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
);
378 void rename_temp_register(int index
, int new_index
);
379 int get_first_temp_read(int index
);
380 int get_first_temp_write(int index
);
381 int get_last_temp_read(int index
);
382 int get_last_temp_write(int index
);
384 void copy_propagate(void);
385 void eliminate_dead_code(void);
386 void merge_registers(void);
387 void renumber_registers(void);
392 static st_src_reg undef_src
= st_src_reg(PROGRAM_UNDEFINED
, 0, NULL
);
394 static st_dst_reg undef_dst
= st_dst_reg(PROGRAM_UNDEFINED
, SWIZZLE_NOOP
, GLSL_TYPE_ERROR
);
396 static st_dst_reg address_reg
= st_dst_reg(PROGRAM_ADDRESS
, WRITEMASK_X
, GLSL_TYPE_FLOAT
);
399 fail_link(struct gl_shader_program
*prog
, const char *fmt
, ...) PRINTFLIKE(2, 3);
402 fail_link(struct gl_shader_program
*prog
, const char *fmt
, ...)
406 ralloc_vasprintf_append(&prog
->InfoLog
, fmt
, args
);
409 prog
->LinkStatus
= GL_FALSE
;
413 swizzle_for_size(int size
)
415 int size_swizzles
[4] = {
416 MAKE_SWIZZLE4(SWIZZLE_X
, SWIZZLE_X
, SWIZZLE_X
, SWIZZLE_X
),
417 MAKE_SWIZZLE4(SWIZZLE_X
, SWIZZLE_Y
, SWIZZLE_Y
, SWIZZLE_Y
),
418 MAKE_SWIZZLE4(SWIZZLE_X
, SWIZZLE_Y
, SWIZZLE_Z
, SWIZZLE_Z
),
419 MAKE_SWIZZLE4(SWIZZLE_X
, SWIZZLE_Y
, SWIZZLE_Z
, SWIZZLE_W
),
422 assert((size
>= 1) && (size
<= 4));
423 return size_swizzles
[size
- 1];
427 is_tex_instruction(unsigned opcode
)
429 const tgsi_opcode_info
* info
= tgsi_get_opcode_info(opcode
);
434 num_inst_dst_regs(unsigned opcode
)
436 const tgsi_opcode_info
* info
= tgsi_get_opcode_info(opcode
);
437 return info
->num_dst
;
441 num_inst_src_regs(unsigned opcode
)
443 const tgsi_opcode_info
* info
= tgsi_get_opcode_info(opcode
);
444 return info
->is_tex
? info
->num_src
- 1 : info
->num_src
;
447 glsl_to_tgsi_instruction
*
448 glsl_to_tgsi_visitor::emit(ir_instruction
*ir
, unsigned op
,
450 st_src_reg src0
, st_src_reg src1
, st_src_reg src2
)
452 glsl_to_tgsi_instruction
*inst
= new(mem_ctx
) glsl_to_tgsi_instruction();
453 int num_reladdr
= 0, i
;
455 op
= get_opcode(ir
, op
, dst
, src0
, src1
);
457 /* If we have to do relative addressing, we want to load the ARL
458 * reg directly for one of the regs, and preload the other reladdr
459 * sources into temps.
461 num_reladdr
+= dst
.reladdr
!= NULL
;
462 num_reladdr
+= src0
.reladdr
!= NULL
;
463 num_reladdr
+= src1
.reladdr
!= NULL
;
464 num_reladdr
+= src2
.reladdr
!= NULL
;
466 reladdr_to_temp(ir
, &src2
, &num_reladdr
);
467 reladdr_to_temp(ir
, &src1
, &num_reladdr
);
468 reladdr_to_temp(ir
, &src0
, &num_reladdr
);
471 emit_arl(ir
, address_reg
, *dst
.reladdr
);
474 assert(num_reladdr
== 0);
483 inst
->function
= NULL
;
485 if (op
== TGSI_OPCODE_ARL
)
486 this->num_address_regs
= 1;
488 /* Update indirect addressing status used by TGSI */
491 case PROGRAM_TEMPORARY
:
492 this->indirect_addr_temps
= true;
494 case PROGRAM_LOCAL_PARAM
:
495 case PROGRAM_ENV_PARAM
:
496 case PROGRAM_STATE_VAR
:
497 case PROGRAM_NAMED_PARAM
:
498 case PROGRAM_CONSTANT
:
499 case PROGRAM_UNIFORM
:
500 this->indirect_addr_consts
= true;
507 for (i
=0; i
<3; i
++) {
508 if(inst
->src
[i
].reladdr
) {
509 switch(inst
->src
[i
].file
) {
510 case PROGRAM_TEMPORARY
:
511 this->indirect_addr_temps
= true;
513 case PROGRAM_LOCAL_PARAM
:
514 case PROGRAM_ENV_PARAM
:
515 case PROGRAM_STATE_VAR
:
516 case PROGRAM_NAMED_PARAM
:
517 case PROGRAM_CONSTANT
:
518 case PROGRAM_UNIFORM
:
519 this->indirect_addr_consts
= true;
528 this->instructions
.push_tail(inst
);
534 glsl_to_tgsi_instruction
*
535 glsl_to_tgsi_visitor::emit(ir_instruction
*ir
, unsigned op
,
536 st_dst_reg dst
, st_src_reg src0
, st_src_reg src1
)
538 return emit(ir
, op
, dst
, src0
, src1
, undef_src
);
541 glsl_to_tgsi_instruction
*
542 glsl_to_tgsi_visitor::emit(ir_instruction
*ir
, unsigned op
,
543 st_dst_reg dst
, st_src_reg src0
)
545 assert(dst
.writemask
!= 0);
546 return emit(ir
, op
, dst
, src0
, undef_src
, undef_src
);
549 glsl_to_tgsi_instruction
*
550 glsl_to_tgsi_visitor::emit(ir_instruction
*ir
, unsigned op
)
552 return emit(ir
, op
, undef_dst
, undef_src
, undef_src
, undef_src
);
556 * Determines whether to use an integer, unsigned integer, or float opcode
557 * based on the operands and input opcode, then emits the result.
559 * TODO: type checking for remaining TGSI opcodes
562 glsl_to_tgsi_visitor::get_opcode(ir_instruction
*ir
, unsigned op
,
564 st_src_reg src0
, st_src_reg src1
)
566 int type
= GLSL_TYPE_FLOAT
;
568 if (src0
.type
== GLSL_TYPE_FLOAT
|| src1
.type
== GLSL_TYPE_FLOAT
)
569 type
= GLSL_TYPE_FLOAT
;
570 else if (glsl_version
>= 130)
573 #define case4(c, f, i, u) \
574 case TGSI_OPCODE_##c: \
575 if (type == GLSL_TYPE_INT) op = TGSI_OPCODE_##i; \
576 else if (type == GLSL_TYPE_UINT) op = TGSI_OPCODE_##u; \
577 else op = TGSI_OPCODE_##f; \
579 #define case3(f, i, u) case4(f, f, i, u)
580 #define case2fi(f, i) case4(f, f, i, i)
581 #define case2iu(i, u) case4(i, LAST, i, u)
587 case3(DIV
, IDIV
, UDIV
);
588 case3(MAX
, IMAX
, UMAX
);
589 case3(MIN
, IMIN
, UMIN
);
594 case3(SGE
, ISGE
, USGE
);
595 case3(SLT
, ISLT
, USLT
);
607 assert(op
!= TGSI_OPCODE_LAST
);
612 glsl_to_tgsi_visitor::emit_dp(ir_instruction
*ir
,
613 st_dst_reg dst
, st_src_reg src0
, st_src_reg src1
,
616 static const unsigned dot_opcodes
[] = {
617 TGSI_OPCODE_DP2
, TGSI_OPCODE_DP3
, TGSI_OPCODE_DP4
620 emit(ir
, dot_opcodes
[elements
- 2], dst
, src0
, src1
);
624 * Emits TGSI scalar opcodes to produce unique answers across channels.
626 * Some TGSI opcodes are scalar-only, like ARB_fp/vp. The src X
627 * channel determines the result across all channels. So to do a vec4
628 * of this operation, we want to emit a scalar per source channel used
629 * to produce dest channels.
