate
= ~op
[0].negate
;
1305 emit(ir
, TGSI_OPCODE_DDY
, result_dst
, op
[0]);
1308 case ir_unop_noise
: {
1309 /* At some point, a motivated person could add a better
1310 * implementation of noise. Currently not even the nvidia
1311 * binary drivers do anything more than this. In any case, the
1312 * place to do this is in the GL state tracker, not the poor
1315 emit(ir
, TGSI_OPCODE_MOV
, result_dst
, st_src_reg_for_float(0.5));
1320 emit(ir
, TGSI_OPCODE_ADD
, result_dst
, op
[0], op
[1]);
1323 emit(ir
, TGSI_OPCODE_SUB
, result_dst
, op
[0], op
[1]);
1327 emit(ir
, TGSI_OPCODE_MUL
, result_dst
, op
[0], op
[1]);
1330 if (result_dst
.type
== GLSL_TYPE_FLOAT
)
1331 assert(!"not reached: should be handled by ir_div_to_mul_rcp");
1333 emit(ir
, TGSI_OPCODE_DIV
, result_dst
, op
[0], op
[1]);
1336 if (result_dst
.type
== GLSL_TYPE_FLOAT
)
1337 assert(!"ir_binop_mod should have been converted to b * fract(a/b)");
1339 emit(ir
, TGSI_OPCODE_MOD
, result_dst
, op
[0], op
[1]);
1343 emit(ir
, TGSI_OPCODE_SLT
, result_dst
, op
[0], op
[1]);
1345 case ir_binop_greater
:
1346 emit(ir
, TGSI_OPCODE_SGT
, result_dst
, op
[0], op
[1]);
1348 case ir_binop_lequal
:
1349 emit(ir
, TGSI_OPCODE_SLE
, result_dst
, op
[0], op
[1]);
1351 case ir_binop_gequal
:
1352 emit(ir
, TGSI_OPCODE_SGE
, result_dst
, op
[0], op
[1]);
1354 case ir_binop_equal
:
1355 emit(ir
, TGSI_OPCODE_SEQ
, result_dst
, op
[0], op
[1]);
1357 case ir_binop_nequal
:
1358 emit(ir
, TGSI_OPCODE_SNE
, result_dst
, op
[0], op
[1]);
1360 case ir_binop_all_equal
:
1361 /* "==" operator producing a scalar boolean. */
1362 if (ir
->operands
[0]->type
->is_vector() ||
1363 ir
->operands
[1]->type
->is_vector()) {
1364 st_src_reg temp
= get_temp(glsl_version
>= 130 ?
1365 glsl_type::get_instance(ir
->operands
[0]->type
->base_type
, 4, 1) :
1366 glsl_type::vec4_type
);
1367 assert(ir
->operands
[0]->type
->base_type
== GLSL_TYPE_FLOAT
);
1368 emit(ir
, TGSI_OPCODE_SNE
, st_dst_reg(temp
), op
[0], op
[1]);
1369 emit_dp(ir
, result_dst
, temp
, temp
, vector_elements
);
1370 emit(ir
, TGSI_OPCODE_SEQ
, result_dst
, result_src
, st_src_reg_for_float(0.0));
1372 emit(ir
, TGSI_OPCODE_SEQ
, result_dst
, op
[0], op
[1]);
1375 case ir_binop_any_nequal
:
1376 /* "!=" operator producing a scalar boolean. */
1377 if (ir
->operands
[0]->type
->is_vector() ||
1378 ir
->operands
[1]->type
->is_vector()) {
1379 st_src_reg temp
= get_temp(glsl_version
>= 130 ?
1380 glsl_type::get_instance(ir
->operands
[0]->type
->base_type
, 4, 1) :
1381 glsl_type::vec4_type
);
1382 assert(ir
->operands
[0]->type
->base_type
== GLSL_TYPE_FLOAT
);
1383 emit(ir
, TGSI_OPCODE_SNE
, st_dst_reg(temp
), op
[0], op
[1]);
1384 emit_dp(ir
, result_dst
, temp
, temp
, vector_elements
);
1385 emit(ir
, TGSI_OPCODE_SNE
, result_dst
, result_src
, st_src_reg_for_float(0.0));
1387 emit(ir
, TGSI_OPCODE_SNE
, result_dst
, op
[0], op
[1]);
1392 assert(ir
->operands
[0]->type
->is_vector());
1393 emit_dp(ir
, result_dst
, op
[0], op
[0],
1394 ir
->operands
[0]->type
->vector_elements
);
1395 emit(ir
, TGSI_OPCODE_SNE
, result_dst
, result_src
, st_src_reg_for_float(0.0));
1398 case ir_binop_logic_xor
:
1399 emit(ir
, TGSI_OPCODE_SNE
, result_dst
, op
[0], op
[1]);
1402 case ir_binop_logic_or
:
1403 /* This could be a saturated add and skip the SNE. */
1404 emit(ir
, TGSI_OPCODE_ADD
, result_dst
, op
[0], op
[1]);
1405 emit(ir
, TGSI_OPCODE_SNE
, result_dst
, result_src
, st_src_reg_for_float(0.0));
1408 case ir_binop_logic_and
:
1409 /* the bool args are stored as float 0.0 or 1.0, so "mul" gives us "and". */
1410 emit(ir
, TGSI_OPCODE_MUL
, result_dst
, op
[0], op
[1]);
1414 assert(ir
->operands
[0]->type
->is_vector());
1415 assert(ir
->operands
[0]->type
== ir
->operands
[1]->type
);
1416 emit_dp(ir
, result_dst
, op
[0], op
[1],
1417 ir
->operands
[0]->type
->vector_elements
);
1421 /* sqrt(x) = x * rsq(x). */
1422 emit_scalar(ir
, TGSI_OPCODE_RSQ
, result_dst
, op
[0]);
1423 emit(ir
, TGSI_OPCODE_MUL
, result_dst
, result_src
, op
[0]);
1424 /* For incoming channels <= 0, set the result to 0. */
1425 op
[0].negate
= ~op
[0].negate
;
1426 emit(ir
, TGSI_OPCODE_CMP
, result_dst
,
1427 op
[0], result_src
, st_src_reg_for_float(0.0));
1430 emit_scalar(ir
, TGSI_OPCODE_RSQ
, result_dst
, op
[0]);
1434 if (glsl_version
>= 130) {
1435 emit(ir
, TGSI_OPCODE_I2F
, result_dst
, op
[0]);
1439 /* Booleans are stored as integers (or floats in GLSL 1.20 and lower). */
1443 if (glsl_version
>= 130)
1444 emit(ir
, TGSI_OPCODE_F2I
, result_dst
, op
[0]);
1446 emit(ir
, TGSI_OPCODE_TRUNC
, result_dst
, op
[0]);
1450 emit(ir
, TGSI_OPCODE_SNE
, result_dst
, op
[0],
1451 st_src_reg_for_type(result_dst
.type
, 0));
1454 emit(ir
, TGSI_OPCODE_TRUNC
, result_dst
, op
[0]);
1457 op
[0].negate
= ~op
[0].negate
;
1458 emit(ir
, TGSI_OPCODE_FLR
, result_dst
, op
[0]);
1459 result_src
.negate
= ~result_src
.negate
;
1462 emit(ir
, TGSI_OPCODE_FLR
, result_dst
, op
[0]);
1465 emit(ir
, TGSI_OPCODE_FRC
, result_dst
, op
[0]);
1469 emit(ir
, TGSI_OPCODE_MIN
, result_dst
, op
[0], op
[1]);
1472 emit(ir
, TGSI_OPCODE_MAX
, result_dst
, op
[0], op
[1]);
1475 emit_scalar(ir
, TGSI_OPCODE_POW
, result_dst
, op
[0], op
[1]);
1478 case ir_unop_bit_not
:
1479 if (glsl_version
>= 130) {
1480 emit(ir
, TGSI_OPCODE_NOT
, result_dst
, op
[0]);
1484 if (glsl_version
>= 130) {
1485 emit(ir
, TGSI_OPCODE_U2F
, result_dst
, op
[0]);
1488 case ir_binop_lshift
:
1489 if (glsl_version
>= 130) {
1490 emit(ir
, TGSI_OPCODE_SHL
, result_dst
, op
[0]);
1493 case ir_binop_rshift
:
1494 if (glsl_version
>= 130) {
1495 emit(ir
, TGSI_OPCODE_ISHR
, result_dst
, op
[0]);
1498 case ir_binop_bit_and
:
1499 if (glsl_version
>= 130) {
1500 emit(ir
, TGSI_OPCODE_AND
, result_dst
, op
[0]);
1503 case ir_binop_bit_xor
:
1504 if (glsl_version
>= 130) {
1505 emit(ir
, TGSI_OPCODE_XOR
, result_dst
, op
[0]);
1508 case ir_binop_bit_or
:
1509 if (glsl_version
>= 130) {
1510 emit(ir
, TGSI_OPCODE_OR
, result_dst
, op
[0]);
1513 case ir_unop_round_even
:
1514 assert(!"GLSL 1.30 features unsupported");
1517 case ir_quadop_vector
:
1518 /* This operation should have already been handled.
1520 assert(!"Should not get here.");
1524 this->result
= result_src
;
1529 glsl_to_tgsi_visitor::visit(ir_swizzle
*ir
)
1535 /* Note that this is only swizzles in expressions, not those on the left
1536 * hand side of an assignment, which do write masking. See ir_assignment
1540 ir
->val
->accept(this);
1542 assert(src
.file
!= PROGRAM_UNDEFINED
);
1544 for (i
= 0; i
< 4; i
++) {
1545 if (i
< ir
->type
->vector_elements
) {
1548 swizzle
[i
] = GET_SWZ(src
.swizzle
, ir
->mask
.x
);
1551 swizzle
[i
] = GET_SWZ(src
.swizzle
, ir
->mask
.y
);
1554 swizzle
[i
] = GET_SWZ(src
.swizzle
, ir
->mask
.z
);
1557 swizzle
[i
] = GET_SWZ(src
.swizzle
, ir
->mask
.w
);
1561 /* If the type is smaller than a vec4, replicate the last
1564 swizzle
[i
] = swizzle
[ir
->type
->vector_elements
- 1];
1568 src
.swizzle
= MAKE_SWIZZLE4(swizzle
[0], swizzle
[1], swizzle
[2], swizzle
[3]);
1574 glsl_to_tgsi_visitor::visit(ir_dereference_variable
*ir
)
1576 variable_storage
*entry
= find_variable_storage(ir
->var
);
1577 ir_variable
*var
= ir
->var
;
1580 switch (var
->mode
) {
1581 case ir_var_uniform
:
1582 entry
= new(mem_ctx
) variable_storage(var
, PROGRAM_UNIFORM
,
1584 this->variables
.push_tail(entry
);
1588 /* The linker assigns locations for varyings and attributes,
1589 * including deprecated builtins (like gl_Color), user-assign
1590 * generic attributes (glBindVertexLocation), and
1591 * user-defined varyings.
1593 * FINISHME: We would hit this path for function arguments. Fix!
1595 assert(var
->location
!= -1);
1596 entry
= new(mem_ctx
) variable_storage(var
,
1599 if (this->prog
->Target
== GL_VERTEX_PROGRAM_ARB
&&
1600 var
->location
>= VERT_ATTRIB_GENERIC0
) {
1601 _mesa_add_attribute(this->prog
->Attributes
,
1603 _mesa_sizeof_glsl_type(var
->type
->gl_type
),
1605 var
->location
- VERT_ATTRIB_GENERIC0
);
1609 assert(var
->location
!= -1);
1610 entry
= new(mem_ctx
) variable_storage(var
,
1614 case ir_var_system_value
:
1615 entry
= new(mem_ctx
) variable_storage(var
,
1616 PROGRAM_SYSTEM_VALUE
,
1620 case ir_var_temporary
:
1621 entry
= new(mem_ctx
) variable_storage(var
, PROGRAM_TEMPORARY
,
1623 this->variables
.push_tail(entry
);
1625 next_temp
+= type_size(var
->type
);
1630 printf("Failed to make storage for %s\n", var
->name
);
1635 this->result
= st_src_reg(entry
->file
, entry
->index
, var
->type
);
1636 if (glsl_version
<= 120)
1637 this->result
.type
= GLSL_TYPE_FLOAT
;
1641 glsl_to_tgsi_visitor::visit(ir_dereference_array
*ir
)
1645 int element_size
= type_size(ir
->type
);
1647 index
= ir
->array_index
->constant_expression_value();
1649 ir
->array
->accept(this);
1653 src
.index
+= index
->value
.i
[0] * element_size
;
1655 st_src_reg array_base
= this->result
;
1656 /* Variable index array dereference. It eats the "vec4" of the
1657 * base of the array and an index that offsets the Mesa register
1660 ir
->array_index
->accept(this);
1662 st_src_reg index_reg
;
1664 if (element_size
== 1) {
1665 index_reg
= this->result
;
1667 index_reg
= get_temp(glsl_type::float_type
);
1669 emit(ir
, TGSI_OPCODE_MUL
, st_dst_reg(index_reg
),
1670 this->result
, st_src_reg_for_float(element_size
));
1673 src
.reladdr
= ralloc(mem_ctx
, st_src_reg
);
1674 memcpy(src
.reladdr
, &index_reg
, sizeof(index_reg
));
1677 /* If the type is smaller than a vec4, replicate the last channel out. */
1678 if (ir
->type
->is_scalar() || ir
->type
->is_vector())
1679 src
.swizzle
= swizzle_for_size(ir
->type
->vector_elements
);
1681 src
.swizzle
= SWIZZLE_NOOP
;
1687 glsl_to_tgsi_visitor::visit(ir_dereference_record
*ir
)
1690 const glsl_type
*struct_type
= ir
->record
->type
;
1693 ir
->record
->accept(this);
1695 for (i
= 0; i
< struct_type
->length
; i
++) {
1696 if (strcmp(struct_type
->fields
.structure
[i
].name
, ir
->field
) == 0)
1698 offset
+= type_size(struct_type
->fields
.structure
[i
].type
);
1701 /* If the type is smaller than a vec4, replicate the last channel out. */
1702 if (ir
->type
->is_scalar() || ir
->type
->is_vector())
1703 this->result
.swizzle
= swizzle_for_size(ir
->type
->vector_elements
);
1705 this->result
.swizzle
= SWIZZLE_NOOP
;
1707 this->result
.index
+= offset
;
1711 * We want to be careful in assignment setup to hit the actual storage
1712 * instead of potentially using a temporary like we might with the
1713 * ir_dereference handler.