632 glsl_to_tgsi_visitor::emit_scalar(ir_instruction
*ir
, unsigned op
,
634 st_src_reg orig_src0
, st_src_reg orig_src1
)
637 int done_mask
= ~dst
.writemask
;
639 /* TGSI RCP is a scalar operation splatting results to all channels,
640 * like ARB_fp/vp. So emit as many RCPs as necessary to cover our
643 for (i
= 0; i
< 4; i
++) {
644 GLuint this_mask
= (1 << i
);
645 glsl_to_tgsi_instruction
*inst
;
646 st_src_reg src0
= orig_src0
;
647 st_src_reg src1
= orig_src1
;
649 if (done_mask
& this_mask
)
652 GLuint src0_swiz
= GET_SWZ(src0
.swizzle
, i
);
653 GLuint src1_swiz
= GET_SWZ(src1
.swizzle
, i
);
654 for (j
= i
+ 1; j
< 4; j
++) {
655 /* If there is another enabled component in the destination that is
656 * derived from the same inputs, generate its value on this pass as
659 if (!(done_mask
& (1 << j
)) &&
660 GET_SWZ(src0
.swizzle
, j
) == src0_swiz
&&
661 GET_SWZ(src1
.swizzle
, j
) == src1_swiz
) {
662 this_mask
|= (1 << j
);
665 src0
.swizzle
= MAKE_SWIZZLE4(src0_swiz
, src0_swiz
,
666 src0_swiz
, src0_swiz
);
667 src1
.swizzle
= MAKE_SWIZZLE4(src1_swiz
, src1_swiz
,
668 src1_swiz
, src1_swiz
);
670 inst
= emit(ir
, op
, dst
, src0
, src1
);
671 inst
->dst
.writemask
= this_mask
;
672 done_mask
|= this_mask
;
677 glsl_to_tgsi_visitor::emit_scalar(ir_instruction
*ir
, unsigned op
,
678 st_dst_reg dst
, st_src_reg src0
)
680 st_src_reg undef
= undef_src
;
682 undef
.swizzle
= SWIZZLE_XXXX
;
684 emit_scalar(ir
, op
, dst
, src0
, undef
);
688 glsl_to_tgsi_visitor::emit_arl(ir_instruction
*ir
,
689 st_dst_reg dst
, st_src_reg src0
)
691 st_src_reg tmp
= get_temp(glsl_type::float_type
);
693 if (src0
.type
== GLSL_TYPE_INT
)
694 emit(ir
, TGSI_OPCODE_I2F
, st_dst_reg(tmp
), src0
);
695 else if (src0
.type
== GLSL_TYPE_UINT
)
696 emit(ir
, TGSI_OPCODE_U2F
, st_dst_reg(tmp
), src0
);
700 emit(ir
, TGSI_OPCODE_ARL
, dst
, tmp
);
704 * Emit an TGSI_OPCODE_SCS instruction
706 * The \c SCS opcode functions a bit differently than the other TGSI opcodes.
707 * Instead of splatting its result across all four components of the
708 * destination, it writes one value to the \c x component and another value to
709 * the \c y component.
711 * \param ir IR instruction being processed
712 * \param op Either \c TGSI_OPCODE_SIN or \c TGSI_OPCODE_COS depending
713 * on which value is desired.
714 * \param dst Destination register
715 * \param src Source register
718 glsl_to_tgsi_visitor::emit_scs(ir_instruction
*ir
, unsigned op
,
720 const st_src_reg
&src
)
722 /* Vertex programs cannot use the SCS opcode.
724 if (this->prog
->Target
== GL_VERTEX_PROGRAM_ARB
) {
725 emit_scalar(ir
, op
, dst
, src
);
729 const unsigned component
= (op
== TGSI_OPCODE_SIN
) ? 0 : 1;
730 const unsigned scs_mask
= (1U << component
);
731 int done_mask
= ~dst
.writemask
;
734 assert(op
== TGSI_OPCODE_SIN
|| op
== TGSI_OPCODE_COS
);
736 /* If there are compnents in the destination that differ from the component
737 * that will be written by the SCS instrution, we'll need a temporary.
739 if (scs_mask
!= unsigned(dst
.writemask
)) {
740 tmp
= get_temp(glsl_type::vec4_type
);
743 for (unsigned i
= 0; i
< 4; i
++) {
744 unsigned this_mask
= (1U << i
);
745 st_src_reg src0
= src
;
747 if ((done_mask
& this_mask
) != 0)
750 /* The source swizzle specified which component of the source generates
751 * sine / cosine for the current component in the destination. The SCS
752 * instruction requires that this value be swizzle to the X component.
753 * Replace the current swizzle with a swizzle that puts the source in
756 unsigned src0_swiz
= GET_SWZ(src
.swizzle
, i
);
758 src0
.swizzle
= MAKE_SWIZZLE4(src0_swiz
, src0_swiz
,
759 src0_swiz
, src0_swiz
);
760 for (unsigned j
= i
+ 1; j
< 4; j
++) {
761 /* If there is another enabled component in the destination that is
762 * derived from the same inputs, generate its value on this pass as
765 if (!(done_mask
& (1 << j
)) &&
766 GET_SWZ(src0
.swizzle
, j
) == src0_swiz
) {
767 this_mask
|= (1 << j
);
771 if (this_mask
!= scs_mask
) {
772 glsl_to_tgsi_instruction
*inst
;
773 st_dst_reg tmp_dst
= st_dst_reg(tmp
);
775 /* Emit the SCS instruction.
777 inst
= emit(ir
, TGSI_OPCODE_SCS
, tmp_dst
, src0
);
778 inst
->dst
.writemask
= scs_mask
;
780 /* Move the result of the SCS instruction to the desired location in
783 tmp
.swizzle
= MAKE_SWIZZLE4(component
, component
,
784 component
, component
);
785 inst
= emit(ir
, TGSI_OPCODE_SCS
, dst
, tmp
);
786 inst
->dst
.writemask
= this_mask
;
788 /* Emit the SCS instruction to write directly to the destination.
790 glsl_to_tgsi_instruction
*inst
= emit(ir
, TGSI_OPCODE_SCS
, dst
, src0
);
791 inst
->dst
.writemask
= scs_mask
;
794 done_mask
|= this_mask
;
799 glsl_to_tgsi_visitor::st_src_reg_for_float(float val
)
801 st_src_reg
src(PROGRAM_CONSTANT
, -1, GLSL_TYPE_FLOAT
);
802 union gl_constant_value uval
;
805 src
.index
= _mesa_add_typed_unnamed_constant(this->prog
->Parameters
,
806 &uval
, 1, GL_FLOAT
, &src
.swizzle
);
812 glsl_to_tgsi_visitor::st_src_reg_for_int(int val
)
814 st_src_reg
src(PROGRAM_CONSTANT
, -1, GLSL_TYPE_INT
);
815 union gl_constant_value uval
;
817 assert(glsl_version
>= 130);
820 src
.index
= _mesa_add_typed_unnamed_constant(this->prog
->Parameters
,
821 &uval
, 1, GL_INT
, &src
.swizzle
);
827 glsl_to_tgsi_visitor::st_src_reg_for_type(int type
, int val
)
829 if (glsl_version
>= 130)
830 return type
== GLSL_TYPE_FLOAT
? st_src_reg_for_float(val
) :
831 st_src_reg_for_int(val
);
833 return st_src_reg_for_float(val
);
837 type_size(const struct glsl_type
*type
)
842 switch (type
->base_type
) {
845 case GLSL_TYPE_FLOAT
:
847 if (type
->is_matrix()) {
848 return type
->matrix_columns
;
850 /* Regardless of size of vector, it gets a vec4. This is bad
851 * packing for things like floats, but otherwise arrays become a
852 * mess. Hopefully a later pass over the code can pack scalars
853 * down if appropriate.
857 case GLSL_TYPE_ARRAY
:
858 assert(type
->length
> 0);
859 return type_size(type
->fields
.array
) * type
->length
;
860 case GLSL_TYPE_STRUCT
:
862 for (i
= 0; i
< type
->length
; i
++) {
863 size
+= type_size(type
->fields
.structure
[i
].type
);
866 case GLSL_TYPE_SAMPLER
:
867 /* Samplers take up one slot in UNIFORMS[], but they're baked in
878 * In the initial pass of codegen, we assign temporary numbers to
879 * intermediate results. (not SSA -- variable assignments will reuse
883 glsl_to_tgsi_visitor::get_temp(const glsl_type
*type
)
889 src
.type
= type
->base_type
;
890 src
.file
= PROGRAM_TEMPORARY
;
891 src
.index
= next_temp
;
893 next_temp
+= type_size(type
);
895 if (type
->is_array() || type
->is_record()) {
896 src
.swizzle
= SWIZZLE_NOOP
;
898 for (i
= 0; i
< type
->vector_elements
; i
++)
901 swizzle
[i
] = type
->vector_elements
- 1;
902 src
.swizzle
= MAKE_SWIZZLE4(swizzle
[0], swizzle
[1],
903 swizzle
[2], swizzle
[3]);
911 glsl_to_tgsi_visitor::find_variable_storage(ir_variable
*var
)
914 variable_storage
*entry
;
916 foreach_iter(exec_list_iterator
, iter
, this->variables
) {
917 entry
= (variable_storage
*)iter
.get();
919 if (entry
->var
== var
)
927 glsl_to_tgsi_visitor::visit(ir_variable
*ir
)
929 if (strcmp(ir
->name
, "gl_FragCoord") == 0) {
930 struct gl_fragment_program
*fp
= (struct gl_fragment_program
*)this->prog
;
932 fp
->OriginUpperLeft
= ir
->origin_upper_left
;
933 fp
->PixelCenterInteger
= ir
->pixel_center_integer
;
935 } else if (strcmp(ir
->name
, "gl_FragDepth") == 0) {
936 struct gl_fragment_program
*fp
= (struct gl_fragment_program
*)this->prog
;
937 switch (ir
->depth_layout
) {
938 case ir_depth_layout_none
:
939 fp
->FragDepthLayout
= FRAG_DEPTH_LAYOUT_NONE
;
941 case ir_depth_layout_any
:
942 fp
->FragDepthLayout
= FRAG_DEPTH_LAYOUT_ANY
;
944 case ir_depth_layout_greater
:
945 fp
->FragDepthLayout
= FRAG_DEPTH_LAYOUT_GREATER
;
947 case ir_depth_layout_less
:
948 fp
->FragDepthLayout
= FRAG_DEPTH_LAYOUT_LESS
;
950 case ir_depth_layout_unchanged
:
951 fp
->FragDepthLayout
= FRAG_DEPTH_LAYOUT_UNCHANGED
;
959 if (ir
->mode
== ir_var_uniform
&& strncmp(ir
->name
, "gl_", 3) == 0) {
961 const ir_state_slot
*const slots
= ir
->state_slots
;
962 assert(ir
->state_slots
!= NULL
);
964 /* Check if this statevar's setup in the STATE file exactly
965 * matches how we'll want to reference it as a
966 * struct/array/whatever. If not, then we need to move it into
967 * temporary storage and hope that it'll get copy-propagated
970 for (i
= 0; i
< ir
->num_state_slots
; i
++) {
971 if (slots
[i
].swizzle
!= SWIZZLE_XYZW
) {
976 struct variable_storage
*storage
;
978 if (i
== ir
->num_state_slots
) {
979 /* We'll set the index later. */
980 storage
= new(mem_ctx
) variable_storage(ir
, PROGRAM_STATE_VAR
, -1);
981 this->variables
.push_tail(storage
);
985 /* The variable_storage constructor allocates slots based on the size
986 * of the type. However, this had better match the number of state
987 * elements that we're going to copy into the new temporary.