1716 get_assignment_lhs(ir_dereference
*ir
, glsl_to_tgsi_visitor
*v
)
1718 /* The LHS must be a dereference. If the LHS is a variable indexed array
1719 * access of a vector, it must be separated into a series conditional moves
1720 * before reaching this point (see ir_vec_index_to_cond_assign).
1722 assert(ir
->as_dereference());
1723 ir_dereference_array
*deref_array
= ir
->as_dereference_array();
1725 assert(!deref_array
->array
->type
->is_vector());
1728 /* Use the rvalue deref handler for the most part. We'll ignore
1729 * swizzles in it and write swizzles using writemask, though.
1732 return st_dst_reg(v
->result
);
1736 * Process the condition of a conditional assignment
1738 * Examines the condition of a conditional assignment to generate the optimal
1739 * first operand of a \c CMP instruction. If the condition is a relational
1740 * operator with 0 (e.g., \c ir_binop_less), the value being compared will be
1741 * used as the source for the \c CMP instruction. Otherwise the comparison
1742 * is processed to a boolean result, and the boolean result is used as the
1743 * operand to the CMP instruction.
1746 glsl_to_tgsi_visitor::process_move_condition(ir_rvalue
*ir
)
1748 ir_rvalue
*src_ir
= ir
;
1750 bool switch_order
= false;
1752 ir_expression
*const expr
= ir
->as_expression();
1753 if ((expr
!= NULL
) && (expr
->get_num_operands() == 2)) {
1754 bool zero_on_left
= false;
1756 if (expr
->operands
[0]->is_zero()) {
1757 src_ir
= expr
->operands
[1];
1758 zero_on_left
= true;
1759 } else if (expr
->operands
[1]->is_zero()) {
1760 src_ir
= expr
->operands
[0];
1761 zero_on_left
= false;
1765 * (a < 0) T F F ( a < 0) T F F
1766 * (0 < a) F F T (-a < 0) F F T
1767 * (a <= 0) T T F (-a < 0) F F T (swap order of other operands)
1768 * (0 <= a) F T T ( a < 0) T F F (swap order of other operands)
1769 * (a > 0) F F T (-a < 0) F F T
1770 * (0 > a) T F F ( a < 0) T F F
1771 * (a >= 0) F T T ( a < 0) T F F (swap order of other operands)
1772 * (0 >= a) T T F (-a < 0) F F T (swap order of other operands)
1774 * Note that exchanging the order of 0 and 'a' in the comparison simply
1775 * means that the value of 'a' should be negated.
1778 switch (expr
->operation
) {
1780 switch_order
= false;
1781 negate
= zero_on_left
;
1784 case ir_binop_greater
:
1785 switch_order
= false;
1786 negate
= !zero_on_left
;
1789 case ir_binop_lequal
:
1790 switch_order
= true;
1791 negate
= !zero_on_left
;
1794 case ir_binop_gequal
:
1795 switch_order
= true;
1796 negate
= zero_on_left
;
1800 /* This isn't the right kind of comparison afterall, so make sure
1801 * the whole condition is visited.
1809 src_ir
->accept(this);
1811 /* We use the TGSI_OPCODE_CMP (a < 0 ? b : c) for conditional moves, and the
1812 * condition we produced is 0.0 or 1.0. By flipping the sign, we can
1813 * choose which value TGSI_OPCODE_CMP produces without an extra instruction
1814 * computing the condition.
1817 this->result
.negate
= ~this->result
.negate
;
1819 return switch_order
;
1823 glsl_to_tgsi_visitor::visit(ir_assignment
*ir
)
1829 ir
->rhs
->accept(this);
1832 l
= get_assignment_lhs(ir
->lhs
, this);
1834 /* FINISHME: This should really set to the correct maximal writemask for each
1835 * FINISHME: component written (in the loops below). This case can only
1836 * FINISHME: occur for matrices, arrays, and structures.
1838 if (ir
->write_mask
== 0) {
1839 assert(!ir
->lhs
->type
->is_scalar() && !ir
->lhs
->type
->is_vector());
1840 l
.writemask
= WRITEMASK_XYZW
;
1841 } else if (ir
->lhs
->type
->is_scalar()) {
1842 /* FINISHME: This hack makes writing to gl_FragDepth, which lives in the
1843 * FINISHME: W component of fragment shader output zero, work correctly.
1845 l
.writemask
= WRITEMASK_XYZW
;
1848 int first_enabled_chan
= 0;
1851 assert(ir
->lhs
->type
->is_vector());
1852 l
.writemask
= ir
->write_mask
;
1854 for (int i
= 0; i
< 4; i
++) {
1855 if (l
.writemask
& (1 << i
)) {
1856 first_enabled_chan
= GET_SWZ(r
.swizzle
, i
);
1861 /* Swizzle a small RHS vector into the channels being written.
1863 * glsl ir treats write_mask as dictating how many channels are
1864 * present on the RHS while Mesa IR treats write_mask as just
1865 * showing which channels of the vec4 RHS get written.
1867 for (int i
= 0; i
< 4; i
++) {
1868 if (l
.writemask
& (1 << i
))
1869 swizzles
[i
] = GET_SWZ(r
.swizzle
, rhs_chan
++);
1871 swizzles
[i
] = first_enabled_chan
;
1873 r
.swizzle
= MAKE_SWIZZLE4(swizzles
[0], swizzles
[1],
1874 swizzles
[2], swizzles
[3]);
1877 assert(l
.file
!= PROGRAM_UNDEFINED
);
1878 assert(r
.file
!= PROGRAM_UNDEFINED
);
1880 if (ir
->condition
) {
1881 const bool switch_order
= this->process_move_condition(ir
->condition
);
1882 st_src_reg condition
= this->result
;
1884 for (i
= 0; i
< type_size(ir
->lhs
->type
); i
++) {
1885 st_src_reg l_src
= st_src_reg(l
);
1886 l_src
.swizzle
= swizzle_for_size(ir
->lhs
->type
->vector_elements
);
1889 emit(ir
, TGSI_OPCODE_CMP
, l
, condition
, l_src
, r
);
1891 emit(ir
, TGSI_OPCODE_CMP
, l
, condition
, r
, l_src
);
1898 for (i
= 0; i
< type_size(ir
->lhs
->type
); i
++) {
1899 emit(ir
, TGSI_OPCODE_MOV
, l
, r
);
1908 glsl_to_tgsi_visitor::visit(ir_constant
*ir
)
1911 GLfloat stack_vals
[4] = { 0 };
1912 gl_constant_value
*values
= (gl_constant_value
*) stack_vals
;
1913 GLenum gl_type
= GL_NONE
;
1916 /* Unfortunately, 4 floats is all we can get into
1917 * _mesa_add_unnamed_constant. So, make a temp to store an
1918 * aggregate constant and move each constant value into it. If we
1919 * get lucky, copy propagation will eliminate the extra moves.
1921 if (ir
->type
->base_type
== GLSL_TYPE_STRUCT
) {
1922 st_src_reg temp_base
= get_temp(ir
->type
);
1923 st_dst_reg temp
= st_dst_reg(temp_base
);
1925 foreach_iter(exec_list_iterator
, iter
, ir
->components
) {
1926 ir_constant
*field_value
= (ir_constant
*)iter
.get();
1927 int size
= type_size(field_value
->type
);
1931 field_value
->accept(this);
1934 for (i
= 0; i
< (unsigned int)size
; i
++) {
1935 emit(ir
, TGSI_OPCODE_MOV
, temp
, src
);
1941 this->result
= temp_base
;
1945 if (ir
->type
->is_array()) {
1946 st_src_reg temp_base
= get_temp(ir
->type
);
1947 st_dst_reg temp
= st_dst_reg(temp_base
);
1948 int size
= type_size(ir
->type
->fields
.array
);
1952 for (i
= 0; i
< ir
->type
->length
; i
++) {
1953 ir
->array_elements
[i
]->accept(this);
1955 for (int j
= 0; j
< size
; j
++) {
1956 emit(ir
, TGSI_OPCODE_MOV
, temp
, src
);
1962 this->result
= temp_base
;
1966 if (ir
->type
->is_matrix()) {
1967 st_src_reg mat
= get_temp(ir
->type
);
1968 st_dst_reg mat_column
= st_dst_reg(mat
);
1970 for (i
= 0; i
< ir
->type
->matrix_columns
; i
++) {
1971 assert(ir
->type
->base_type
== GLSL_TYPE_FLOAT
);
1972 values
= (gl_constant_value
*) &ir
->value
.f
[i
* ir
->type
->vector_elements
];
1974 src
= st_src_reg(PROGRAM_CONSTANT
, -1, ir
->type
->base_type
);
1975 src
.index
= _mesa_add_typed_unnamed_constant(this->prog
->Parameters
,
1977 ir
->type
->vector_elements
,
1980 emit(ir
, TGSI_OPCODE_MOV
, mat_column
, src
);
1989 src
.file
= PROGRAM_CONSTANT
;
1990 switch (ir
->type
->base_type
) {
1991 case GLSL_TYPE_FLOAT
:
1993 for (i
= 0; i
< ir
->type
->vector_elements
; i
++) {
1994 values
[i
].f
= ir
->value
.f
[i
];
1997 case GLSL_TYPE_UINT
:
1998 gl_type
= glsl_version
>= 130 ? GL_UNSIGNED_INT
: GL_FLOAT
;
1999 for (i
= 0; i
< ir
->type
->vector_elements
; i
++) {
2000 if (glsl_version
>= 130)
2001 values
[i
].u
= ir
->value
.u
[i
];
2003 values
[i
].f
= ir
->value
.u
[i
];
2007 gl_type
= glsl_version
>= 130 ? GL_INT
: GL_FLOAT
;
2008 for (i
= 0; i
< ir
->type
->vector_elements
; i
++) {
2009 if (glsl_version
>= 130)
2010 values
[i
].i
= ir
->value
.i
[i
];
2012 values
[i
].f
= ir
->value
.i
[i
];
2015 case GLSL_TYPE_BOOL
:
2016 gl_type
= glsl_version
>= 130 ? GL_BOOL
: GL_FLOAT
;
2017 for (i
= 0; i
< ir
->type
->vector_elements
; i
++) {
2018 if (glsl_version
>= 130)
2019 values
[i
].b
= ir
->value
.b
[i
];
2021 values
[i
].f
= ir
->value
.b
[i
];
2025 assert(!"Non-float/uint/int/bool constant");
2028 this->result
= st_src_reg(PROGRAM_CONSTANT
, -1, ir
->type
);
2029 this->result
.index
= _mesa_add_typed_unnamed_constant(this->prog
->Parameters
,
2030 values
, ir
->type
->vector_elements
, gl_type
,
2031 &this->result
.swizzle
);
2035 glsl_to_tgsi_visitor::get_function_signature(ir_function_signature
*sig
)
2037 function_entry
*entry
;
2039 foreach_iter(exec_list_iterator
, iter
, this->function_signatures
) {
2040 entry
= (function_entry
*)iter
.get();
2042 if (entry
->sig
== sig
)
2046 entry
= ralloc(mem_ctx
, function_entry
);
2048 entry
->sig_id
= this->next_signature_id
++;
2049 entry
->bgn_inst
= NULL
;
2051 /* Allocate storage for all the parameters. */
2052 foreach_iter(exec_list_iterator
, iter
, sig
->parameters
) {
2053 ir_variable
*param
= (ir_variable
*)iter
.get();
2054 variable_storage
*storage
;
2056 storage
= find_variable_storage(param
);
2059 storage
= new(mem_ctx
) variable_storage(param
, PROGRAM_TEMPORARY
,
2061 this->variables
.push_tail(storage
);
2063 this->next_temp
+= type_size(param
->type
);
2066 if (!sig
->return_type
->is_void()) {
2067 entry
->return_reg
= get_temp(sig
->return_type
);
2069 entry
->return_reg
= undef_src
;
2072 this->function_signatures
.push_tail(entry
);
2077 glsl_to_tgsi_visitor::visit(ir_call
*ir
)
2079 glsl_to_tgsi_instruction
*call_inst
;
2080 ir_function_signature
*sig
= ir
->get_callee();
2081 function_entry
*entry
= get_function_signature(sig
);
2084 /* Process in parameters. */
2085 exec_list_iterator sig_iter
= sig
->parameters
.iterator();
2086 foreach_iter(exec_list_iterator
, iter
, *ir
) {
2087 ir_rvalue
*param_rval
= (ir_rvalue
*)iter
.get();
2088 ir_variable
*param
= (ir_variable
*)sig_iter
.get();
2090 if (param
->mode
== ir_var_in
||
2091 param
->mode
== ir_var_inout
) {
2092 variable_storage
*storage
= find_variable_storage(param
);
2095 param_rval
->accept(this);
2096 st_src_reg r
= this->result
;
2099 l
.file
= storage
->file
;
2100 l
.index
= storage
->index
;
2102 l
.writemask
= WRITEMASK_XYZW
;
2103 l
.cond_mask
= COND_TR
;
2105 for (i
= 0; i
< type_size(param
->type
); i
++) {
2106 emit(ir
, TGSI_OPCODE_MOV
, l
, r
);
2114 assert(!sig_iter
.has_next());
2116 /* Emit call instruction */
2117 call_inst
= emit(ir
, TGSI_OPCODE_CAL
);
2118 call_inst
->function
= entry
;
2120 /* Process out parameters. */
2121 sig_iter
= sig
->parameters
.iterator();
2122 foreach_iter(exec_list_iterator
, iter
, *ir
) {
2123 ir_rvalue
*param_rval
= (ir_rvalue
*)iter
.get();
2124 ir_variable
*param
= (ir_variable
*)sig_iter
.get();
2126 if (param
->mode
== ir_var_out
||
2127 param
->mode
== ir_var_inout
) {
2128 variable_storage
*storage
= find_variable_storage(param
);
2132 r
.file
= storage
->file
;
2133 r
.index
= storage
->index
;
2135 r
.swizzle
= SWIZZLE_NOOP
;
2138 param_rval
->accept(this);
2139 st_dst_reg l
= st_dst_reg(this->result
);
2141 for (i
= 0; i
< type_size(param
->type
); i
++) {
2142 emit(ir
, TGSI_OPCODE_MOV
, l
, r
);
2150 assert(!sig_iter
.has_next());
2152 /* Process return value. */
2153 this->result
= entry
->return_reg
;
2157 glsl_to_tgsi_visitor::visit(ir_texture
*ir
)
2159 st_src_reg result_src
, coord
, lod_info
, projector
, dx
, dy
;
2160 st_dst_reg result_dst
, coord_dst
;
2161 glsl_to_tgsi_instruction
*inst
= NULL
;
2162 unsigned opcode
= TGSI_OPCODE_NOP
;
2164 ir
->coordinate
->accept(this);
2166 /* Put our coords in a temp. We'll need to modify them for shadow,
2167 * projection, or LOD, so the only case we'd use it as is is if
2168 * we're doing plain old texturing. Mesa IR optimization should
2169 * handle cleaning up our mess in that case.
2171 coord
= get_temp(glsl_type::vec4_type
);
2172 coord_dst
= st_dst_reg(coord
);
2173 emit(ir
, TGSI_OPCODE_MOV
, coord_dst
, this->result
);
2175 if (ir
->projector
) {
2176 ir
->projector
->accept(this);
2177 projector
= this->result
;
2180 /* Storage for our result. Ideally for an assignment we'd be using
2181 * the actual storage for the result here, instead.
2183 result_src
= get_temp(glsl_type::vec4_type
);
2184 result_dst
= st_dst_reg(result_src
);
2188 opcode
= TGSI_OPCODE_TEX
;
2191 opcode
= TGSI_OPCODE_TXB
;
2192 ir
->lod_info
.bias
->accept(this);
2193 lod_info
= this->result
;
2196 opcode
= TGSI_OPCODE_TXL
;
2197 ir
->lod_info
.lod
->accept(this);
2198 lod_info
= this->result
;
2201 opcode
= TGSI_OPCODE_TXD
;
2202 ir
->lod_info
.grad
.dPdx
->accept(this);
2204 ir
->lod_info
.grad
.dPdy
->accept(this);
2207 case ir_txf
: /* TODO: use TGSI_OPCODE_TXF here */
2208 assert(!"GLSL 1.30 features unsupported");
2212 if (ir
->projector
) {
2213 if (opcode
== TGSI_OPCODE_TEX
) {
2214 /* Slot the projector in as the last component of the coord. */
2215 coord_dst
.writemask
= WRITEMASK_W
;
2216 emit(ir
, TGSI_OPCODE_MOV
, coord_dst
, projector
);
2217 coord_dst
.writemask
= WRITEMASK_XYZW
;
2218 opcode
= TGSI_OPCODE_TXP
;
2220 st_src_reg coord_w
= coord
;
2221 coord_w
.swizzle
= SWIZZLE_WWWW
;
2223 /* For the other TEX opcodes there's no projective version
2224 * since the last slot is taken up by LOD info. Do the
2225 * projective divide now.
2227 coord_dst
.writemask
= WRITEMASK_W
;
2228 emit(ir
, TGSI_OPCODE_RCP
, coord_dst
, projector
);
2230 /* In the case where we have to project the coordinates "by hand,"
2231 * the shadow comparator value must also be projected.
2233 st_src_reg tmp_src
= coord
;
2234 if (ir
->shadow_comparitor
) {
2235 /* Slot the shadow value in as the second to last component of the
2238 ir
->shadow_comparitor
->accept(this);
2240 tmp_src
= get_temp(glsl_type::vec4_type
);
2241 st_dst_reg tmp_dst
= st_dst_reg(tmp_src
);
2243 tmp_dst
.writemask
= WRITEMASK_Z
;
2244 emit(ir
, TGSI_OPCODE_MOV
, tmp_dst
, this->result
);
2246 tmp_dst
.writemask
= WRITEMASK_XY
;
2247 emit(ir
, TGSI_OPCODE_MOV
, tmp_dst
, coord
);
2250 coord_dst
.writemask
= WRITEMASK_XYZ
;
2251 emit(ir
, TGSI_OPCODE_MUL
, coord_dst
, tmp_src
, coord_w
);
2253 coord_dst
.writemask
= WRITEMASK_XYZW
;
2254 coord
.swizzle
= SWIZZLE_XYZW
;
2258 /* If projection is done and the opcode is not TGSI_OPCODE_TXP, then the shadow
2259 * comparator was put in the correct place (and projected) by the code,
2260 * above, that handles by-hand projection.
2262 if (ir
->shadow_comparitor
&& (!ir
->projector
|| opcode
== TGSI_OPCODE_TXP
)) {
2263 /* Slot the shadow value in as the second to last component of the
2266 ir
->shadow_comparitor
->accept(this);
2267 coord_dst
.writemask
= WRITEMASK_Z
;
2268 emit(ir
, TGSI_OPCODE_MOV
, coord_dst
, this->result
);
2269 coord_dst
.writemask
= WRITEMASK_XYZW
;
2272 if (opcode
== TGSI_OPCODE_TXL
|| opcode
== TGSI_OPCODE_TXB
) {
2273 /* TGSI stores LOD or LOD bias in the last channel of the coords. */
2274 coord_dst
.writemask
= WRITEMASK_W
;
2275 emit(ir
, TGSI_OPCODE_MOV
, coord_dst
, lod_info
);
2276 coord_dst
.writemask
= WRITEMASK_XYZW
;
2279 if (opcode
== TGSI_OPCODE_TXD
)
2280 inst
= emit(ir
, opcode
, result_dst
, coord
, dx
, dy
);
2282 inst
= emit(ir
, opcode
, result_dst
, coord
);
2284 if (ir
->shadow_comparitor
)
2285 inst
->tex_shadow
= GL_TRUE
;
2287 inst
->sampler
= _mesa_get_sampler_uniform_value(ir
->sampler
,
2288 this->shader_program
,
2291 const glsl_type
*sampler_type
= ir
->sampler
->type
;
2293 switch (sampler_type
->sampler_dimensionality
) {
2294 case GLSL_SAMPLER_DIM_1D
:
2295 inst
->tex_target
= (sampler_type
->sampler_array
)
2296 ? TEXTURE_1D_ARRAY_INDEX
: TEXTURE_1D_INDEX
;
2298 case GLSL_SAMPLER_DIM_2D
:
2299 inst
->tex_target
= (sampler_type
->sampler_array
)
2300 ? TEXTURE_2D_ARRAY_INDEX
: TEXTURE_2D_INDEX
;
2302 case GLSL_SAMPLER_DIM_3D
:
2303 inst
->tex_target
= TEXTURE_3D_INDEX
;
2305 case GLSL_SAMPLER_DIM_CUBE
:
2306 inst
->tex_target
= TEXTURE_CUBE_INDEX
;
2308 case GLSL_SAMPLER_DIM_RECT
:
2309 inst
->tex_target
= TEXTURE_RECT_INDEX
;
2311 case GLSL_SAMPLER_DIM_BUF
:
2312 assert(!"FINISHME: Implement ARB_texture_buffer_object");
2315 assert(!"Should not get here.");
2318 this->result
= result_src
;
2322 glsl_to_tgsi_visitor::visit(ir_return
*ir
)
2324 if (ir
->get_value()) {
2328 assert(current_function
);
2330 ir
->get_value()->accept(this);
2331 st_src_reg r
= this->result
;
2333 l
= st_dst_reg(current_function
->return_reg
);
2335 for (i
= 0; i
< type_size(current_function
->sig
->return_type
); i
++) {
2336 emit(ir
, TGSI_OPCODE_MOV
, l
, r
);
2342 emit(ir
, TGSI_OPCODE_RET
);
2346 glsl_to_tgsi_visitor::visit(ir_discard
*ir
)
2348 struct gl_fragment_program
*fp
= (struct gl_fragment_program
*)this->prog
;
2350 if (ir
->condition
) {
2351 ir
->condition
->accept(this);
2352 this->result
.negate
= ~this->result
.negate
;
2353 emit(ir
, TGSI_OPCODE_KIL
, undef_dst
, this->result
);
2355 emit(ir
, TGSI_OPCODE_KILP
);
2358 fp
->UsesKill
= GL_TRUE
;
2362 glsl_to_tgsi_visitor::visit(ir_if
*ir
)
2364 glsl_to_tgsi_instruction
*cond_inst
, *if_inst
, *else_inst
= NULL
;
2365 glsl_to_tgsi_instruction
*prev_inst
;
2367 prev_inst
= (glsl_to_tgsi_instruction
*)this->instructions
.get_tail();
2369 ir
->condition
->accept(this);
2370 assert(this->result
.file
!= PROGRAM_UNDEFINED
);
2372 if (this->options
->EmitCondCodes
) {
2373 cond_inst
= (glsl_to_tgsi_instruction
*)this->instructions
.get_tail();
2375 /* See if we actually generated any instruction for generating
2376 * the condition. If not, then cook up a move to a temp so we
2377 * have something to set cond_update on.
2379 if (cond_inst
== prev_inst
) {
2380 st_src_reg temp
= get_temp(glsl_type::bool_type
);
2381 cond_inst
= emit(ir
->condition
, TGSI_OPCODE_MOV
, st_dst_reg(temp
), result
);
2383 cond_inst
->cond_update
= GL_TRUE
;
2385 if_inst
= emit(ir
->condition
, TGSI_OPCODE_IF
);
2386 if_inst
->dst
.cond_mask
= COND_NE
;
2388 if_inst
= emit(ir
->condition
, TGSI_OPCODE_IF
, undef_dst
, this->result
);
2391 this->instructions
.push_tail(if_inst
);
2393 visit_exec_list(&ir
->then_instructions
, this);
2395 if (!ir
->else_instructions
.is_empty()) {
2396 else_inst
= emit(ir
->condition
, TGSI_OPCODE_ELSE
);
2397 visit_exec_list(&ir
->else_instructions
, this);
2400 if_inst
= emit(ir
->condition
, TGSI_OPCODE_ENDIF
);
2403 glsl_to_tgsi_visitor::glsl_to_tgsi_visitor()
2405 result
.file
= PROGRAM_UNDEFINED
;
2407 next_signature_id
= 1;
2408 current_function
= NULL
;
2409 num_address_regs
= 0;
2410 indirect_addr_temps
= false;
2411 indirect_addr_consts
= false;
2412 mem_ctx
= ralloc_context(NULL
);
2415 glsl_to_tgsi_visitor::~glsl_to_tgsi_visitor()
2417 ralloc_free(mem_ctx
);
2420 extern "C" void free_glsl_to_tgsi_visitor(glsl_to_tgsi_visitor
*v
)
2427 * Count resources used by the given gpu program (number of texture
2431 count_resources(glsl_to_tgsi_visitor
*v
, gl_program
*prog
)
2433 v
->samplers_used
= 0;
2435 foreach_iter(exec_list_iterator
, iter
, v
->instructions
) {
2436 glsl_to_tgsi_instruction
*inst
= (glsl_to_tgsi_instruction
*)iter
.get();
2438 if (is_tex_instruction(inst
->op
)) {
2439 v
->samplers_used
|= 1 << inst
->sampler
;
2441 prog
->SamplerTargets
[inst
->sampler
] =
2442 (gl_texture_index
)inst
->tex_target
;
2443 if (inst
->tex_shadow
) {
2444 prog
->ShadowSamplers
|= 1 << inst
->sampler
;
2449 prog
->SamplersUsed
= v
->samplers_used
;
2450 _mesa_update_shader_textures_used(prog
);
2455 * Check if the given vertex/fragment/shader program is within the
2456 * resource limits of the context (number of texture units, etc).
2457 * If any of those checks fail, record a linker error.
2459 * XXX more checks are needed...
2462 check_resources(const struct gl_context
*ctx
,
2463 struct gl_shader_program
*shader_program
,
2464 glsl_to_tgsi_visitor
*prog
,
2465 struct gl_program
*proginfo
)
2467 switch (proginfo
->Target
) {
2468 case GL_VERTEX_PROGRAM_ARB
:
2469 if (_mesa_bitcount(prog
->samplers_used
) >
2470 ctx
->Const
.MaxVertexTextureImageUnits
) {
2471 fail_link(shader_program
, "Too many vertex shader texture samplers");
2473 if (proginfo
->Parameters
->NumParameters
> MAX_UNIFORMS
) {
2474 fail_link(shader_program
, "Too many vertex shader constants");
2477 case MESA_GEOMETRY_PROGRAM
:
2478 if (_mesa_bitcount(prog
->samplers_used
) >
2479 ctx
->Const
.MaxGeometryTextureImageUnits
) {
2480 fail_link(shader_program
, "Too many geometry shader texture samplers");
2482 if (proginfo
->Parameters
->NumParameters
>
2483 MAX_GEOMETRY_UNIFORM_COMPONENTS
/ 4) {
2484 fail_link(shader_program
, "Too many geometry shader constants");
2487 case GL_FRAGMENT_PROGRAM_ARB
:
2488 if (_mesa_bitcount(prog
->samplers_used
) >
2489 ctx
->Const
.MaxTextureImageUnits
) {
2490 fail_link(shader_program
, "Too many fragment shader texture samplers");
2492 if (proginfo
->Parameters
->NumParameters
> MAX_UNIFORMS
) {
2493 fail_link(shader_program
, "Too many fragment shader constants");
2497 _mesa_problem(ctx
, "unexpected program type in check_resources()");
2503 struct uniform_sort
{
2504 struct gl_uniform
*u
;
2508 /* The shader_program->Uniforms list is almost sorted in increasing
2509 * uniform->{Frag,Vert}Pos locations, but not quite when there are
2510 * uniforms shared between targets. We need to add parameters in
2511 * increasing order for the targets.
2514 sort_uniforms(const void *a
, const void *b
)
2516 struct uniform_sort
*u1
= (struct uniform_sort
*)a
;
2517 struct uniform_sort
*u2
= (struct uniform_sort
*)b
;
2519 return u1
->pos
- u2
->pos
;
2522 /* Add the uniforms to the parameters. The linker chose locations
2523 * in our parameters lists (which weren't created yet), which the
2524 * uniforms code will use to poke values into our parameters list
2525 * when uniforms are updated.