989 assert((int) ir
->num_state_slots
== type_size(ir
->type
));
991 storage
= new(mem_ctx
) variable_storage(ir
, PROGRAM_TEMPORARY
,
993 this->variables
.push_tail(storage
);
994 this->next_temp
+= type_size(ir
->type
);
996 dst
= st_dst_reg(st_src_reg(PROGRAM_TEMPORARY
, storage
->index
,
997 glsl_version
>= 130 ? ir
->type
->base_type
: GLSL_TYPE_FLOAT
));
1001 for (unsigned int i
= 0; i
< ir
->num_state_slots
; i
++) {
1002 int index
= _mesa_add_state_reference(this->prog
->Parameters
,
1003 (gl_state_index
*)slots
[i
].tokens
);
1005 if (storage
->file
== PROGRAM_STATE_VAR
) {
1006 if (storage
->index
== -1) {
1007 storage
->index
= index
;
1009 assert(index
== storage
->index
+ (int)i
);
1012 st_src_reg
src(PROGRAM_STATE_VAR
, index
,
1013 glsl_version
>= 130 ? ir
->type
->base_type
: GLSL_TYPE_FLOAT
);
1014 src
.swizzle
= slots
[i
].swizzle
;
1015 emit(ir
, TGSI_OPCODE_MOV
, dst
, src
);
1016 /* even a float takes up a whole vec4 reg in a struct/array. */
1021 if (storage
->file
== PROGRAM_TEMPORARY
&&
1022 dst
.index
!= storage
->index
+ (int) ir
->num_state_slots
) {
1023 fail_link(this->shader_program
,
1024 "failed to load builtin uniform `%s' (%d/%d regs loaded)\n",
1025 ir
->name
, dst
.index
- storage
->index
,
1026 type_size(ir
->type
));
1032 glsl_to_tgsi_visitor::visit(ir_loop
*ir
)
1034 ir_dereference_variable
*counter
= NULL
;
1036 if (ir
->counter
!= NULL
)
1037 counter
= new(ir
) ir_dereference_variable(ir
->counter
);
1039 if (ir
->from
!= NULL
) {
1040 assert(ir
->counter
!= NULL
);
1042 ir_assignment
*a
= new(ir
) ir_assignment(counter
, ir
->from
, NULL
);
1048 emit(NULL
, TGSI_OPCODE_BGNLOOP
);
1052 new(ir
) ir_expression(ir
->cmp
, glsl_type::bool_type
,
1054 ir_if
*if_stmt
= new(ir
) ir_if(e
);
1056 ir_loop_jump
*brk
= new(ir
) ir_loop_jump(ir_loop_jump::jump_break
);
1058 if_stmt
->then_instructions
.push_tail(brk
);
1060 if_stmt
->accept(this);
1067 visit_exec_list(&ir
->body_instructions
, this);
1069 if (ir
->increment
) {
1071 new(ir
) ir_expression(ir_binop_add
, counter
->type
,
1072 counter
, ir
->increment
);
1074 ir_assignment
*a
= new(ir
) ir_assignment(counter
, e
, NULL
);
1081 emit(NULL
, TGSI_OPCODE_ENDLOOP
);
1085 glsl_to_tgsi_visitor::visit(ir_loop_jump
*ir
)
1088 case ir_loop_jump::jump_break
:
1089 emit(NULL
, TGSI_OPCODE_BRK
);
1091 case ir_loop_jump::jump_continue
:
1092 emit(NULL
, TGSI_OPCODE_CONT
);
1099 glsl_to_tgsi_visitor::visit(ir_function_signature
*ir
)
1106 glsl_to_tgsi_visitor::visit(ir_function
*ir
)
1108 /* Ignore function bodies other than main() -- we shouldn't see calls to
1109 * them since they should all be inlined before we get to glsl_to_tgsi.
1111 if (strcmp(ir
->name
, "main") == 0) {
1112 const ir_function_signature
*sig
;
1115 sig
= ir
->matching_signature(&empty
);
1119 foreach_iter(exec_list_iterator
, iter
, sig
->body
) {
1120 ir_instruction
*ir
= (ir_instruction
*)iter
.get();
1128 glsl_to_tgsi_visitor::try_emit_mad(ir_expression
*ir
, int mul_operand
)
1130 int nonmul_operand
= 1 - mul_operand
;
1133 ir_expression
*expr
= ir
->operands
[mul_operand
]->as_expression();
1134 if (!expr
|| expr
->operation
!= ir_binop_mul
)
1137 expr
->operands
[0]->accept(this);
1139 expr
->operands
[1]->accept(this);
1141 ir
->operands
[nonmul_operand
]->accept(this);
1144 this->result
= get_temp(ir
->type
);
1145 emit(ir
, TGSI_OPCODE_MAD
, st_dst_reg(this->result
), a
, b
, c
);
1151 glsl_to_tgsi_visitor::try_emit_sat(ir_expression
*ir
)
1153 /* Saturates were only introduced to vertex programs in
1154 * NV_vertex_program3, so don't give them to drivers in the VP.
1156 if (this->prog
->Target
== GL_VERTEX_PROGRAM_ARB
)
1159 ir_rvalue
*sat_src
= ir
->as_rvalue_to_saturate();
1163 sat_src
->accept(this);
1164 st_src_reg src
= this->result
;
1166 this->result
= get_temp(ir
->type
);
1167 glsl_to_tgsi_instruction
*inst
;
1168 inst
= emit(ir
, TGSI_OPCODE_MOV
, st_dst_reg(this->result
), src
);
1169 inst
->saturate
= true;
1175 glsl_to_tgsi_visitor::reladdr_to_temp(ir_instruction
*ir
,
1176 st_src_reg
*reg
, int *num_reladdr
)
1181 emit_arl(ir
, address_reg
, *reg
->reladdr
);
1183 if (*num_reladdr
!= 1) {
1184 st_src_reg temp
= get_temp(glsl_type::vec4_type
);
1186 emit(ir
, TGSI_OPCODE_MOV
, st_dst_reg(temp
), *reg
);
1194 glsl_to_tgsi_visitor::visit(ir_expression
*ir
)
1196 unsigned int operand
;
1197 st_src_reg op
[Elements(ir
->operands
)];
1198 st_src_reg result_src
;
1199 st_dst_reg result_dst
;
1201 /* Quick peephole: Emit MAD(a, b, c) instead of ADD(MUL(a, b), c)
1203 if (ir
->operation
== ir_binop_add
) {
1204 if (try_emit_mad(ir
, 1))
1206 if (try_emit_mad(ir
, 0))
1209 if (try_emit_sat(ir
))
1212 if (ir
->operation
== ir_quadop_vector
)
1213 assert(!"ir_quadop_vector should have been lowered");
1215 for (operand
= 0; operand
< ir
->get_num_operands(); operand
++) {
1216 this->result
.file
= PROGRAM_UNDEFINED
;
1217 ir
->operands
[operand
]->accept(this);
1218 if (this->result
.file
== PROGRAM_UNDEFINED
) {
1220 printf("Failed to get tree for expression operand:\n");
1221 ir
->operands
[operand
]->accept(&v
);
1224 op
[operand
] = this->result
;
1226 /* Matrix expression operands should have been broken down to vector
1227 * operations already.
1229 assert(!ir
->operands
[operand
]->type
->is_matrix());
1232 int vector_elements
= ir
->operands
[0]->type
->vector_elements
;
1233 if (ir
->operands
[1]) {
1234 vector_elements
= MAX2(vector_elements
,
1235 ir
->operands
[1]->type
->vector_elements
);
1238 this->result
.file
= PROGRAM_UNDEFINED
;
1240 /* Storage for our result. Ideally for an assignment we'd be using
1241 * the actual storage for the result here, instead.
1243 result_src
= get_temp(ir
->type
);
1244 /* convenience for the emit functions below. */
1245 result_dst
= st_dst_reg(result_src
);
1246 /* Limit writes to the channels that will be used by result_src later.