2528 add_uniforms_to_parameters_list(struct gl_shader_program
*shader_program
,
2529 struct gl_shader
*shader
,
2530 struct gl_program
*prog
)
2533 unsigned int next_sampler
= 0, num_uniforms
= 0;
2534 struct uniform_sort
*sorted_uniforms
;
2536 sorted_uniforms
= ralloc_array(NULL
, struct uniform_sort
,
2537 shader_program
->Uniforms
->NumUniforms
);
2539 for (i
= 0; i
< shader_program
->Uniforms
->NumUniforms
; i
++) {
2540 struct gl_uniform
*uniform
= shader_program
->Uniforms
->Uniforms
+ i
;
2541 int parameter_index
= -1;
2543 switch (shader
->Type
) {
2544 case GL_VERTEX_SHADER
:
2545 parameter_index
= uniform
->VertPos
;
2547 case GL_FRAGMENT_SHADER
:
2548 parameter_index
= uniform
->FragPos
;
2550 case GL_GEOMETRY_SHADER
:
2551 parameter_index
= uniform
->GeomPos
;
2555 /* Only add uniforms used in our target. */
2556 if (parameter_index
!= -1) {
2557 sorted_uniforms
[num_uniforms
].pos
= parameter_index
;
2558 sorted_uniforms
[num_uniforms
].u
= uniform
;
2563 qsort(sorted_uniforms
, num_uniforms
, sizeof(struct uniform_sort
),
2566 for (i
= 0; i
< num_uniforms
; i
++) {
2567 struct gl_uniform
*uniform
= sorted_uniforms
[i
].u
;
2568 int parameter_index
= sorted_uniforms
[i
].pos
;
2569 const glsl_type
*type
= uniform
->Type
;
2572 if (type
->is_vector() ||
2573 type
->is_scalar()) {
2574 size
= type
->vector_elements
;
2576 size
= type_size(type
) * 4;
2579 gl_register_file file
;
2580 if (type
->is_sampler() ||
2581 (type
->is_array() && type
->fields
.array
->is_sampler())) {
2582 file
= PROGRAM_SAMPLER
;
2584 file
= PROGRAM_UNIFORM
;
2587 GLint index
= _mesa_lookup_parameter_index(prog
->Parameters
, -1,
2591 index
= _mesa_add_parameter(prog
->Parameters
, file
,
2592 uniform
->Name
, size
, type
->gl_type
,
2595 /* Sampler uniform values are stored in prog->SamplerUnits,
2596 * and the entry in that array is selected by this index we
2597 * store in ParameterValues[].
2599 if (file
== PROGRAM_SAMPLER
) {
2600 for (unsigned int j
= 0; j
< size
/ 4; j
++)
2601 prog
->Parameters
->ParameterValues
[index
+ j
][0].f
= next_sampler
++;
2604 /* The location chosen in the Parameters list here (returned
2605 * from _mesa_add_uniform) has to match what the linker chose.
2607 if (index
!= parameter_index
) {
2608 fail_link(shader_program
, "Allocation of uniform `%s' to target "
2609 "failed (%d vs %d)\n",
2610 uniform
->Name
, index
, parameter_index
);
2615 ralloc_free(sorted_uniforms
);
2619 set_uniform_initializer(struct gl_context
*ctx
, void *mem_ctx
,
2620 struct gl_shader_program
*shader_program
,
2621 const char *name
, const glsl_type
*type
,
2624 if (type
->is_record()) {
2625 ir_constant
*field_constant
;
2627 field_constant
= (ir_constant
*)val
->components
.get_head();
2629 for (unsigned int i
= 0; i
< type
->length
; i
++) {
2630 const glsl_type
*field_type
= type
->fields
.structure
[i
].type
;
2631 const char *field_name
= ralloc_asprintf(mem_ctx
, "%s.%s", name
,
2632 type
->fields
.structure
[i
].name
);
2633 set_uniform_initializer(ctx
, mem_ctx
, shader_program
, field_name
,
2634 field_type
, field_constant
);
2635 field_constant
= (ir_constant
*)field_constant
->next
;
2640 int loc
= _mesa_get_uniform_location(ctx
, shader_program
, name
);
2643 fail_link(shader_program
,
2644 "Couldn't find uniform for initializer %s\n", name
);
2648 for (unsigned int i
= 0; i
< (type
->is_array() ? type
->length
: 1); i
++) {
2649 ir_constant
*element
;
2650 const glsl_type
*element_type
;
2651 if (type
->is_array()) {
2652 element
= val
->array_elements
[i
];
2653 element_type
= type
->fields
.array
;
2656 element_type
= type
;
2661 if (element_type
->base_type
== GLSL_TYPE_BOOL
) {
2662 int *conv
= ralloc_array(mem_ctx
, int, element_type
->components());
2663 for (unsigned int j
= 0; j
< element_type
->components(); j
++) {
2664 conv
[j
] = element
->value
.b
[j
];
2666 values
= (void *)conv
;
2667 element_type
= glsl_type::get_instance(GLSL_TYPE_INT
,
2668 element_type
->vector_elements
,
2671 values
= &element
->value
;
2674 if (element_type
->is_matrix()) {
2675 _mesa_uniform_matrix(ctx
, shader_program
,
2676 element_type
->matrix_columns
,
2677 element_type
->vector_elements
,
2678 loc
, 1, GL_FALSE
, (GLfloat
*)values
);
2679 loc
+= element_type
->matrix_columns
;
2681 _mesa_uniform(ctx
, shader_program
, loc
, element_type
->matrix_columns
,
2682 values
, element_type
->gl_type
);
2683 loc
+= type_size(element_type
);
2689 set_uniform_initializers(struct gl_context
*ctx
,
2690 struct gl_shader_program
*shader_program
)
2692 void *mem_ctx
= NULL
;
2694 for (unsigned int i
= 0; i
< MESA_SHADER_TYPES
; i
++) {
2695 struct gl_shader
*shader
= shader_program
->_LinkedShaders
[i
];
2700 foreach_iter(exec_list_iterator
, iter
, *shader
->ir
) {
2701 ir_instruction
*ir
= (ir_instruction
*)iter
.get();
2702 ir_variable
*var
= ir
->as_variable();
2704 if (!var
|| var
->mode
!= ir_var_uniform
|| !var
->constant_value
)
2708 mem_ctx
= ralloc_context(NULL
);
2710 set_uniform_initializer(ctx
, mem_ctx
, shader_program
, var
->name
,
2711 var
->type
, var
->constant_value
);
2715 ralloc_free(mem_ctx
);
2719 * Scan/rewrite program to remove reads of custom (output) registers.
2720 * The passed type has to be either PROGRAM_OUTPUT or PROGRAM_VARYING
2721 * (for vertex shaders).
2722 * In GLSL shaders, varying vars can be read and written.
2723 * On some hardware, trying to read an output register causes trouble.
2724 * So, rewrite the program to use a temporary register in this case.
2726 * Based on _mesa_remove_output_reads from programopt.c.
2729 glsl_to_tgsi_visitor::remove_output_reads(gl_register_file type
)
2732 GLint outputMap
[VERT_RESULT_MAX
];
2733 GLint outputTypes
[VERT_RESULT_MAX
];
2734 GLuint numVaryingReads
= 0;
2735 GLboolean usedTemps
[MAX_PROGRAM_TEMPS
];
2736 GLuint firstTemp
= 0;
2738 _mesa_find_used_registers(prog
, PROGRAM_TEMPORARY
,
2739 usedTemps
, MAX_PROGRAM_TEMPS
);
2741 assert(type
== PROGRAM_VARYING
|| type
== PROGRAM_OUTPUT
);
2742 assert(prog
->Target
== GL_VERTEX_PROGRAM_ARB
|| type
!= PROGRAM_VARYING
);
2744 for (i
= 0; i
< VERT_RESULT_MAX
; i
++)
2747 /* look for instructions which read from varying vars */
2748 foreach_iter(exec_list_iterator
, iter
, this->instructions
) {
2749 glsl_to_tgsi_instruction
*inst
= (glsl_to_tgsi_instruction
*)iter
.get();
2750 const GLuint numSrc
= num_inst_src_regs(inst
->op
);
2752 for (j
= 0; j
< numSrc
; j
++) {
2753 if (inst
->src
[j
].file
== type
) {
2754 /* replace the read with a temp reg */
2755 const GLuint var
= inst
->src
[j
].index
;
2756 if (outputMap
[var
] == -1) {
2758 outputMap
[var
] = _mesa_find_free_register(usedTemps
,
2761 outputTypes
[var
] = inst
->src
[j
].type
;
2762 firstTemp
= outputMap
[var
] + 1;
2764 inst
->src
[j
].file
= PROGRAM_TEMPORARY
;
2765 inst
->src
[j
].index
= outputMap
[var
];
2770 if (numVaryingReads
== 0)
2771 return; /* nothing to be done */
2773 /* look for instructions which write to the varying vars identified above */
2774 foreach_iter(exec_list_iterator
, iter
, this->instructions
) {
2775 glsl_to_tgsi_instruction
*inst
= (glsl_to_tgsi_instruction
*)iter
.get();
2776 if (inst
->dst
.file
== type
&& outputMap
[inst
->dst
.index
] >= 0) {
2777 /* change inst to write to the temp reg, instead of the varying */
2778 inst
->dst
.file
= PROGRAM_TEMPORARY
;
2779 inst
->dst
.index
= outputMap
[inst
->dst
.index
];
2783 /* insert new MOV instructions at the end */
2784 for (i
= 0; i
< VERT_RESULT_MAX
; i
++) {
2785 if (outputMap
[i
] >= 0) {
2786 /* MOV VAR[i], TEMP[tmp]; */
2787 st_src_reg src
= st_src_reg(PROGRAM_TEMPORARY
, outputMap
[i
], outputTypes
[i
]);
2788 st_dst_reg dst
= st_dst_reg(type
, WRITEMASK_XYZW
, outputTypes
[i
]);
2790 this->emit(NULL
, TGSI_OPCODE_MOV
, dst
, src
);
2796 * Returns the mask of channels (bitmask of WRITEMASK_X,Y,Z,W) which
2797 * are read from the given src in this instruction
2800 get_src_arg_mask(st_dst_reg dst
, st_src_reg src
)
2802 int read_mask
= 0, comp
;
2804 /* Now, given the src swizzle and the written channels, find which
2805 * components are actually read
2807 for (comp
= 0; comp
< 4; ++comp
) {
2808 const unsigned coord
= GET_SWZ(src
.swizzle
, comp
);
2810 if (dst
.writemask
& (1 << comp
) && coord
<= SWIZZLE_W
)
2811 read_mask
|= 1 << coord
;
2818 * This pass replaces CMP T0, T1 T2 T0 with MOV T0, T2 when the CMP
2819 * instruction is the first instruction to write to register T0. There are
2820 * several lowering passes done in GLSL IR (e.g. branches and
2821 * relative addressing) that create a large number of conditional assignments
2822 * that ir_to_mesa converts to CMP instructions like the one mentioned above.
2824 * Here is why this conversion is safe:
2825 * CMP T0, T1 T2 T0 can be expanded to:
2831 * If (T1 < 0.0) evaluates to true then our replacement MOV T0, T2 is the same
2832 * as the original program. If (T1 < 0.0) evaluates to false, executing
2833 * MOV T0, T0 will store a garbage value in T0 since T0 is uninitialized.
2834 * Therefore, it doesn't matter that we are replacing MOV T0, T0 with MOV T0, T2
2835 * because any instruction that was going to read from T0 after this was going
2836 * to read a garbage value anyway.
2839 glsl_to_tgsi_visitor::simplify_cmp(void)
2841 unsigned tempWrites
[MAX_PROGRAM_TEMPS
];
2842 unsigned outputWrites
[MAX_PROGRAM_OUTPUTS
];
2844 memset(tempWrites
, 0, sizeof(tempWrites
));
2845 memset(outputWrites
, 0, sizeof(outputWrites
));
2847 foreach_iter(exec_list_iterator
, iter
, this->instructions
) {
2848 glsl_to_tgsi_instruction
*inst
= (glsl_to_tgsi_instruction
*)iter
.get();
2849 unsigned prevWriteMask
= 0;
2851 /* Give up if we encounter relative addressing or flow control. */
2852 if (inst
->dst
.reladdr
||
2853 tgsi_get_opcode_info(inst
->op
)->is_branch
||
2854 inst
->op
== TGSI_OPCODE_BGNSUB
||
2855 inst
->op
== TGSI_OPCODE_CONT
||
2856 inst
->op
== TGSI_OPCODE_END
||
2857 inst
->op
== TGSI_OPCODE_ENDSUB
||
2858 inst
->op
== TGSI_OPCODE_RET
) {
2862 if (inst
->dst
.file
== PROGRAM_OUTPUT
) {
2863 assert(inst
->dst
.index
< MAX_PROGRAM_OUTPUTS
);
2864 prevWriteMask
= outputWrites
[inst
->dst
.index
];
2865 outputWrites
[inst
->dst
.index
] |= inst
->dst
.writemask
;
2866 } else if (inst
->dst
.file
== PROGRAM_TEMPORARY
) {
2867 assert(inst
->dst
.index
< MAX_PROGRAM_TEMPS
);
2868 prevWriteMask
= tempWrites
[inst
->dst
.index
];
2869 tempWrites
[inst
->dst
.index
] |= inst
->dst
.writemask
;
2872 /* For a CMP to be considered a conditional write, the destination
2873 * register and source register two must be the same. */
2874 if (inst
->op
== TGSI_OPCODE_CMP
2875 && !(inst
->dst
.writemask
& prevWriteMask
)
2876 && inst
->src
[2].file
== inst
->dst
.file
2877 && inst
->src
[2].index
== inst
->dst
.index
2878 && inst
->dst
.writemask
== get_src_arg_mask(inst
->dst
, inst
->src
[2])) {
2880 inst
->op
= TGSI_OPCODE_MOV
;
2881 inst
->src
[0] = inst
->src
[1];
2886 /* Replaces all references to a temporary register index with another index. */
2888 glsl_to_tgsi_visitor::rename_temp_register(int index
, int new_index
)
2890 foreach_iter(exec_list_iterator
, iter
, this->instructions
) {
2891 glsl_to_tgsi_instruction
*inst
= (glsl_to_tgsi_instruction
*)iter
.get();
2894 for (j
=0; j
< num_inst_src_regs(inst
->op
); j
++) {
2895 if (inst
->src
[j
].file
== PROGRAM_TEMPORARY
&&
2896 inst
->src
[j
].index
== index
) {
2897 inst
->src
[j
].index
= new_index
;
2901 if (inst
->dst
.file
== PROGRAM_TEMPORARY
&& inst
->dst
.index
== index
) {
2902 inst
->dst
.index
= new_index
;
2908 glsl_to_tgsi_visitor::get_first_temp_read(int index
)
2910 int depth
= 0; /* loop depth */
2911 int loop_start
= -1; /* index of the first active BGNLOOP (if any) */
2914 foreach_iter(exec_list_iterator
, iter
, this->instructions
) {
2915 glsl_to_tgsi_instruction
*inst
= (glsl_to_tgsi_instruction
*)iter
.get();
2917 for (j
=0; j
< num_inst_src_regs(inst
->op
); j
++) {
2918 if (inst
->src
[j
].file
== PROGRAM_TEMPORARY
&&
2919 inst
->src
[j
].index
== index
) {
2920 return (depth
== 0) ? i
: loop_start
;
2924 if (inst
->op
== TGSI_OPCODE_BGNLOOP
) {
2927 } else if (inst
->op
== TGSI_OPCODE_ENDLOOP
) {
2940 glsl_to_tgsi_visitor::get_first_temp_write(int index
)
2942 int depth
= 0; /* loop depth */
2943 int loop_start
= -1; /* index of the first active BGNLOOP (if any) */
2946 foreach_iter(exec_list_iterator
, iter
, this->instructions
) {
2947 glsl_to_tgsi_instruction
*inst
= (glsl_to_tgsi_instruction
*)iter
.get();
2949 if (inst
->dst
.file
== PROGRAM_TEMPORARY
&& inst
->dst
.index
== index
) {
2950 return (depth
== 0) ? i
: loop_start
;
2953 if (inst
->op
== TGSI_OPCODE_BGNLOOP
) {
2956 } else if (inst
->op
== TGSI_OPCODE_ENDLOOP
) {
2969 glsl_to_tgsi_visitor::get_last_temp_read(int index
)
2971 int depth
= 0; /* loop depth */
2972 int last
= -1; /* index of last instruction that reads the temporary */
2975 foreach_iter(exec_list_iterator
, iter
, this->instructions
) {
2976 glsl_to_tgsi_instruction
*inst
= (glsl_to_tgsi_instruction
*)iter
.get();
2978 for (j
=0; j
< num_inst_src_regs(inst
->op
); j
++) {
2979 if (inst
->src
[j
].file
== PROGRAM_TEMPORARY
&&
2980 inst
->src
[j
].index
== index
) {
2981 last
= (depth
== 0) ? i
: -2;
2985 if (inst
->op
== TGSI_OPCODE_BGNLOOP
)
2987 else if (inst
->op
== TGSI_OPCODE_ENDLOOP
)
2988 if (--depth
== 0 && last
== -2)
3000 glsl_to_tgsi_visitor::get_last_temp_write(int index
)
3002 int depth
= 0; /* loop depth */
3003 int last
= -1; /* index of last instruction that writes to the temporary */
3006 foreach_iter(exec_list_iterator
, iter
, this->instructions
) {
3007 glsl_to_tgsi_instruction
*inst
= (glsl_to_tgsi_instruction
*)iter
.get();
3009 if (inst
->dst
.file
== PROGRAM_TEMPORARY
&& inst
->dst
.index
== index
)
3010 last
= (depth
== 0) ? i
: -2;
3012 if (inst
->op
== TGSI_OPCODE_BGNLOOP
)
3014 else if (inst
->op
== TGSI_OPCODE_ENDLOOP
)
3015 if (--depth
== 0 && last
== -2)
3027 * On a basic block basis, tracks available PROGRAM_TEMPORARY register
3028 * channels for copy propagation and updates following instructions to
3029 * use the original versions.
3031 * The glsl_to_tgsi_visitor lazily produces code assuming that this pass
3032 * will occur. As an example, a TXP production before this pass:
3034 * 0: MOV TEMP[1], INPUT[4].xyyy;
3035 * 1: MOV TEMP[1].w, INPUT[4].wwww;
3036 * 2: TXP TEMP[2], TEMP[1], texture[0], 2D;
3040 * 0: MOV TEMP[1], INPUT[4].xyyy;
3041 * 1: MOV TEMP[1].w, INPUT[4].wwww;
3042 * 2: TXP TEMP[2], INPUT[4].xyyw, texture[0], 2D;
3044 * which allows for dead code elimination on TEMP[1]'s writes.
3047 glsl_to_tgsi_visitor::copy_propagate(void)
3049 glsl_to_tgsi_instruction
**acp
= rzalloc_array(mem_ctx
,
3050 glsl_to_tgsi_instruction
*,
3051 this->next_temp
* 4);
3052 int *acp_level
= rzalloc_array(mem_ctx
, int, this->next_temp
* 4);
3055 foreach_iter(exec_list_iterator
, iter
, this->instructions
) {
3056 glsl_to_tgsi_instruction
*inst
= (glsl_to_tgsi_instruction
*)iter
.get();
3058 assert(inst
->dst
.file
!= PROGRAM_TEMPORARY
3059 || inst
->dst
.index
< this->next_temp
);
3061 /* First, do any copy propagation possible into the src regs. */
3062 for (int r
= 0; r
< 3; r
++) {
3063 glsl_to_tgsi_instruction
*first
= NULL
;
3065 int acp_base
= inst
->src
[r
].index
* 4;
3067 if (inst
->src
[r
].file
!= PROGRAM_TEMPORARY
||
3068 inst
->src
[r
].reladdr
)
3071 /* See if we can find entries in the ACP consisting of MOVs
3072 * from the same src register for all the swizzled channels
3073 * of this src register reference.
3075 for (int i
= 0; i
< 4; i
++) {
3076 int src_chan
= GET_SWZ(inst
->src
[r
].swizzle
, i
);
3077 glsl_to_tgsi_instruction
*copy_chan
= acp
[acp_base
+ src_chan
];
3084 assert(acp_level
[acp_base
+ src_chan
] <= level
);
3089 if (first
->src
[0].file
!= copy_chan
->src
[0].file
||
3090 first
->src
[0].index
!= copy_chan
->src
[0].index
) {
3098 /* We've now validated that we can copy-propagate to
3099 * replace this src register reference. Do it.
3101 inst
->src
[r
].file
= first
->src
[0].file
;
3102 inst
->src
[r
].index
= first
->src
[0].index
;
3105 for (int i
= 0; i
< 4; i
++) {
3106 int src_chan
= GET_SWZ(inst
->src
[r
].swizzle
, i
);
3107 glsl_to_tgsi_instruction
*copy_inst
= acp
[acp_base
+ src_chan
];
3108 swizzle
|= (GET_SWZ(copy_inst
->src
[0].swizzle
, src_chan
) <<
3111 inst
->src
[r
].swizzle
= swizzle
;
3116 case TGSI_OPCODE_BGNLOOP
:
3117 case TGSI_OPCODE_ENDLOOP
:
3118 /* End of a basic block, clear the ACP entirely. */
3119 memset(acp
, 0, sizeof(*acp
) * this->next_temp
* 4);
3122 case TGSI_OPCODE_IF
:
3126 case TGSI_OPCODE_ENDIF
:
3127 case TGSI_OPCODE_ELSE
:
3128 /* Clear all channels written inside the block from the ACP, but
3129 * leaving those that were not touched.
3131 for (int r
= 0; r
< this->next_temp
; r
++) {
3132 for (int c
= 0; c
< 4; c
++) {
3133 if (!acp
[4 * r
+ c
])
3136 if (acp_level
[4 * r
+ c
] >= level
)
3137 acp
[4 * r
+ c
] = NULL
;
3140 if (inst
->op
== TGSI_OPCODE_ENDIF
)
3145 /* Continuing the block, clear any written channels from
3148 if (inst
->dst
.file
== PROGRAM_TEMPORARY
&& inst
->dst
.reladdr
) {
3149 /* Any temporary might be written, so no copy propagation
3150 * across this instruction.
3152 memset(acp
, 0, sizeof(*acp
) * this->next_temp
* 4);
3153 } else if (inst
->dst
.file
== PROGRAM_OUTPUT
&&
3154 inst
->dst
.reladdr
) {
3155 /* Any output might be written, so no copy propagation
3156 * from outputs across this instruction.
3158 for (int r
= 0; r
< this->next_temp
; r
++) {
3159 for (int c
= 0; c
< 4; c
++) {
3160 if (!acp
[4 * r
+ c
])
3163 if (acp
[4 * r
+ c
]->src
[0].file
== PROGRAM_OUTPUT
)
3164 acp
[4 * r
+ c
] = NULL
;
3167 } else if (inst
->dst
.file
== PROGRAM_TEMPORARY
||
3168 inst
->dst
.file
== PROGRAM_OUTPUT
) {
3169 /* Clear where it's used as dst. */
3170 if (inst
->dst
.file
== PROGRAM_TEMPORARY
) {
3171 for (int c
= 0; c
< 4; c
++) {
3172 if (inst
->dst
.writemask
& (1 << c
)) {
3173 acp
[4 * inst
->dst
.index
+ c
] = NULL
;
3178 /* Clear where it's used as src. */
3179 for (int r
= 0; r
< this->next_temp
; r
++) {
3180 for (int c
= 0; c
< 4; c
++) {
3181 if (!acp
[4 * r
+ c
])
3184 int src_chan
= GET_SWZ(acp
[4 * r
+ c
]->src
[0].swizzle
, c
);
3186 if (acp
[4 * r
+ c
]->src
[0].file
== inst
->dst
.file
&&
3187 acp
[4 * r
+ c
]->src
[0].index
== inst
->dst
.index
&&
3188 inst
->dst
.writemask
& (1 << src_chan
))
3190 acp
[4 * r
+ c
] = NULL
;
3198 /* If this is a copy, add it to the ACP. */
3199 if (inst
->op
== TGSI_OPCODE_MOV
&&
3200 inst
->dst
.file
== PROGRAM_TEMPORARY
&&
3201 !inst
->dst
.reladdr
&&
3203 !inst
->src
[0].reladdr
&&
3204 !inst
->src
[0].negate
) {
3205 for (int i
= 0; i
< 4; i
++) {
3206 if (inst
->dst
.writemask
& (1 << i
)) {
3207 acp
[4 * inst
->dst
.index
+ i
] = inst
;
3208 acp_level
[4 * inst
->dst
.index
+ i
] = level
;
3214 ralloc_free(acp_level
);
3219 * Tracks available PROGRAM_TEMPORARY registers for dead code elimination.
3221 * The glsl_to_tgsi_visitor lazily produces code assuming that this pass
3222 * will occur. As an example, a TXP production after copy propagation but
3225 * 0: MOV TEMP[1], INPUT[4].xyyy;
3226 * 1: MOV TEMP[1].w, INPUT[4].wwww;
3227 * 2: TXP TEMP[2], INPUT[4].xyyw, texture[0], 2D;
3229 * and after this pass:
3231 * 0: TXP TEMP[2], INPUT[4].xyyw, texture[0], 2D;
3233 * FIXME: assumes that all functions are inlined (no support for BGNSUB/ENDSUB)
3234 * FIXME: doesn't eliminate all dead code inside of loops; it steps around them
3237 glsl_to_tgsi_visitor::eliminate_dead_code(void)
3241 for (i
=0; i
< this->next_temp
; i
++) {
3242 int last_read
= get_last_temp_read(i
);
3245 foreach_iter(exec_list_iterator
, iter
, this->instructions
) {
3246 glsl_to_tgsi_instruction
*inst
= (glsl_to_tgsi_instruction
*)iter
.get();
3248 if (inst
->dst
.file
== PROGRAM_TEMPORARY
&& inst
->dst
.index
== i
&&
3260 /* Merges temporary registers together where possible to reduce the number of
3261 * registers needed to run a program.
3263 * Produces optimal code only after copy propagation and dead code elimination
3266 glsl_to_tgsi_visitor::merge_registers(void)
3268 int *last_reads
= rzalloc_array(mem_ctx
, int, this->next_temp
);
3269 int *first_writes
= rzalloc_array(mem_ctx
, int, this->next_temp
);
3272 /* Read the indices of the last read and first write to each temp register
3273 * into an array so that we don't have to traverse the instruction list as
3275 for (i
=0; i
< this->next_temp
; i
++) {
3276 last_reads
[i
] = get_last_temp_read(i
);
3277 first_writes
[i
] = get_first_temp_write(i
);
3280 /* Start looking for registers with non-overlapping usages that can be
3281 * merged together. */
3282 for (i
=0; i
< this->next_temp
; i
++) {
3283 /* Don't touch unused registers. */
3284 if (last_reads
[i
] < 0 || first_writes
[i
] < 0) continue;
3286 for (j
=0; j
< this->next_temp
; j
++) {
3287 /* Don't touch unused registers. */
3288 if (last_reads
[j
] < 0 || first_writes
[j
] < 0) continue;
3290 /* We can merge the two registers if the first write to j is after or
3291 * in the same instruction as the last read from i. Note that the
3292 * register at index i will always be used earlier or at the same time
3293 * as the register at index j. */
3294 if (first_writes
[i
] <= first_writes
[j
] &&
3295 last_reads
[i
] <= first_writes
[j
])
3297 rename_temp_register(j
, i
); /* Replace all references to j with i.*/
3299 /* Update the first_writes and last_reads arrays with the new
3300 * values for the merged register index, and mark the newly unused
3301 * register index as such. */
3302 last_reads
[i
] = last_reads
[j
];
3303 first_writes
[j
] = -1;
3309 ralloc_free(last_reads
);
3310 ralloc_free(first_writes
);
3313 /* Reassign indices to temporary registers by reusing unused indices created
3314 * by optimization passes. */
3316 glsl_to_tgsi_visitor::renumber_registers(void)
3321 for (i
=0; i
< this->next_temp
; i
++) {
3322 if (get_first_temp_read(i
) < 0) continue;
3324 rename_temp_register(i
, new_index
);
3328 this->next_temp
= new_index
;
3331 /* ------------------------- TGSI conversion stuff -------------------------- */
3333 unsigned branch_target
;
3338 * Intermediate state used during shader translation.