1247 * This does limit this temp's use as a temporary for multi-instruction
1250 result_dst
.writemask
= (1 << ir
->type
->vector_elements
) - 1;
1252 switch (ir
->operation
) {
1253 case ir_unop_logic_not
:
1254 emit(ir
, TGSI_OPCODE_SEQ
, result_dst
, op
[0], st_src_reg_for_type(result_dst
.type
, 0));
1257 assert(result_dst
.type
== GLSL_TYPE_FLOAT
|| result_dst
.type
== GLSL_TYPE_INT
);
1258 if (result_dst
.type
== GLSL_TYPE_INT
)
1259 emit(ir
, TGSI_OPCODE_INEG
, result_dst
, op
[0]);
1261 op
[0].negate
= ~op
[0].negate
;
1266 assert(result_dst
.type
== GLSL_TYPE_FLOAT
);
1267 emit(ir
, TGSI_OPCODE_ABS
, result_dst
, op
[0]);
1270 emit(ir
, TGSI_OPCODE_SSG
, result_dst
, op
[0]);
1273 emit_scalar(ir
, TGSI_OPCODE_RCP
, result_dst
, op
[0]);
1277 emit_scalar(ir
, TGSI_OPCODE_EX2
, result_dst
, op
[0]);
1281 assert(!"not reached: should be handled by ir_explog_to_explog2");
1284 emit_scalar(ir
, TGSI_OPCODE_LG2
, result_dst
, op
[0]);
1287 emit_scalar(ir
, TGSI_OPCODE_SIN
, result_dst
, op
[0]);
1290 emit_scalar(ir
, TGSI_OPCODE_COS
, result_dst
, op
[0]);
1292 case ir_unop_sin_reduced
:
1293 emit_scs(ir
, TGSI_OPCODE_SIN
, result_dst
, op
[0]);
1295 case ir_unop_cos_reduced
:
1296 emit_scs(ir
, TGSI_OPCODE_COS
, result_dst
, op
[0]);
1300 emit(ir
, TGSI_OPCODE_DDX
, result_dst
, op
[0]);
1303 op
[0].negate
= ~op
[0].negate
;
1304 emit(ir
, TGSI_OPCODE_DDY
, result_dst
, op
[0]);
1307 case ir_unop_noise
: {
1308 /* At some point, a motivated person could add a better
1309 * implementation of noise. Currently not even the nvidia
1310 * binary drivers do anything more than this. In any case, the
1311 * place to do this is in the GL state tracker, not the poor
1314 emit(ir
, TGSI_OPCODE_MOV
, result_dst
, st_src_reg_for_float(0.5));
1319 emit(ir
, TGSI_OPCODE_ADD
, result_dst
, op
[0], op
[1]);
1322 emit(ir
, TGSI_OPCODE_SUB
, result_dst
, op
[0], op
[1]);
1326 emit(ir
, TGSI_OPCODE_MUL
, result_dst
, op
[0], op
[1]);
1329 if (result_dst
.type
== GLSL_TYPE_FLOAT
)
1330 assert(!"not reached: should be handled by ir_div_to_mul_rcp");
1332 emit(ir
, TGSI_OPCODE_DIV
, result_dst
, op
[0], op
[1]);
1335 if (result_dst
.type
== GLSL_TYPE_FLOAT
)
1336 assert(!"ir_binop_mod should have been converted to b * fract(a/b)");
1338 emit(ir
, TGSI_OPCODE_MOD
, result_dst
, op
[0], op
[1]);
1342 emit(ir
, TGSI_OPCODE_SLT
, result_dst
, op
[0], op
[1]);
1344 case ir_binop_greater
:
1345 emit(ir
, TGSI_OPCODE_SGT
, result_dst
, op
[0], op
[1]);
1347 case ir_binop_lequal
:
1348 emit(ir
, TGSI_OPCODE_SLE
, result_dst
, op
[0], op
[1]);
1350 case ir_binop_gequal
:
1351 emit(ir
, TGSI_OPCODE_SGE
, result_dst
, op
[0], op
[1]);
1353 case ir_binop_equal
:
1354 emit(ir
, TGSI_OPCODE_SEQ
, result_dst
, op
[0], op
[1]);
1356 case ir_binop_nequal
:
1357 emit(ir
, TGSI_OPCODE_SNE
, result_dst
, op
[0], op
[1]);
1359 case ir_binop_all_equal
:
1360 /* "==" operator producing a scalar boolean. */
1361 if (ir
->operands
[0]->type
->is_vector() ||
1362 ir
->operands
[1]->type
->is_vector()) {
1363 st_src_reg temp
= get_temp(glsl_version
>= 130 ?
1364 glsl_type::get_vec4_type(ir
->operands
[0]->type
->base_type
) :
1365 glsl_type::vec4_type
);
1366 assert(ir
->operands
[0]->type
->base_type
== GLSL_TYPE_FLOAT
);
1367 emit(ir
, TGSI_OPCODE_SNE
, st_dst_reg(temp
), op
[0], op
[1]);
1368 emit_dp(ir
, result_dst
, temp
, temp
, vector_elements
);
1369 emit(ir
, TGSI_OPCODE_SEQ
, result_dst
, result_src
, st_src_reg_for_float(0.0));
1371 emit(ir
, TGSI_OPCODE_SEQ
, result_dst
, op
[0], op
[1]);
1374 case ir_binop_any_nequal
:
1375 /* "!=" operator producing a scalar boolean. */
1376 if (ir
->operands
[0]->type
->is_vector() ||
1377 ir
->operands
[1]->type
->is_vector()) {
1378 st_src_reg temp
= get_temp(glsl_version
>= 130 ?
1379 glsl_type::get_vec4_type(ir
->operands
[0]->type
->base_type
) :
1380 glsl_type::vec4_type
);
1381 assert(ir
->operands
[0]->type
->base_type
== GLSL_TYPE_FLOAT
);
1382 emit(ir
, TGSI_OPCODE_SNE
, st_dst_reg(temp
), op
[0], op
[1]);
1383 emit_dp(ir
, result_dst
, temp
, temp
, vector_elements
);
1384 emit(ir
, TGSI_OPCODE_SNE
, result_dst
, result_src
, st_src_reg_for_float(0.0));
1386 emit(ir
, TGSI_OPCODE_SNE
, result_dst
, op
[0], op
[1]);
1391 assert(ir
->operands
[0]->type
->is_vector());
1392 emit_dp(ir
, result_dst
, op
[0], op
[0],
1393 ir
->operands
[0]->type
->vector_elements
);
1394 emit(ir
, TGSI_OPCODE_SNE
, result_dst
, result_src
, st_src_reg_for_float(0.0));
1397 case ir_binop_logic_xor
:
1398 emit(ir
, TGSI_OPCODE_SNE
, result_dst
, op
[0], op
[1]);
1401 case ir_binop_logic_or
:
1402 /* This could be a saturated add and skip the SNE. */
1403 emit(ir
, TGSI_OPCODE_ADD
, result_dst
, op
[0], op
[1]);
1404 emit(ir
, TGSI_OPCODE_SNE
, result_dst
, result_src
, st_src_reg_for_float(0.0));
1407 case ir_binop_logic_and
:
1408 /* the bool args are stored as float 0.0 or 1.0, so "mul" gives us "and". */
1409 emit(ir
, TGSI_OPCODE_MUL
, result_dst
, op
[0], op
[1]);
1413 assert(ir
->operands
[0]->type
->is_vector());
1414 assert(ir
->operands
[0]->type
== ir
->operands
[1]->type
);
1415 emit_dp(ir
, result_dst
, op
[0], op
[1],
1416 ir
->operands
[0]->type
->vector_elements
);
1420 /* sqrt(x) = x * rsq(x). */
1421 emit_scalar(ir
, TGSI_OPCODE_RSQ
, result_dst
, op
[0]);
1422 emit(ir
, TGSI_OPCODE_MUL
, result_dst
, result_src
, op
[0]);
1423 /* For incoming channels <= 0, set the result to 0. */
1424 op
[0].negate
= ~op
[0].negate
;
1425 emit(ir
, TGSI_OPCODE_CMP
, result_dst
,
1426 op
[0], result_src
, st_src_reg_for_float(0.0));
1429 emit_scalar(ir
, TGSI_OPCODE_RSQ
, result_dst
, op
[0]);
1433 if (glsl_version
>= 130) {
1434 emit(ir
, TGSI_OPCODE_I2F
, result_dst
, op
[0]);
1438 /* Booleans are stored as integers (or floats in GLSL 1.20 and lower). */
1442 if (glsl_version
>= 130)
1443 emit(ir
, TGSI_OPCODE_F2I
, result_dst
, op
[0]);
1445 emit(ir
, TGSI_OPCODE_TRUNC
, result_dst
, op
[0]);
1449 emit(ir
, TGSI_OPCODE_SNE
, result_dst
, op
[0],
1450 st_src_reg_for_type(result_dst
.type
, 0));
1453 emit(ir
, TGSI_OPCODE_TRUNC
, result_dst
, op
[0]);
1456 op
[0].negate
= ~op
[0].negate
;
1457 emit(ir
, TGSI_OPCODE_FLR
, result_dst
, op
[0]);
1458 result_src
.negate
= ~result_src
.negate
;
1461 emit(ir
, TGSI_OPCODE_FLR
, result_dst
, op
[0]);
1464 emit(ir
, TGSI_OPCODE_FRC
, result_dst
, op
[0]);
1468 emit(ir
, TGSI_OPCODE_MIN
, result_dst
, op
[0], op
[1]);
1471 emit(ir
, TGSI_OPCODE_MAX
, result_dst
, op
[0], op
[1]);
1474 emit_scalar(ir
, TGSI_OPCODE_POW
, result_dst
, op
[0], op
[1]);
1477 case ir_unop_bit_not
:
1478 if (glsl_version
>= 130) {
1479 emit(ir
, TGSI_OPCODE_NOT
, result_dst
, op
[0]);
1483 if (glsl_version
>= 130) {
1484 emit(ir
, TGSI_OPCODE_U2F
, result_dst
, op
[0]);
1487 case ir_binop_lshift
:
1488 if (glsl_version
>= 130) {
1489 emit(ir
, TGSI_OPCODE_SHL
, result_dst
, op
[0]);
1492 case ir_binop_rshift
:
1493 if (glsl_version
>= 130) {
1494 emit(ir
, TGSI_OPCODE_ISHR
, result_dst
, op
[0]);
1497 case ir_binop_bit_and
:
1498 if (glsl_version
>= 130) {
1499 emit(ir
, TGSI_OPCODE_AND
, result_dst
, op
[0]);
1502 case ir_binop_bit_xor
:
1503 if (glsl_version
>= 130) {
1504 emit(ir
, TGSI_OPCODE_XOR
, result_dst
, op
[0]);
1507 case ir_binop_bit_or
:
1508 if (glsl_version
>= 130) {
1509 emit(ir
, TGSI_OPCODE_OR
, result_dst
, op
[0]);
1512 case ir_unop_round_even
:
1513 assert(!"GLSL 1.30 features unsupported");
1516 case ir_quadop_vector
:
1517 /* This operation should have already been handled.