3340 struct st_translate
{
3341 struct ureg_program
*ureg
;
3343 struct ureg_dst temps
[MAX_PROGRAM_TEMPS
];
3344 struct ureg_src
*constants
;
3345 struct ureg_dst outputs
[PIPE_MAX_SHADER_OUTPUTS
];
3346 struct ureg_src inputs
[PIPE_MAX_SHADER_INPUTS
];
3347 struct ureg_dst address
[1];
3348 struct ureg_src samplers
[PIPE_MAX_SAMPLERS
];
3349 struct ureg_src systemValues
[SYSTEM_VALUE_MAX
];
3351 /* Extra info for handling point size clamping in vertex shader */
3352 struct ureg_dst pointSizeResult
; /**< Actual point size output register */
3353 struct ureg_src pointSizeConst
; /**< Point size range constant register */
3354 GLint pointSizeOutIndex
; /**< Temp point size output register */
3355 GLboolean prevInstWrotePointSize
;
3357 const GLuint
*inputMapping
;
3358 const GLuint
*outputMapping
;
3360 /* For every instruction that contains a label (eg CALL), keep
3361 * details so that we can go back afterwards and emit the correct
3362 * tgsi instruction number for each label.
3364 struct label
*labels
;
3365 unsigned labels_size
;
3366 unsigned labels_count
;
3368 /* Keep a record of the tgsi instruction number that each mesa
3369 * instruction starts at, will be used to fix up labels after
3374 unsigned insn_count
;
3376 unsigned procType
; /**< TGSI_PROCESSOR_VERTEX/FRAGMENT */
3381 /** Map Mesa's SYSTEM_VALUE_x to TGSI_SEMANTIC_x */
3382 static unsigned mesa_sysval_to_semantic
[SYSTEM_VALUE_MAX
] = {
3384 TGSI_SEMANTIC_INSTANCEID
3388 * Make note of a branch to a label in the TGSI code.
3389 * After we've emitted all instructions, we'll go over the list
3390 * of labels built here and patch the TGSI code with the actual
3391 * location of each label.
3393 static unsigned *get_label( struct st_translate
*t
,
3394 unsigned branch_target
)
3398 if (t
->labels_count
+ 1 >= t
->labels_size
) {
3399 t
->labels_size
= 1 << (util_logbase2(t
->labels_size
) + 1);
3400 t
->labels
= (struct label
*)realloc(t
->labels
,
3401 t
->labels_size
* sizeof t
->labels
[0]);
3402 if (t
->labels
== NULL
) {
3403 static unsigned dummy
;
3409 i
= t
->labels_count
++;
3410 t
->labels
[i
].branch_target
= branch_target
;
3411 return &t
->labels
[i
].token
;
3415 * Called prior to emitting the TGSI code for each Mesa instruction.
3416 * Allocate additional space for instructions if needed.
3417 * Update the insn[] array so the next Mesa instruction points to
3418 * the next TGSI instruction.
3420 static void set_insn_start( struct st_translate
*t
,
3423 if (t
->insn_count
+ 1 >= t
->insn_size
) {
3424 t
->insn_size
= 1 << (util_logbase2(t
->insn_size
) + 1);
3425 t
->insn
= (unsigned *)realloc(t
->insn
, t
->insn_size
* sizeof t
->insn
[0]);
3426 if (t
->insn
== NULL
) {
3432 t
->insn
[t
->insn_count
++] = start
;
3436 * Map a Mesa dst register to a TGSI ureg_dst register.
3438 static struct ureg_dst
3439 dst_register( struct st_translate
*t
,
3440 gl_register_file file
,
3444 case PROGRAM_UNDEFINED
:
3445 return ureg_dst_undef();
3447 case PROGRAM_TEMPORARY
:
3448 if (ureg_dst_is_undef(t
->temps
[index
]))
3449 t
->temps
[index
] = ureg_DECL_temporary( t
->ureg
);
3451 return t
->temps
[index
];
3453 case PROGRAM_OUTPUT
:
3454 if (t
->procType
== TGSI_PROCESSOR_VERTEX
&& index
== VERT_RESULT_PSIZ
)
3455 t
->prevInstWrotePointSize
= GL_TRUE
;
3457 if (t
->procType
== TGSI_PROCESSOR_VERTEX
)
3458 assert(index
< VERT_RESULT_MAX
);
3459 else if (t
->procType
== TGSI_PROCESSOR_FRAGMENT
)
3460 assert(index
< FRAG_RESULT_MAX
);
3462 assert(index
< GEOM_RESULT_MAX
);
3464 assert(t
->outputMapping
[index
] < Elements(t
->outputs
));
3466 return t
->outputs
[t
->outputMapping
[index
]];
3468 case PROGRAM_ADDRESS
:
3469 return t
->address
[index
];
3473 return ureg_dst_undef();
3478 * Map a Mesa src register to a TGSI ureg_src register.
3480 static struct ureg_src
3481 src_register( struct st_translate
*t
,
3482 gl_register_file file
,
3486 case PROGRAM_UNDEFINED
:
3487 return ureg_src_undef();
3489 case PROGRAM_TEMPORARY
:
3491 assert(index
< Elements(t
->temps
));
3492 if (ureg_dst_is_undef(t
->temps
[index
]))
3493 t
->temps
[index
] = ureg_DECL_temporary( t
->ureg
);
3494 return ureg_src(t
->temps
[index
]);
3496 case PROGRAM_NAMED_PARAM
:
3497 case PROGRAM_ENV_PARAM
:
3498 case PROGRAM_LOCAL_PARAM
:
3499 case PROGRAM_UNIFORM
:
3501 return t
->constants
[index
];
3502 case PROGRAM_STATE_VAR
:
3503 case PROGRAM_CONSTANT
: /* ie, immediate */
3505 return ureg_DECL_constant( t
->ureg
, 0 );
3507 return t
->constants
[index
];
3510 assert(t
->inputMapping
[index
] < Elements(t
->inputs
));
3511 return t
->inputs
[t
->inputMapping
[index
]];
3513 case PROGRAM_OUTPUT
:
3514 assert(t
->outputMapping
[index
] < Elements(t
->outputs
));
3515 return ureg_src(t
->outputs
[t
->outputMapping
[index
]]); /* not needed? */
3517 case PROGRAM_ADDRESS
:
3518 return ureg_src(t
->address
[index
]);
3520 case PROGRAM_SYSTEM_VALUE
:
3521 assert(index
< Elements(t
->systemValues
));
3522 return t
->systemValues
[index
];
3526 return ureg_src_undef();
3531 * Create a TGSI ureg_dst register from an st_dst_reg.
3533 static struct ureg_dst
3534 translate_dst( struct st_translate
*t
,
3535 const st_dst_reg
*dst_reg
,
3538 struct ureg_dst dst
= dst_register( t
,
3542 dst
= ureg_writemask( dst
,
3543 dst_reg
->writemask
);
3546 dst
= ureg_saturate( dst
);
3548 if (dst_reg
->reladdr
!= NULL
)
3549 dst
= ureg_dst_indirect( dst
, ureg_src(t
->address
[0]) );
3555 * Create a TGSI ureg_src register from an st_src_reg.
3557 static struct ureg_src
3558 translate_src( struct st_translate
*t
,
3559 const st_src_reg
*src_reg
)
3561 struct ureg_src src
= src_register( t
, src_reg
->file
, src_reg
->index
);
3563 src
= ureg_swizzle( src
,
3564 GET_SWZ( src_reg
->swizzle
, 0 ) & 0x3,
3565 GET_SWZ( src_reg
->swizzle
, 1 ) & 0x3,
3566 GET_SWZ( src_reg
->swizzle
, 2 ) & 0x3,
3567 GET_SWZ( src_reg
->swizzle
, 3 ) & 0x3);
3569 if ((src_reg
->negate
& 0xf) == NEGATE_XYZW
)
3570 src
= ureg_negate(src
);
3572 if (src_reg
->reladdr
!= NULL
) {
3573 /* Normally ureg_src_indirect() would be used here, but a stupid compiler
3574 * bug in g++ makes ureg_src_indirect (an inline C function) erroneously
3575 * set the bit for src.Negate. So we have to do the operation manually
3576 * here to work around the compiler's problems. */
3577 /*src = ureg_src_indirect(src, ureg_src(t->address[0]));*/
3578 struct ureg_src addr
= ureg_src(t
->address
[0]);
3580 src
.IndirectFile
= addr
.File
;
3581 src
.IndirectIndex
= addr
.Index
;
3582 src
.IndirectSwizzle
= addr
.SwizzleX
;
3584 if (src_reg
->file
!= PROGRAM_INPUT
&&
3585 src_reg
->file
!= PROGRAM_OUTPUT
) {
3586 /* If src_reg->index was negative, it was set to zero in
3587 * src_register(). Reassign it now. But don't do this
3588 * for input/output regs since they get remapped while
3589 * const buffers don't.
3591 src
.Index
= src_reg
->index
;
3599 compile_tgsi_instruction(struct st_translate
*t
,
3600 const struct glsl_to_tgsi_instruction
*inst
)
3602 struct ureg_program
*ureg
= t
->ureg
;
3604 struct ureg_dst dst
[1];
3605 struct ureg_src src
[4];
3609 num_dst
= num_inst_dst_regs( inst
->op
);
3610 num_src
= num_inst_src_regs( inst
->op
);
3613 dst
[0] = translate_dst( t
,
3617 for (i
= 0; i
< num_src
; i
++)
3618 src
[i
] = translate_src( t
, &inst
->src
[i
] );
3620 switch( inst
->op
) {
3621 case TGSI_OPCODE_BGNLOOP
:
3622 case TGSI_OPCODE_CAL
:
3623 case TGSI_OPCODE_ELSE
:
3624 case TGSI_OPCODE_ENDLOOP
:
3625 case TGSI_OPCODE_IF
:
3626 debug_assert(num_dst
== 0);
3627 ureg_label_insn( ureg
,
3631 inst
->op
== TGSI_OPCODE_CAL
? inst
->function
->sig_id
: 0 ));
3634 case TGSI_OPCODE_TEX
:
3635 case TGSI_OPCODE_TXB
:
3636 case TGSI_OPCODE_TXD
:
3637 case TGSI_OPCODE_TXL
:
3638 case TGSI_OPCODE_TXP
:
3639 src
[num_src
++] = t
->samplers
[inst
->sampler
];
3640 ureg_tex_insn( ureg
,
3643 translate_texture_target( inst
->tex_target
,
3648 case TGSI_OPCODE_SCS
:
3649 dst
[0] = ureg_writemask(dst
[0], TGSI_WRITEMASK_XY
);
3666 * Emit the TGSI instructions to adjust the WPOS pixel center convention
3667 * Basically, add (adjX, adjY) to the fragment position.
3670 emit_adjusted_wpos( struct st_translate
*t
,
3671 const struct gl_program
*program
,
3672 GLfloat adjX
, GLfloat adjY
)
3674 struct ureg_program
*ureg
= t
->ureg
;
3675 struct ureg_dst wpos_temp
= ureg_DECL_temporary(ureg
);
3676 struct ureg_src wpos_input
= t
->inputs
[t
->inputMapping
[FRAG_ATTRIB_WPOS
]];
3678 /* Note that we bias X and Y and pass Z and W through unchanged.
3679 * The shader might also use gl_FragCoord.w and .z.
3681 ureg_ADD(ureg
, wpos_temp
, wpos_input
,
3682 ureg_imm4f(ureg
, adjX
, adjY
, 0.0f
, 0.0f
));
3684 t
->inputs
[t
->inputMapping
[FRAG_ATTRIB_WPOS
]] = ureg_src(wpos_temp
);
3689 * Emit the TGSI instructions for inverting the WPOS y coordinate.
3690 * This code is unavoidable because it also depends on whether
3691 * a FBO is bound (STATE_FB_WPOS_Y_TRANSFORM).
3694 emit_wpos_inversion( struct st_translate
*t
,
3695 const struct gl_program
*program
,
3698 struct ureg_program
*ureg
= t
->ureg
;
3700 /* Fragment program uses fragment position input.
3701 * Need to replace instances of INPUT[WPOS] with temp T
3702 * where T = INPUT[WPOS] by y is inverted.
3704 static const gl_state_index wposTransformState
[STATE_LENGTH
]
3705 = { STATE_INTERNAL
, STATE_FB_WPOS_Y_TRANSFORM
,
3706 (gl_state_index
)0, (gl_state_index
)0, (gl_state_index
)0 };
3708 /* XXX: note we are modifying the incoming shader here! Need to
3709 * do this before emitting the constant decls below, or this
3712 unsigned wposTransConst
= _mesa_add_state_reference(program
->Parameters
,
3713 wposTransformState
);
3715 struct ureg_src wpostrans
= ureg_DECL_constant( ureg
, wposTransConst
);
3716 struct ureg_dst wpos_temp
;
3717 struct ureg_src wpos_input
= t
->inputs
[t
->inputMapping
[FRAG_ATTRIB_WPOS
]];
3719 /* MOV wpos_temp, input[wpos]
3721 if (wpos_input
.File
== TGSI_FILE_TEMPORARY
)
3722 wpos_temp
= ureg_dst(wpos_input
);
3724 wpos_temp
= ureg_DECL_temporary( ureg
);
3725 ureg_MOV( ureg
, wpos_temp
, wpos_input
);
3729 /* MAD wpos_temp.y, wpos_input, wpostrans.xxxx, wpostrans.yyyy
3732 ureg_writemask(wpos_temp
, TGSI_WRITEMASK_Y
),
3734 ureg_scalar(wpostrans
, 0),
3735 ureg_scalar(wpostrans
, 1));
3737 /* MAD wpos_temp.y, wpos_input, wpostrans.zzzz, wpostrans.wwww
3740 ureg_writemask(wpos_temp
, TGSI_WRITEMASK_Y
),
3742 ureg_scalar(wpostrans
, 2),
3743 ureg_scalar(wpostrans
, 3));
3746 /* Use wpos_temp as position input from here on:
3748 t
->inputs
[t
->inputMapping
[FRAG_ATTRIB_WPOS
]] = ureg_src(wpos_temp
);
3753 * Emit fragment position/ooordinate code.
3756 emit_wpos(struct st_context
*st
,
3757 struct st_translate
*t
,
3758 const struct gl_program
*program
,
3759 struct ureg_program
*ureg
)
3761 const struct gl_fragment_program
*fp
=
3762 (const struct gl_fragment_program
*) program
;
3763 struct pipe_screen
*pscreen
= st
->pipe
->screen
;
3764 boolean invert
= FALSE
;
3766 if (fp
->OriginUpperLeft
) {
3767 /* Fragment shader wants origin in upper-left */
3768 if (pscreen
->get_param(pscreen
, PIPE_CAP_TGSI_FS_COORD_ORIGIN_UPPER_LEFT
)) {
3769 /* the driver supports upper-left origin */
3771 else if (pscreen
->get_param(pscreen
, PIPE_CAP_TGSI_FS_COORD_ORIGIN_LOWER_LEFT
)) {
3772 /* the driver supports lower-left origin, need to invert Y */
3773 ureg_property_fs_coord_origin(ureg
, TGSI_FS_COORD_ORIGIN_LOWER_LEFT
);
3780 /* Fragment shader wants origin in lower-left */
3781 if (pscreen
->get_param(pscreen
, PIPE_CAP_TGSI_FS_COORD_ORIGIN_LOWER_LEFT
))
3782 /* the driver supports lower-left origin */
3783 ureg_property_fs_coord_origin(ureg
, TGSI_FS_COORD_ORIGIN_LOWER_LEFT
);
3784 else if (pscreen
->get_param(pscreen
, PIPE_CAP_TGSI_FS_COORD_ORIGIN_UPPER_LEFT
))
3785 /* the driver supports upper-left origin, need to invert Y */
3791 if (fp
->PixelCenterInteger
) {
3792 /* Fragment shader wants pixel center integer */
3793 if (pscreen
->get_param(pscreen
, PIPE_CAP_TGSI_FS_COORD_PIXEL_CENTER_INTEGER
))
3794 /* the driver supports pixel center integer */
3795 ureg_property_fs_coord_pixel_center(ureg
, TGSI_FS_COORD_PIXEL_CENTER_INTEGER
);
3796 else if (pscreen
->get_param(pscreen
, PIPE_CAP_TGSI_FS_COORD_PIXEL_CENTER_HALF_INTEGER
))
3797 /* the driver supports pixel center half integer, need to bias X,Y */
3798 emit_adjusted_wpos(t
, program
, 0.5f
, invert
? 0.5f
: -0.5f
);
3803 /* Fragment shader wants pixel center half integer */
3804 if (pscreen
->get_param(pscreen
, PIPE_CAP_TGSI_FS_COORD_PIXEL_CENTER_HALF_INTEGER
)) {
3805 /* the driver supports pixel center half integer */
3807 else if (pscreen
->get_param(pscreen
, PIPE_CAP_TGSI_FS_COORD_PIXEL_CENTER_INTEGER
)) {
3808 /* the driver supports pixel center integer, need to bias X,Y */
3809 ureg_property_fs_coord_pixel_center(ureg
, TGSI_FS_COORD_PIXEL_CENTER_INTEGER
);
3810 emit_adjusted_wpos(t
, program
, 0.5f
, invert
? -0.5f
: 0.5f
);
3816 /* we invert after adjustment so that we avoid the MOV to temporary,
3817 * and reuse the adjustment ADD instead */
3818 emit_wpos_inversion(t
, program
, invert
);
3822 * OpenGL's fragment gl_FrontFace input is 1 for front-facing, 0 for back.
3823 * TGSI uses +1 for front, -1 for back.
3824 * This function converts the TGSI value to the GL value. Simply clamping/
3825 * saturating the value to [0,1] does the job.
3828 emit_face_var(struct st_translate
*t
)
3830 struct ureg_program
*ureg
= t
->ureg
;
3831 struct ureg_dst face_temp
= ureg_DECL_temporary(ureg
);
3832 struct ureg_src face_input
= t
->inputs
[t
->inputMapping
[FRAG_ATTRIB_FACE
]];
3834 /* MOV_SAT face_temp, input[face] */
3835 face_temp
= ureg_saturate(face_temp
);
3836 ureg_MOV(ureg
, face_temp
, face_input
);
3838 /* Use face_temp as face input from here on: */
3839 t
->inputs
[t
->inputMapping
[FRAG_ATTRIB_FACE
]] = ureg_src(face_temp
);
3843 emit_edgeflags(struct st_translate
*t
)
3845 struct ureg_program
*ureg
= t
->ureg
;
3846 struct ureg_dst edge_dst
= t
->outputs
[t
->outputMapping
[VERT_RESULT_EDGE
]];
3847 struct ureg_src edge_src
= t
->inputs
[t
->inputMapping
[VERT_ATTRIB_EDGEFLAG
]];
3849 ureg_MOV(ureg
, edge_dst
, edge_src
);
3853 * Translate intermediate IR (glsl_to_tgsi_instruction) to TGSI format.
3854 * \param program the program to translate
3855 * \param numInputs number of input registers used
3856 * \param inputMapping maps Mesa fragment program inputs to TGSI generic
3858 * \param inputSemanticName the TGSI_SEMANTIC flag for each input
3859 * \param inputSemanticIndex the semantic index (ex: which texcoord) for
3861 * \param interpMode the TGSI_INTERPOLATE_LINEAR/PERSP mode for each input
3862 * \param numOutputs number of output registers used
3863 * \param outputMapping maps Mesa fragment program outputs to TGSI
3865 * \param outputSemanticName the TGSI_SEMANTIC flag for each output
3866 * \param outputSemanticIndex the semantic index (ex: which texcoord) for
3869 * \return PIPE_OK or PIPE_ERROR_OUT_OF_MEMORY
3871 extern "C" enum pipe_error
3872 st_translate_program(
3873 struct gl_context
*ctx
,
3875 struct ureg_program
*ureg
,
3876 glsl_to_tgsi_visitor
*program
,
3877 const struct gl_program
*proginfo
,
3879 const GLuint inputMapping
[],
3880 const ubyte inputSemanticName
[],
3881 const ubyte inputSemanticIndex
[],
3882 const GLuint interpMode
[],
3884 const GLuint outputMapping
[],
3885 const ubyte outputSemanticName
[],
3886 const ubyte outputSemanticIndex
[],
3887 boolean passthrough_edgeflags
)
3889 struct st_translate translate
, *t
;
3891 enum pipe_error ret
= PIPE_OK
;
3893 assert(numInputs
<= Elements(t
->inputs
));
3894 assert(numOutputs
<= Elements(t
->outputs
));
3897 memset(t
, 0, sizeof *t
);
3899 t
->procType
= procType
;
3900 t
->inputMapping
= inputMapping
;
3901 t
->outputMapping
= outputMapping
;
3903 t
->pointSizeOutIndex
= -1;
3904 t
->prevInstWrotePointSize
= GL_FALSE
;
3907 * Declare input attributes.
3909 if (procType
== TGSI_PROCESSOR_FRAGMENT
) {
3910 for (i
= 0; i
< numInputs
; i
++) {
3911 t
->inputs
[i
] = ureg_DECL_fs_input(ureg
,
3912 inputSemanticName
[i
],
3913 inputSemanticIndex
[i
],
3917 if (proginfo
->InputsRead
& FRAG_BIT_WPOS
) {
3918 /* Must do this after setting up t->inputs, and before
3919 * emitting constant references, below:
3921 emit_wpos(st_context(ctx
), t
, proginfo
, ureg
);
3924 if (proginfo
->InputsRead
& FRAG_BIT_FACE
)
3928 * Declare output attributes.
3930 for (i
= 0; i
< numOutputs
; i
++) {
3931 switch (outputSemanticName
[i
]) {
3932 case TGSI_SEMANTIC_POSITION
:
3933 t
->outputs
[i
] = ureg_DECL_output( ureg
,
3934 TGSI_SEMANTIC_POSITION
, /* Z / Depth */
3935 outputSemanticIndex
[i
] );
3937 t
->outputs
[i
] = ureg_writemask( t
->outputs
[i
],
3940 case TGSI_SEMANTIC_STENCIL
:
3941 t
->outputs
[i
] = ureg_DECL_output( ureg
,
3942 TGSI_SEMANTIC_STENCIL
, /* Stencil */
3943 outputSemanticIndex
[i
] );
3944 t
->outputs
[i
] = ureg_writemask( t
->outputs
[i
],
3947 case TGSI_SEMANTIC_COLOR
:
3948 t
->outputs
[i
] = ureg_DECL_output( ureg
,
3949 TGSI_SEMANTIC_COLOR
,
3950 outputSemanticIndex
[i
] );
3954 return PIPE_ERROR_BAD_INPUT
;
3958 else if (procType
== TGSI_PROCESSOR_GEOMETRY
) {
3959 for (i
= 0; i
< numInputs
; i
++) {
3960 t
->inputs
[i
] = ureg_DECL_gs_input(ureg
,
3962 inputSemanticName
[i
],
3963 inputSemanticIndex
[i
]);
3966 for (i
= 0; i
< numOutputs
; i
++) {
3967 t
->outputs
[i
] = ureg_DECL_output( ureg
,
3968 outputSemanticName
[i
],
3969 outputSemanticIndex
[i
] );
3973 assert(procType
== TGSI_PROCESSOR_VERTEX
);
3975 for (i
= 0; i
< numInputs
; i
++) {
3976 t
->inputs
[i
] = ureg_DECL_vs_input(ureg
, i
);
3979 for (i
= 0; i
< numOutputs
; i
++) {
3980 t
->outputs
[i
] = ureg_DECL_output( ureg
,
3981 outputSemanticName
[i
],
3982 outputSemanticIndex
[i
] );
3983 if ((outputSemanticName
[i
] == TGSI_SEMANTIC_PSIZE
) && proginfo
->Id
) {
3984 /* Writing to the point size result register requires special
3985 * handling to implement clamping.
3987 static const gl_state_index pointSizeClampState
[STATE_LENGTH
]
3988 = { STATE_INTERNAL
, STATE_POINT_SIZE_IMPL_CLAMP
, (gl_state_index
)0, (gl_state_index
)0, (gl_state_index
)0 };
3989 /* XXX: note we are modifying the incoming shader here! Need to
3990 * do this before emitting the constant decls below, or this
3993 unsigned pointSizeClampConst
=
3994 _mesa_add_state_reference(proginfo
->Parameters
,
3995 pointSizeClampState
);
3996 struct ureg_dst psizregtemp
= ureg_DECL_temporary( ureg
);
3997 t
->pointSizeConst
= ureg_DECL_constant( ureg
, pointSizeClampConst
);
3998 t
->pointSizeResult
= t
->outputs
[i
];
3999 t
->pointSizeOutIndex
= i
;
4000 t
->outputs
[i
] = psizregtemp
;
4003 if (passthrough_edgeflags
)
4007 /* Declare address register.
4009 if (program
->num_address_regs
> 0) {
4010 debug_assert( program
->num_address_regs
== 1 );
4011 t
->address
[0] = ureg_DECL_address( ureg
);
4014 /* Declare misc input registers
4017 GLbitfield sysInputs
= proginfo
->SystemValuesRead
;
4018 unsigned numSys
= 0;
4019 for (i
= 0; sysInputs
; i
++) {
4020 if (sysInputs
& (1 << i
)) {
4021 unsigned semName
= mesa_sysval_to_semantic
[i
];
4022 t
->systemValues
[i
] = ureg_DECL_system_value(ureg
, numSys
, semName
, 0);
4024 sysInputs
&= ~(1 << i
);
4029 if (program
->indirect_addr_temps
) {
4030 /* If temps are accessed with indirect addressing, declare temporaries
4031 * in sequential order. Else, we declare them on demand elsewhere.
4032 * (Note: the number of temporaries is equal to program->next_temp)
4034 for (i
= 0; i
< (unsigned)program
->next_temp
; i
++) {
4035 /* XXX use TGSI_FILE_TEMPORARY_ARRAY when it's supported by ureg */
4036 t
->temps
[i
] = ureg_DECL_temporary( t
->ureg
);
4040 /* Emit constants and immediates. Mesa uses a single index space
4041 * for these, so we put all the translated regs in t->constants.
4042 * XXX: this entire if block depends on proginfo->Parameters from Mesa IR
4044 if (proginfo
->Parameters
) {
4045 t
->constants
= (struct ureg_src
*)CALLOC( proginfo
->Parameters
->NumParameters
* sizeof t
->constants
[0] );
4046 if (t
->constants
== NULL
) {
4047 ret
= PIPE_ERROR_OUT_OF_MEMORY
;
4051 for (i
= 0; i
< proginfo
->Parameters
->NumParameters
; i
++) {
4052 switch (proginfo
->Parameters
->Parameters
[i
].Type
) {
4053 case PROGRAM_ENV_PARAM
:
4054 case PROGRAM_LOCAL_PARAM
:
4055 case PROGRAM_STATE_VAR
:
4056 case PROGRAM_NAMED_PARAM
:
4057 case PROGRAM_UNIFORM
:
4058 t
->constants
[i
] = ureg_DECL_constant( ureg
, i
);
4061 /* Emit immediates only when there's no indirect addressing of
4063 * FIXME: Be smarter and recognize param arrays:
4064 * indirect addressing is only valid within the referenced
4067 case PROGRAM_CONSTANT
:
4068 if (program
->indirect_addr_consts
)
4069 t
->constants
[i
] = ureg_DECL_constant( ureg
, i
);
4071 switch(proginfo
->Parameters
->Parameters
[i
].DataType
)
4077 t
->constants
[i
] = ureg_DECL_immediate(ureg
, (float *)proginfo
->Parameters
->ParameterValues
[i
], 4);
4083 t
->constants
[i
] = ureg_DECL_immediate_int(ureg
, (int *)proginfo
->Parameters
->ParameterValues
[i
], 4);
4085 case GL_UNSIGNED_INT
:
4086 case GL_UNSIGNED_INT_VEC2
:
4087 case GL_UNSIGNED_INT_VEC3
:
4088 case GL_UNSIGNED_INT_VEC4
:
4093 t
->constants
[i
] = ureg_DECL_immediate_uint(ureg
, (unsigned *)proginfo
->Parameters
->ParameterValues
[i
], 4);
4096 assert(!"should not get here");
4105 /* texture samplers */
4106 for (i
= 0; i
< ctx
->Const
.MaxTextureImageUnits
; i
++) {
4107 if (program
->samplers_used
& (1 << i
)) {
4108 t
->samplers
[i
] = ureg_DECL_sampler( ureg
, i
);
4112 /* Emit each instruction in turn:
4114 foreach_iter(exec_list_iterator
, iter
, program
->instructions
) {
4115 set_insn_start( t
, ureg_get_instruction_number( ureg
));
4116 compile_tgsi_instruction( t
, (glsl_to_tgsi_instruction
*)iter
.get() );
4118 if (t
->prevInstWrotePointSize
&& proginfo
->Id
) {
4119 /* The previous instruction wrote to the (fake) vertex point size
4120 * result register. Now we need to clamp that value to the min/max
4121 * point size range, putting the result into the real point size
4123 * Note that we can't do this easily at the end of program due to
4124 * possible early return.