1519 assert(!"Should not get here.");
1523 this->result
= result_src
;
1528 glsl_to_tgsi_visitor::visit(ir_swizzle
*ir
)
1534 /* Note that this is only swizzles in expressions, not those on the left
1535 * hand side of an assignment, which do write masking. See ir_assignment
1539 ir
->val
->accept(this);
1541 assert(src
.file
!= PROGRAM_UNDEFINED
);
1543 for (i
= 0; i
< 4; i
++) {
1544 if (i
< ir
->type
->vector_elements
) {
1547 swizzle
[i
] = GET_SWZ(src
.swizzle
, ir
->mask
.x
);
1550 swizzle
[i
] = GET_SWZ(src
.swizzle
, ir
->mask
.y
);
1553 swizzle
[i
] = GET_SWZ(src
.swizzle
, ir
->mask
.z
);
1556 swizzle
[i
] = GET_SWZ(src
.swizzle
, ir
->mask
.w
);
1560 /* If the type is smaller than a vec4, replicate the last
1563 swizzle
[i
] = swizzle
[ir
->type
->vector_elements
- 1];
1567 src
.swizzle
= MAKE_SWIZZLE4(swizzle
[0], swizzle
[1], swizzle
[2], swizzle
[3]);
1573 glsl_to_tgsi_visitor::visit(ir_dereference_variable
*ir
)
1575 variable_storage
*entry
= find_variable_storage(ir
->var
);
1576 ir_variable
*var
= ir
->var
;
1579 switch (var
->mode
) {
1580 case ir_var_uniform
:
1581 entry
= new(mem_ctx
) variable_storage(var
, PROGRAM_UNIFORM
,
1583 this->variables
.push_tail(entry
);
1587 /* The linker assigns locations for varyings and attributes,
1588 * including deprecated builtins (like gl_Color), user-assign
1589 * generic attributes (glBindVertexLocation), and
1590 * user-defined varyings.
1592 * FINISHME: We would hit this path for function arguments. Fix!
1594 assert(var
->location
!= -1);
1595 entry
= new(mem_ctx
) variable_storage(var
,
1598 if (this->prog
->Target
== GL_VERTEX_PROGRAM_ARB
&&
1599 var
->location
>= VERT_ATTRIB_GENERIC0
) {
1600 _mesa_add_attribute(this->prog
->Attributes
,
1602 _mesa_sizeof_glsl_type(var
->type
->gl_type
),
1604 var
->location
- VERT_ATTRIB_GENERIC0
);
1608 assert(var
->location
!= -1);
1609 entry
= new(mem_ctx
) variable_storage(var
,
1613 case ir_var_system_value
:
1614 entry
= new(mem_ctx
) variable_storage(var
,
1615 PROGRAM_SYSTEM_VALUE
,
1619 case ir_var_temporary
:
1620 entry
= new(mem_ctx
) variable_storage(var
, PROGRAM_TEMPORARY
,
1622 this->variables
.push_tail(entry
);
1624 next_temp
+= type_size(var
->type
);
1629 printf("Failed to make storage for %s\n", var
->name
);
1634 this->result
= st_src_reg(entry
->file
, entry
->index
, var
->type
);
1638 glsl_to_tgsi_visitor::visit(ir_dereference_array
*ir
)
1642 int element_size
= type_size(ir
->type
);
1644 index
= ir
->array_index
->constant_expression_value();
1646 ir
->array
->accept(this);
1650 src
.index
+= index
->value
.i
[0] * element_size
;
1652 st_src_reg array_base
= this->result
;
1653 /* Variable index array dereference. It eats the "vec4" of the
1654 * base of the array and an index that offsets the Mesa register
1657 ir
->array_index
->accept(this);
1659 st_src_reg index_reg
;
1661 if (element_size
== 1) {
1662 index_reg
= this->result
;
1664 index_reg
= get_temp(glsl_type::float_type
);
1666 emit(ir
, TGSI_OPCODE_MUL
, st_dst_reg(index_reg
),
1667 this->result
, st_src_reg_for_float(element_size
));
1670 src
.reladdr
= ralloc(mem_ctx
, st_src_reg
);
1671 memcpy(src
.reladdr
, &index_reg
, sizeof(index_reg
));
1674 /* If the type is smaller than a vec4, replicate the last channel out. */
1675 if (ir
->type
->is_scalar() || ir
->type
->is_vector())
1676 src
.swizzle
= swizzle_for_size(ir
->type
->vector_elements
);
1678 src
.swizzle
= SWIZZLE_NOOP
;
1684 glsl_to_tgsi_visitor::visit(ir_dereference_record
*ir
)
1687 const glsl_type
*struct_type
= ir
->record
->type
;
1690 ir
->record
->accept(this);
1692 for (i
= 0; i
< struct_type
->length
; i
++) {
1693 if (strcmp(struct_type
->fields
.structure
[i
].name
, ir
->field
) == 0)
1695 offset
+= type_size(struct_type
->fields
.structure
[i
].type
);
1698 /* If the type is smaller than a vec4, replicate the last channel out. */
1699 if (ir
->type
->is_scalar() || ir
->type
->is_vector())
1700 this->result
.swizzle
= swizzle_for_size(ir
->type
->vector_elements
);
1702 this->result
.swizzle
= SWIZZLE_NOOP
;
1704 this->result
.index
+= offset
;
1708 * We want to be careful in assignment setup to hit the actual storage
1709 * instead of potentially using a temporary like we might with the
1710 * ir_dereference handler.
1713 get_assignment_lhs(ir_dereference
*ir
, glsl_to_tgsi_visitor
*v
)
1715 /* The LHS must be a dereference. If the LHS is a variable indexed array
1716 * access of a vector, it must be separated into a series conditional moves
1717 * before reaching this point (see ir_vec_index_to_cond_assign).
1719 assert(ir
->as_dereference());
1720 ir_dereference_array
*deref_array
= ir
->as_dereference_array();
1722 assert(!deref_array
->array
->type
->is_vector());
1725 /* Use the rvalue deref handler for the most part. We'll ignore
1726 * swizzles in it and write swizzles using writemask, though.
1729 return st_dst_reg(v
->result
);
1733 * Process the condition of a conditional assignment
1735 * Examines the condition of a conditional assignment to generate the optimal
1736 * first operand of a \c CMP instruction. If the condition is a relational
1737 * operator with 0 (e.g., \c ir_binop_less), the value being compared will be
1738 * used as the source for the \c CMP instruction. Otherwise the comparison
1739 * is processed to a boolean result, and the boolean result is used as the
1740 * operand to the CMP instruction.
1743 glsl_to_tgsi_visitor::process_move_condition(ir_rvalue
*ir
)
1745 ir_rvalue
*src_ir
= ir
;
1747 bool switch_order
= false;
1749 ir_expression
*const expr
= ir
->as_expression();
1750 if ((expr
!= NULL
) && (expr
->get_num_operands() == 2)) {
1751 bool zero_on_left
= false;
1753 if (expr
->operands
[0]->is_zero()) {
1754 src_ir
= expr
->operands
[1];
1755 zero_on_left
= true;
1756 } else if (expr
->operands
[1]->is_zero()) {
1757 src_ir
= expr
->operands
[0];
1758 zero_on_left
= false;
1762 * (a < 0) T F F ( a < 0) T F F
1763 * (0 < a) F F T (-a < 0) F F T
1764 * (a <= 0) T T F (-a < 0) F F T (swap order of other operands)
1765 * (0 <= a) F T T ( a < 0) T F F (swap order of other operands)
1766 * (a > 0) F F T (-a < 0) F F T
1767 * (0 > a) T F F ( a < 0) T F F
1768 * (a >= 0) F T T ( a < 0) T F F (swap order of other operands)
1769 * (0 >= a) T T F (-a < 0) F F T (swap order of other operands)
1771 * Note that exchanging the order of 0 and 'a' in the comparison simply
1772 * means that the value of 'a' should be negated.
1775 switch (expr
->operation
) {
1777 switch_order
= false;
1778 negate
= zero_on_left
;
1781 case ir_binop_greater
:
1782 switch_order
= false;
1783 negate
= !zero_on_left
;
1786 case ir_binop_lequal
:
1787 switch_order
= true;
1788 negate
= !zero_on_left
;
1791 case ir_binop_gequal
:
1792 switch_order
= true;
1793 negate
= zero_on_left
;
1797 /* This isn't the right kind of comparison afterall, so make sure
1798 * the whole condition is visited.
1806 src_ir
->accept(this);
1808 /* We use the TGSI_OPCODE_CMP (a < 0 ? b : c) for conditional moves, and the
1809 * condition we produced is 0.0 or 1.0. By flipping the sign, we can
1810 * choose which value TGSI_OPCODE_CMP produces without an extra instruction
1811 * computing the condition.
1814 this->result
.negate
= ~this->result
.negate
;
1816 return switch_order
;
1820 glsl_to_tgsi_visitor::visit(ir_assignment
*ir
)
1826 ir
->rhs
->accept(this);
1829 l
= get_assignment_lhs(ir
->lhs
, this);
1831 /* FINISHME: This should really set to the correct maximal writemask for each
1832 * FINISHME: component written (in the loops below). This case can only
1833 * FINISHME: occur for matrices, arrays, and structures.
1835 if (ir
->write_mask
== 0) {
1836 assert(!ir
->lhs
->type
->is_scalar() && !ir
->lhs
->type
->is_vector());
1837 l
.writemask
= WRITEMASK_XYZW
;
1838 } else if (ir
->lhs
->type
->is_scalar()) {
1839 /* FINISHME: This hack makes writing to gl_FragDepth, which lives in the
1840 * FINISHME: W component of fragment shader output zero, work correctly.
1842 l
.writemask
= WRITEMASK_XYZW
;
1845 int first_enabled_chan
= 0;
1848 assert(ir
->lhs
->type
->is_vector());
1849 l
.writemask
= ir
->write_mask
;
1851 for (int i
= 0; i
< 4; i
++) {
1852 if (l
.writemask
& (1 << i
)) {
1853 first_enabled_chan
= GET_SWZ(r
.swizzle
, i
);
1858 /* Swizzle a small RHS vector into the channels being written.
1860 * glsl ir treats write_mask as dictating how many channels are
1861 * present on the RHS while Mesa IR treats write_mask as just
1862 * showing which channels of the vec4 RHS get written.
1864 for (int i
= 0; i
< 4; i
++) {
1865 if (l
.writemask
& (1 << i
))
1866 swizzles
[i
] = GET_SWZ(r
.swizzle
, rhs_chan
++);
1868 swizzles
[i
] = first_enabled_chan
;
1870 r
.swizzle
= MAKE_SWIZZLE4(swizzles
[0], swizzles
[1],
1871 swizzles
[2], swizzles
[3]);
1874 assert(l
.file
!= PROGRAM_UNDEFINED
);
1875 assert(r
.file
!= PROGRAM_UNDEFINED
);
1877 if (ir
->condition
) {
1878 const bool switch_order
= this->process_move_condition(ir
->condition
);
1879 st_src_reg condition
= this->result
;
1881 for (i
= 0; i
< type_size(ir
->lhs
->type
); i
++) {
1883 emit(ir
, TGSI_OPCODE_CMP
, l
, condition
, st_src_reg(l
), r
);
1885 emit(ir
, TGSI_OPCODE_CMP
, l
, condition
, r
, st_src_reg(l
));
1892 for (i
= 0; i
< type_size(ir
->lhs
->type
); i
++) {
1893 emit(ir
, TGSI_OPCODE_MOV
, l
, r
);
1902 glsl_to_tgsi_visitor::visit(ir_constant
*ir
)
1905 GLfloat stack_vals
[4] = { 0 };
1906 gl_constant_value
*values
= (gl_constant_value
*) stack_vals
;
1907 GLenum gl_type
= GL_NONE
;
1910 /* Unfortunately, 4 floats is all we can get into
1911 * _mesa_add_unnamed_constant. So, make a temp to store an
1912 * aggregate constant and move each constant value into it. If we
1913 * get lucky, copy propagation will eliminate the extra moves.
1915 if (ir
->type
->base_type
== GLSL_TYPE_STRUCT
) {
1916 st_src_reg temp_base
= get_temp(ir
->type
);
1917 st_dst_reg temp
= st_dst_reg(temp_base
);
1919 foreach_iter(exec_list_iterator
, iter
, ir
->components
) {
1920 ir_constant
*field_value
= (ir_constant
*)iter
.get();
1921 int size
= type_size(field_value
->type
);
1925 field_value
->accept(this);
1928 for (i
= 0; i
< (unsigned int)size
; i
++) {
1929 emit(ir
, TGSI_OPCODE_MOV
, temp
, src
);
1935 this->result
= temp_base
;
1939 if (ir
->type
->is_array()) {
1940 st_src_reg temp_base
= get_temp(ir
->type
);
1941 st_dst_reg temp
= st_dst_reg(temp_base
);
1942 int size
= type_size(ir
->type
->fields
.array
);
1946 for (i
= 0; i
< ir
->type
->length
; i
++) {
1947 ir
->array_elements
[i
]->accept(this);
1949 for (int j
= 0; j
< size
; j
++) {
1950 emit(ir
, TGSI_OPCODE_MOV
, temp
, src
);
1956 this->result
= temp_base
;
1960 if (ir
->type
->is_matrix()) {
1961 st_src_reg mat
= get_temp(ir
->type
);
1962 st_dst_reg mat_column
= st_dst_reg(mat
);
1964 for (i
= 0; i
< ir
->type
->matrix_columns
; i
++) {
1965 assert(ir
->type
->base_type
== GLSL_TYPE_FLOAT
);
1966 values
= (gl_constant_value
*) &ir
->value
.f
[i
* ir
->type
->vector_elements
];
1968 src
= st_src_reg(PROGRAM_CONSTANT
, -1, ir
->type
->base_type
);
1969 src
.index
= _mesa_add_unnamed_constant(this->prog
->Parameters
,
1971 ir
->type
->vector_elements
,
1973 emit(ir
, TGSI_OPCODE_MOV
, mat_column
, src
);
1982 src
.file
= PROGRAM_CONSTANT
;
1983 switch (ir
->type
->base_type
) {
1984 case GLSL_TYPE_FLOAT
:
1986 for (i
= 0; i
< ir
->type
->vector_elements
; i
++) {
1987 values
[i
].f
= ir
->value
.f
[i
];
1990 case GLSL_TYPE_UINT
:
1991 gl_type
= glsl_version
>= 130 ? GL_UNSIGNED_INT
: GL_FLOAT
;
1992 for (i
= 0; i
< ir
->type
->vector_elements
; i
++) {
1993 if (glsl_version
>= 130)
1994 values
[i
].u
= ir
->value
.u
[i
];
1996 values
[i
].f
= ir
->value
.u
[i
];
2000 gl_type
= glsl_version
>= 130 ? GL_INT
: GL_FLOAT
;
2001 for (i
= 0; i
< ir
->type
->vector_elements
; i
++) {
2002 if (glsl_version
>= 130)
2003 values
[i
].i
= ir
->value
.i
[i
];
2005 values
[i
].f
= ir
->value
.i
[i
];
2008 case GLSL_TYPE_BOOL
:
2009 gl_type
= glsl_version
>= 130 ? GL_BOOL
: GL_FLOAT
;
2010 for (i
= 0; i
< ir
->type
->vector_elements
; i
++) {
2011 if (glsl_version
>= 130)
2012 values
[i
].b
= ir
->value
.b
[i
];
2014 values
[i
].f
= ir
->value
.b
[i
];
2018 assert(!"Non-float/uint/int/bool constant");
2021 this->result
= st_src_reg(PROGRAM_CONSTANT
, -1, ir
->type
);
2022 this->result
.index
= _mesa_add_typed_unnamed_constant(this->prog
->Parameters
,
2023 values
, ir
->type
->vector_elements
, gl_type
,
2024 &this->result
.swizzle
);
2028 glsl_to_tgsi_visitor::get_function_signature(ir_function_signature
*sig
)
2030 function_entry
*entry
;
2032 foreach_iter(exec_list_iterator
, iter
, this->function_signatures
) {
2033 entry
= (function_entry
*)iter
.get();
2035 if (entry
->sig
== sig
)
2039 entry
= ralloc(mem_ctx
, function_entry
);
2041 entry
->sig_id
= this->next_signature_id
++;
2042 entry
->bgn_inst
= NULL
;
2044 /* Allocate storage for all the parameters. */
2045 foreach_iter(exec_list_iterator
, iter
, sig
->parameters
) {
2046 ir_variable
*param
= (ir_variable
*)iter
.get();
2047 variable_storage
*storage
;
2049 storage
= find_variable_storage(param
);
2052 storage
= new(mem_ctx
) variable_storage(param
, PROGRAM_TEMPORARY
,
2054 this->variables
.push_tail(storage
);
2056 this->next_temp
+= type_size(param
->type
);
2059 if (!sig
->return_type
->is_void()) {
2060 entry
->return_reg
= get_temp(sig
->return_type
);
2062 entry
->return_reg
= undef_src
;
2065 this->function_signatures
.push_tail(entry
);
2070 glsl_to_tgsi_visitor::visit(ir_call
*ir
)
2072 glsl_to_tgsi_instruction
*call_inst
;
2073 ir_function_signature
*sig
= ir
->get_callee();
2074 function_entry
*entry
= get_function_signature(sig
);
2077 /* Process in parameters. */
2078 exec_list_iterator sig_iter
= sig
->parameters
.iterator();
2079 foreach_iter(exec_list_iterator
, iter
, *ir
) {
2080 ir_rvalue
*param_rval
= (ir_rvalue
*)iter
.get();
2081 ir_variable
*param
= (ir_variable
*)sig_iter
.get();
2083 if (param
->mode
== ir_var_in
||
2084 param
->mode
== ir_var_inout
) {
2085 variable_storage
*storage
= find_variable_storage(param
);
2088 param_rval
->accept(this);
2089 st_src_reg r
= this->result
;
2092 l
.file
= storage
->file
;
2093 l
.index
= storage
->index
;
2095 l
.writemask
= WRITEMASK_XYZW
;
2096 l
.cond_mask
= COND_TR
;
2098 for (i
= 0; i
< type_size(param
->type
); i
++) {
2099 emit(ir
, TGSI_OPCODE_MOV
, l
, r
);
2107 assert(!sig_iter
.has_next());
2109 /* Emit call instruction */
2110 call_inst
= emit(ir
, TGSI_OPCODE_CAL
);
2111 call_inst
->function
= entry
;
2113 /* Process out parameters. */
2114 sig_iter
= sig
->parameters
.iterator();
2115 foreach_iter(exec_list_iterator
, iter
, *ir
) {
2116 ir_rvalue
*param_rval
= (ir_rvalue
*)iter
.get();
2117 ir_variable
*param
= (ir_variable
*)sig_iter
.get();
2119 if (param
->mode
== ir_var_out
||
2120 param
->mode
== ir_var_inout
) {
2121 variable_storage
*storage
= find_variable_storage(param
);
2125 r
.file
= storage
->file
;
2126 r
.index
= storage
->index
;
2128 r
.swizzle
= SWIZZLE_NOOP
;
2131 param_rval
->accept(this);
2132 st_dst_reg l
= st_dst_reg(this->result
);
2134 for (i
= 0; i
< type_size(param
->type
); i
++) {
2135 emit(ir
, TGSI_OPCODE_MOV
, l
, r
);
2143 assert(!sig_iter
.has_next());
2145 /* Process return value. */
2146 this->result
= entry
->return_reg
;
2150 glsl_to_tgsi_visitor::visit(ir_texture
*ir
)
2152 st_src_reg result_src
, coord
, lod_info
, projector
, dx
, dy
;
2153 st_dst_reg result_dst
, coord_dst
;
2154 glsl_to_tgsi_instruction
*inst
= NULL
;
2155 unsigned opcode
= TGSI_OPCODE_NOP
;
2157 ir
->coordinate
->accept(this);
2159 /* Put our coords in a temp. We'll need to modify them for shadow,
2160 * projection, or LOD, so the only case we'd use it as is is if
2161 * we're doing plain old texturing. Mesa IR optimization should
2162 * handle cleaning up our mess in that case.