4126 set_insn_start( t
, ureg_get_instruction_number( ureg
));
4128 ureg_writemask(t
->outputs
[t
->pointSizeOutIndex
], WRITEMASK_X
),
4129 ureg_src(t
->outputs
[t
->pointSizeOutIndex
]),
4130 ureg_swizzle(t
->pointSizeConst
, 1,1,1,1));
4131 ureg_MIN( t
->ureg
, ureg_writemask(t
->pointSizeResult
, WRITEMASK_X
),
4132 ureg_src(t
->outputs
[t
->pointSizeOutIndex
]),
4133 ureg_swizzle(t
->pointSizeConst
, 2,2,2,2));
4135 t
->prevInstWrotePointSize
= GL_FALSE
;
4138 /* Fix up all emitted labels:
4140 for (i
= 0; i
< t
->labels_count
; i
++) {
4141 ureg_fixup_label( ureg
,
4143 t
->insn
[t
->labels
[i
].branch_target
] );
4152 debug_printf("%s: translate error flag set\n", __FUNCTION__
);
4157 /* ----------------------------- End TGSI code ------------------------------ */
4160 * Convert a shader's GLSL IR into a Mesa gl_program, although without
4161 * generating Mesa IR.
4163 static struct gl_program
*
4164 get_mesa_program(struct gl_context
*ctx
,
4165 struct gl_shader_program
*shader_program
,
4166 struct gl_shader
*shader
)
4168 glsl_to_tgsi_visitor
* v
= new glsl_to_tgsi_visitor();
4169 struct gl_program
*prog
;
4171 const char *target_string
;
4173 struct gl_shader_compiler_options
*options
=
4174 &ctx
->ShaderCompilerOptions
[_mesa_shader_type_to_index(shader
->Type
)];
4176 switch (shader
->Type
) {
4177 case GL_VERTEX_SHADER
:
4178 target
= GL_VERTEX_PROGRAM_ARB
;
4179 target_string
= "vertex";
4181 case GL_FRAGMENT_SHADER
:
4182 target
= GL_FRAGMENT_PROGRAM_ARB
;
4183 target_string
= "fragment";
4185 case GL_GEOMETRY_SHADER
:
4186 target
= GL_GEOMETRY_PROGRAM_NV
;
4187 target_string
= "geometry";
4190 assert(!"should not be reached");
4194 validate_ir_tree(shader
->ir
);
4196 prog
= ctx
->Driver
.NewProgram(ctx
, target
, shader_program
->Name
);
4199 prog
->Parameters
= _mesa_new_parameter_list();
4200 prog
->Varying
= _mesa_new_parameter_list();
4201 prog
->Attributes
= _mesa_new_parameter_list();
4204 v
->shader_program
= shader_program
;
4205 v
->options
= options
;
4206 v
->glsl_version
= ctx
->Const
.GLSLVersion
;
4208 add_uniforms_to_parameters_list(shader_program
, shader
, prog
);
4210 /* Emit intermediate IR for main(). */
4211 visit_exec_list(shader
->ir
, v
);
4213 /* Now emit bodies for any functions that were used. */
4215 progress
= GL_FALSE
;
4217 foreach_iter(exec_list_iterator
, iter
, v
->function_signatures
) {
4218 function_entry
*entry
= (function_entry
*)iter
.get();
4220 if (!entry
->bgn_inst
) {
4221 v
->current_function
= entry
;
4223 entry
->bgn_inst
= v
->emit(NULL
, TGSI_OPCODE_BGNSUB
);
4224 entry
->bgn_inst
->function
= entry
;
4226 visit_exec_list(&entry
->sig
->body
, v
);
4228 glsl_to_tgsi_instruction
*last
;
4229 last
= (glsl_to_tgsi_instruction
*)v
->instructions
.get_tail();
4230 if (last
->op
!= TGSI_OPCODE_RET
)
4231 v
->emit(NULL
, TGSI_OPCODE_RET
);
4233 glsl_to_tgsi_instruction
*end
;
4234 end
= v
->emit(NULL
, TGSI_OPCODE_ENDSUB
);
4235 end
->function
= entry
;
4243 /* Print out some information (for debugging purposes) used by the
4244 * optimization passes. */
4245 for (i
=0; i
< v
->next_temp
; i
++) {
4246 int fr
= v
->get_first_temp_read(i
);
4247 int fw
= v
->get_first_temp_write(i
);
4248 int lr
= v
->get_last_temp_read(i
);
4249 int lw
= v
->get_last_temp_write(i
);
4251 printf("Temp %d: FR=%3d FW=%3d LR=%3d LW=%3d\n", i
, fr
, fw
, lr
, lw
);
4256 /* Remove reads to output registers, and to varyings in vertex shaders. */
4257 v
->remove_output_reads(PROGRAM_OUTPUT
);
4258 if (target
== GL_VERTEX_PROGRAM_ARB
)
4259 v
->remove_output_reads(PROGRAM_VARYING
);
4261 /* Perform the simplify_cmp optimization, which is required by r300g. */
4264 /* Perform optimizations on the instructions in the glsl_to_tgsi_visitor.
4265 * FIXME: These passes to optimize temporary registers don't work when there
4266 * is indirect addressing of the temporary register space. We need proper
4267 * array support so that we don't have to give up these passes in every
4268 * shader that uses arrays.
4270 if (!v
->indirect_addr_temps
) {
4271 v
->copy_propagate();
4272 v
->eliminate_dead_code();
4273 v
->merge_registers();
4274 v
->renumber_registers();
4277 /* Write the END instruction. */
4278 v
->emit(NULL
, TGSI_OPCODE_END
);
4280 if (ctx
->Shader
.Flags
& GLSL_DUMP
) {
4282 printf("GLSL IR for linked %s program %d:\n", target_string
,
4283 shader_program
->Name
);
4284 _mesa_print_ir(shader
->ir
, NULL
);
4289 prog
->Instructions
= NULL
;
4290 prog
->NumInstructions
= 0;
4292 do_set_program_inouts(shader
->ir
, prog
);
4293 count_resources(v
, prog
);
4295 check_resources(ctx
, shader_program
, v
, prog
);
4297 _mesa_reference_program(ctx
, &shader
->Program
, prog
);
4299 struct st_vertex_program
*stvp
;
4300 struct st_fragment_program
*stfp
;
4301 struct st_geometry_program
*stgp
;
4303 switch (shader
->Type
) {
4304 case GL_VERTEX_SHADER
:
4305 stvp
= (struct st_vertex_program
*)prog
;
4306 stvp
->glsl_to_tgsi
= v
;
4308 case GL_FRAGMENT_SHADER
:
4309 stfp
= (struct st_fragment_program
*)prog
;
4310 stfp
->glsl_to_tgsi
= v
;
4312 case GL_GEOMETRY_SHADER
:
4313 stgp
= (struct st_geometry_program
*)prog
;
4314 stgp
->glsl_to_tgsi
= v
;
4317 assert(!"should not be reached");
4327 st_new_shader(struct gl_context
*ctx
, GLuint name
, GLuint type
)
4329 struct gl_shader
*shader
;
4330 assert(type
== GL_FRAGMENT_SHADER
|| type
== GL_VERTEX_SHADER
||
4331 type
== GL_GEOMETRY_SHADER_ARB
);
4332 shader
= rzalloc(NULL
, struct gl_shader
);
4334 shader
->Type
= type
;
4335 shader
->Name
= name
;
4336 _mesa_init_shader(ctx
, shader
);
4341 struct gl_shader_program
*
4342 st_new_shader_program(struct gl_context
*ctx
, GLuint name
)
4344 struct gl_shader_program
*shProg
;
4345 shProg
= rzalloc(NULL
, struct gl_shader_program
);
4347 shProg
->Name
= name
;
4348 _mesa_init_shader_program(ctx
, shProg
);
4355 * Called via ctx->Driver.LinkShader()
4356 * This actually involves converting GLSL IR into an intermediate TGSI-like IR
4357 * with code lowering and other optimizations.
4360 st_link_shader(struct gl_context
*ctx
, struct gl_shader_program
*prog
)
4362 assert(prog
->LinkStatus
);
4364 for (unsigned i
= 0; i
< MESA_SHADER_TYPES
; i
++) {
4365 if (prog
->_LinkedShaders
[i
] == NULL
)
4369 exec_list
*ir
= prog
->_LinkedShaders
[i
]->ir
;
4370 const struct gl_shader_compiler_options
*options
=
4371 &ctx
->ShaderCompilerOptions
[_mesa_shader_type_to_index(prog
->_LinkedShaders
[i
]->Type
)];
4377 do_mat_op_to_vec(ir
);
4378 lower_instructions(ir
, (MOD_TO_FRACT
| DIV_TO_MUL_RCP
| EXP_TO_EXP2
4380 | ((options
->EmitNoPow
) ? POW_TO_EXP2
: 0)));
4382 progress
= do_lower_jumps(ir
, true, true, options
->EmitNoMainReturn
, options
->EmitNoCont
, options
->EmitNoLoops
) || progress
;
4384 progress
= do_common_optimization(ir
, true, options
->MaxUnrollIterations
) || progress
;
4386 progress
= lower_quadop_vector(ir
, true) || progress
;
4388 if (options
->EmitNoIfs
) {
4389 progress
= lower_discard(ir
) || progress
;
4390 progress
= lower_if_to_cond_assign(ir
) || progress
;
4393 if (options
->EmitNoNoise
)
4394 progress
= lower_noise(ir
) || progress
;
4396 /* If there are forms of indirect addressing that the driver
4397 * cannot handle, perform the lowering pass.
4399 if (options
->EmitNoIndirectInput
|| options
->EmitNoIndirectOutput
4400 || options
->EmitNoIndirectTemp
|| options
->EmitNoIndirectUniform
)
4402 lower_variable_index_to_cond_assign(ir
,
4403 options
->EmitNoIndirectInput
,
4404 options
->EmitNoIndirectOutput
,
4405 options
->EmitNoIndirectTemp
,
4406 options
->EmitNoIndirectUniform
)
4409 progress
= do_vec_index_to_cond_assign(ir
) || progress
;
4412 validate_ir_tree(ir
);
4415 for (unsigned i
= 0; i
< MESA_SHADER_TYPES
; i
++) {
4416 struct gl_program
*linked_prog
;
4418 if (prog
->_LinkedShaders
[i
] == NULL
)
4421 linked_prog
= get_mesa_program(ctx
, prog
, prog
->_LinkedShaders
[i
]);
4426 switch (prog
->_LinkedShaders
[i
]->Type
) {
4427 case GL_VERTEX_SHADER
:
4428 _mesa_reference_vertprog(ctx
, &prog
->VertexProgram
,
4429 (struct gl_vertex_program
*)linked_prog
);
4430 ok
= ctx
->Driver
.ProgramStringNotify(ctx
, GL_VERTEX_PROGRAM_ARB
,
4433 case GL_FRAGMENT_SHADER
:
4434 _mesa_reference_fragprog(ctx
, &prog
->FragmentProgram
,
4435 (struct gl_fragment_program
*)linked_prog
);
4436 ok
= ctx
->Driver
.ProgramStringNotify(ctx
, GL_FRAGMENT_PROGRAM_ARB
,
4439 case GL_GEOMETRY_SHADER
:
4440 _mesa_reference_geomprog(ctx
, &prog
->GeometryProgram
,
4441 (struct gl_geometry_program
*)linked_prog
);
4442 ok
= ctx
->Driver
.ProgramStringNotify(ctx
, GL_GEOMETRY_PROGRAM_NV
,
4451 _mesa_reference_program(ctx
, &linked_prog
, NULL
);
4459 * Link a GLSL shader program. Called via glLinkProgram().
4462 st_glsl_link_shader(struct gl_context
*ctx
, struct gl_shader_program
*prog
)
4466 _mesa_clear_shader_program_data(ctx
, prog
);
4468 prog
->LinkStatus
= GL_TRUE
;
4470 for (i
= 0; i
< prog
->NumShaders
; i
++) {
4471 if (!prog
->Shaders
[i
]->CompileStatus
) {
4472 fail_link(prog
, "linking with uncompiled shader");
4473 prog
->LinkStatus
= GL_FALSE
;
4477 prog
->Varying
= _mesa_new_parameter_list();
4478 _mesa_reference_vertprog(ctx
, &prog
->VertexProgram
, NULL
);
4479 _mesa_reference_fragprog(ctx
, &prog
->FragmentProgram
, NULL
);
4480 _mesa_reference_geomprog(ctx
, &prog
->GeometryProgram
, NULL
);
4482 if (prog
->LinkStatus
) {
4483 link_shaders(ctx
, prog
);
4486 if (prog
->LinkStatus
) {
4487 if (!ctx
->Driver
.LinkShader(ctx
, prog
)) {
4488 prog
->LinkStatus
= GL_FALSE
;
4492 set_uniform_initializers(ctx
, prog
);
4494 if (ctx
->Shader
.Flags
& GLSL_DUMP
) {
4495 if (!prog
->LinkStatus
) {
4496 printf("GLSL shader program %d failed to link\n", prog
->Name
);
4499 if (prog
->InfoLog
&& prog
->InfoLog
[0] != 0) {
4500 printf("GLSL shader program %d info log:\n", prog
->Name
);
4501 printf("%s\n", prog
->InfoLog
);