2164 coord
= get_temp(glsl_type::vec4_type
);
2165 coord_dst
= st_dst_reg(coord
);
2166 emit(ir
, TGSI_OPCODE_MOV
, coord_dst
, this->result
);
2168 if (ir
->projector
) {
2169 ir
->projector
->accept(this);
2170 projector
= this->result
;
2173 /* Storage for our result. Ideally for an assignment we'd be using
2174 * the actual storage for the result here, instead.
2176 result_src
= get_temp(glsl_type::vec4_type
);
2177 result_dst
= st_dst_reg(result_src
);
2181 opcode
= TGSI_OPCODE_TEX
;
2184 opcode
= TGSI_OPCODE_TXB
;
2185 ir
->lod_info
.bias
->accept(this);
2186 lod_info
= this->result
;
2189 opcode
= TGSI_OPCODE_TXL
;
2190 ir
->lod_info
.lod
->accept(this);
2191 lod_info
= this->result
;
2194 opcode
= TGSI_OPCODE_TXD
;
2195 ir
->lod_info
.grad
.dPdx
->accept(this);
2197 ir
->lod_info
.grad
.dPdy
->accept(this);
2200 case ir_txf
: // TODO: use TGSI_OPCODE_TXF here
2201 assert(!"GLSL 1.30 features unsupported");
2205 if (ir
->projector
) {
2206 if (opcode
== TGSI_OPCODE_TEX
) {
2207 /* Slot the projector in as the last component of the coord. */
2208 coord_dst
.writemask
= WRITEMASK_W
;
2209 emit(ir
, TGSI_OPCODE_MOV
, coord_dst
, projector
);
2210 coord_dst
.writemask
= WRITEMASK_XYZW
;
2211 opcode
= TGSI_OPCODE_TXP
;
2213 st_src_reg coord_w
= coord
;
2214 coord_w
.swizzle
= SWIZZLE_WWWW
;
2216 /* For the other TEX opcodes there's no projective version
2217 * since the last slot is taken up by LOD info. Do the
2218 * projective divide now.
2220 coord_dst
.writemask
= WRITEMASK_W
;
2221 emit(ir
, TGSI_OPCODE_RCP
, coord_dst
, projector
);
2223 /* In the case where we have to project the coordinates "by hand,"
2224 * the shadow comparator value must also be projected.
2226 st_src_reg tmp_src
= coord
;
2227 if (ir
->shadow_comparitor
) {
2228 /* Slot the shadow value in as the second to last component of the
2231 ir
->shadow_comparitor
->accept(this);
2233 tmp_src
= get_temp(glsl_type::vec4_type
);
2234 st_dst_reg tmp_dst
= st_dst_reg(tmp_src
);
2236 tmp_dst
.writemask
= WRITEMASK_Z
;
2237 emit(ir
, TGSI_OPCODE_MOV
, tmp_dst
, this->result
);
2239 tmp_dst
.writemask
= WRITEMASK_XY
;
2240 emit(ir
, TGSI_OPCODE_MOV
, tmp_dst
, coord
);
2243 coord_dst
.writemask
= WRITEMASK_XYZ
;
2244 emit(ir
, TGSI_OPCODE_MUL
, coord_dst
, tmp_src
, coord_w
);
2246 coord_dst
.writemask
= WRITEMASK_XYZW
;
2247 coord
.swizzle
= SWIZZLE_XYZW
;
2251 /* If projection is done and the opcode is not TGSI_OPCODE_TXP, then the shadow
2252 * comparator was put in the correct place (and projected) by the code,
2253 * above, that handles by-hand projection.
2255 if (ir
->shadow_comparitor
&& (!ir
->projector
|| opcode
== TGSI_OPCODE_TXP
)) {
2256 /* Slot the shadow value in as the second to last component of the
2259 ir
->shadow_comparitor
->accept(this);
2260 coord_dst
.writemask
= WRITEMASK_Z
;
2261 emit(ir
, TGSI_OPCODE_MOV
, coord_dst
, this->result
);
2262 coord_dst
.writemask
= WRITEMASK_XYZW
;
2265 if (opcode
== TGSI_OPCODE_TXL
|| opcode
== TGSI_OPCODE_TXB
) {
2266 /* TGSI stores LOD or LOD bias in the last channel of the coords. */
2267 coord_dst
.writemask
= WRITEMASK_W
;
2268 emit(ir
, TGSI_OPCODE_MOV
, coord_dst
, lod_info
);
2269 coord_dst
.writemask
= WRITEMASK_XYZW
;
2272 if (opcode
== TGSI_OPCODE_TXD
)
2273 inst
= emit(ir
, opcode
, result_dst
, coord
, dx
, dy
);
2275 inst
= emit(ir
, opcode
, result_dst
, coord
);
2277 if (ir
->shadow_comparitor
)
2278 inst
->tex_shadow
= GL_TRUE
;
2280 inst
->sampler
= _mesa_get_sampler_uniform_value(ir
->sampler
,
2281 this->shader_program
,
2284 const glsl_type
*sampler_type
= ir
->sampler
->type
;
2286 switch (sampler_type
->sampler_dimensionality
) {
2287 case GLSL_SAMPLER_DIM_1D
:
2288 inst
->tex_target
= (sampler_type
->sampler_array
)
2289 ? TEXTURE_1D_ARRAY_INDEX
: TEXTURE_1D_INDEX
;
2291 case GLSL_SAMPLER_DIM_2D
:
2292 inst
->tex_target
= (sampler_type
->sampler_array
)
2293 ? TEXTURE_2D_ARRAY_INDEX
: TEXTURE_2D_INDEX
;
2295 case GLSL_SAMPLER_DIM_3D
:
2296 inst
->tex_target
= TEXTURE_3D_INDEX
;
2298 case GLSL_SAMPLER_DIM_CUBE
:
2299 inst
->tex_target
= TEXTURE_CUBE_INDEX
;
2301 case GLSL_SAMPLER_DIM_RECT
:
2302 inst
->tex_target
= TEXTURE_RECT_INDEX
;
2304 case GLSL_SAMPLER_DIM_BUF
:
2305 assert(!"FINISHME: Implement ARB_texture_buffer_object");
2308 assert(!"Should not get here.");
2311 this->result
= result_src
;
2315 glsl_to_tgsi_visitor::visit(ir_return
*ir
)
2317 if (ir
->get_value()) {
2321 assert(current_function
);
2323 ir
->get_value()->accept(this);
2324 st_src_reg r
= this->result
;
2326 l
= st_dst_reg(current_function
->return_reg
);
2328 for (i
= 0; i
< type_size(current_function
->sig
->return_type
); i
++) {
2329 emit(ir
, TGSI_OPCODE_MOV
, l
, r
);
2335 emit(ir
, TGSI_OPCODE_RET
);
2339 glsl_to_tgsi_visitor::visit(ir_discard
*ir
)
2341 struct gl_fragment_program
*fp
= (struct gl_fragment_program
*)this->prog
;
2343 if (ir
->condition
) {
2344 ir
->condition
->accept(this);
2345 this->result
.negate
= ~this->result
.negate
;
2346 emit(ir
, TGSI_OPCODE_KIL
, undef_dst
, this->result
);
2348 emit(ir
, TGSI_OPCODE_KILP
);
2351 fp
->UsesKill
= GL_TRUE
;
2355 glsl_to_tgsi_visitor::visit(ir_if
*ir
)
2357 glsl_to_tgsi_instruction
*cond_inst
, *if_inst
, *else_inst
= NULL
;
2358 glsl_to_tgsi_instruction
*prev_inst
;
2360 prev_inst
= (glsl_to_tgsi_instruction
*)this->instructions
.get_tail();
2362 ir
->condition
->accept(this);
2363 assert(this->result
.file
!= PROGRAM_UNDEFINED
);
2365 if (this->options
->EmitCondCodes
) {
2366 cond_inst
= (glsl_to_tgsi_instruction
*)this->instructions
.get_tail();
2368 /* See if we actually generated any instruction for generating
2369 * the condition. If not, then cook up a move to a temp so we
2370 * have something to set cond_update on.
2372 if (cond_inst
== prev_inst
) {
2373 st_src_reg temp
= get_temp(glsl_type::bool_type
);
2374 cond_inst
= emit(ir
->condition
, TGSI_OPCODE_MOV
, st_dst_reg(temp
), result
);
2376 cond_inst
->cond_update
= GL_TRUE
;
2378 if_inst
= emit(ir
->condition
, TGSI_OPCODE_IF
);
2379 if_inst
->dst
.cond_mask
= COND_NE
;
2381 if_inst
= emit(ir
->condition
, TGSI_OPCODE_IF
, undef_dst
, this->result
);
2384 this->instructions
.push_tail(if_inst
);
2386 visit_exec_list(&ir
->then_instructions
, this);
2388 if (!ir
->else_instructions
.is_empty()) {
2389 else_inst
= emit(ir
->condition
, TGSI_OPCODE_ELSE
);
2390 visit_exec_list(&ir
->else_instructions
, this);
2393 if_inst
= emit(ir
->condition
, TGSI_OPCODE_ENDIF
);
2396 glsl_to_tgsi_visitor::glsl_to_tgsi_visitor()
2398 result
.file
= PROGRAM_UNDEFINED
;
2400 next_signature_id
= 1;
2401 current_function
= NULL
;
2402 num_address_regs
= 0;
2403 indirect_addr_temps
= false;
2404 indirect_addr_consts
= false;
2405 mem_ctx
= ralloc_context(NULL
);
2408 glsl_to_tgsi_visitor::~glsl_to_tgsi_visitor()
2410 ralloc_free(mem_ctx
);
2413 extern "C" void free_glsl_to_tgsi_visitor(glsl_to_tgsi_visitor
*v
)
2420 * Count resources used by the given gpu program (number of texture
2424 count_resources(glsl_to_tgsi_visitor
*v
, gl_program
*prog
)
2426 v
->samplers_used
= 0;
2428 foreach_iter(exec_list_iterator
, iter
, v
->instructions
) {
2429 glsl_to_tgsi_instruction
*inst
= (glsl_to_tgsi_instruction
*)iter
.get();
2431 if (is_tex_instruction(inst
->op
)) {
2432 v
->samplers_used
|= 1 << inst
->sampler
;
2434 prog
->SamplerTargets
[inst
->sampler
] =
2435 (gl_texture_index
)inst
->tex_target
;
2436 if (inst
->tex_shadow
) {
2437 prog
->ShadowSamplers
|= 1 << inst
->sampler
;
2442 prog
->SamplersUsed
= v
->samplers_used
;
2443 _mesa_update_shader_textures_used(prog
);
2448 * Check if the given vertex/fragment/shader program is within the
2449 * resource limits of the context (number of texture units, etc).
2450 * If any of those checks fail, record a linker error.
2452 * XXX more checks are needed...
2455 check_resources(const struct gl_context
*ctx
,
2456 struct gl_shader_program
*shader_program
,
2457 glsl_to_tgsi_visitor
*prog
,
2458 struct gl_program
*proginfo
)
2460 switch (proginfo
->Target
) {
2461 case GL_VERTEX_PROGRAM_ARB
:
2462 if (_mesa_bitcount(prog
->samplers_used
) >
2463 ctx
->Const
.MaxVertexTextureImageUnits
) {
2464 fail_link(shader_program
, "Too many vertex shader texture samplers");
2466 if (proginfo
->Parameters
->NumParameters
> MAX_UNIFORMS
) {
2467 fail_link(shader_program
, "Too many vertex shader constants");
2470 case MESA_GEOMETRY_PROGRAM
:
2471 if (_mesa_bitcount(prog
->samplers_used
) >
2472 ctx
->Const
.MaxGeometryTextureImageUnits
) {
2473 fail_link(shader_program
, "Too many geometry shader texture samplers");
2475 if (proginfo
->Parameters
->NumParameters
>
2476 MAX_GEOMETRY_UNIFORM_COMPONENTS
/ 4) {
2477 fail_link(shader_program
, "Too many geometry shader constants");
2480 case GL_FRAGMENT_PROGRAM_ARB
:
2481 if (_mesa_bitcount(prog
->samplers_used
) >
2482 ctx
->Const
.MaxTextureImageUnits
) {
2483 fail_link(shader_program
, "Too many fragment shader texture samplers");
2485 if (proginfo
->Parameters
->NumParameters
> MAX_UNIFORMS
) {
2486 fail_link(shader_program
, "Too many fragment shader constants");
2490 _mesa_problem(ctx
, "unexpected program type in check_resources()");
2496 struct uniform_sort
{
2497 struct gl_uniform
*u
;
2501 /* The shader_program->Uniforms list is almost sorted in increasing
2502 * uniform->{Frag,Vert}Pos locations, but not quite when there are
2503 * uniforms shared between targets. We need to add parameters in
2504 * increasing order for the targets.
2507 sort_uniforms(const void *a
, const void *b
)
2509 struct uniform_sort
*u1
= (struct uniform_sort
*)a
;
2510 struct uniform_sort
*u2
= (struct uniform_sort
*)b
;
2512 return u1
->pos
- u2
->pos
;
2515 /* Add the uniforms to the parameters. The linker chose locations
2516 * in our parameters lists (which weren't created yet), which the
2517 * uniforms code will use to poke values into our parameters list
2518 * when uniforms are updated.
2521 add_uniforms_to_parameters_list(struct gl_shader_program
*shader_program
,
2522 struct gl_shader
*shader
,
2523 struct gl_program
*prog
)
2526 unsigned int next_sampler
= 0, num_uniforms
= 0;
2527 struct uniform_sort
*sorted_uniforms
;
2529 sorted_uniforms
= ralloc_array(NULL
, struct uniform_sort
,
2530 shader_program
->Uniforms
->NumUniforms
);
2532 for (i
= 0; i
< shader_program
->Uniforms
->NumUniforms
; i
++) {
2533 struct gl_uniform
*uniform
= shader_program
->Uniforms
->Uniforms
+ i
;
2534 int parameter_index
= -1;
2536 switch (shader
->Type
) {
2537 case GL_VERTEX_SHADER
:
2538 parameter_index
= uniform
->VertPos
;
2540 case GL_FRAGMENT_SHADER
:
2541 parameter_index
= uniform
->FragPos
;
2543 case GL_GEOMETRY_SHADER
:
2544 parameter_index
= uniform
->GeomPos
;
2548 /* Only add uniforms used in our target. */
2549 if (parameter_index
!= -1) {
2550 sorted_uniforms
[num_uniforms
].pos
= parameter_index
;
2551 sorted_uniforms
[num_uniforms
].u
= uniform
;
2556 qsort(sorted_uniforms
, num_uniforms
, sizeof(struct uniform_sort
),
2559 for (i
= 0; i
< num_uniforms
; i
++) {
2560 struct gl_uniform
*uniform
= sorted_uniforms
[i
].u
;
2561 int parameter_index
= sorted_uniforms
[i
].pos
;
2562 const glsl_type
*type
= uniform
->Type
;
2565 if (type
->is_vector() ||
2566 type
->is_scalar()) {
2567 size
= type
->vector_elements
;
2569 size
= type_size(type
) * 4;
2572 gl_register_file file
;
2573 if (type
->is_sampler() ||
2574 (type
->is_array() && type
->fields
.array
->is_sampler())) {
2575 file
= PROGRAM_SAMPLER
;
2577 file
= PROGRAM_UNIFORM
;
2580 GLint index
= _mesa_lookup_parameter_index(prog
->Parameters
, -1,
2584 index
= _mesa_add_parameter(prog
->Parameters
, file
,
2585 uniform
->Name
, size
, type
->gl_type
,
2588 /* Sampler uniform values are stored in prog->SamplerUnits,
2589 * and the entry in that array is selected by this index we
2590 * store in ParameterValues[].
2592 if (file
== PROGRAM_SAMPLER
) {
2593 for (unsigned int j
= 0; j
< size
/ 4; j
++)
2594 prog
->Parameters
->ParameterValues
[index
+ j
][0].f
= next_sampler
++;
2597 /* The location chosen in the Parameters list here (returned
2598 * from _mesa_add_uniform) has to match what the linker chose.
2600 if (index
!= parameter_index
) {
2601 fail_link(shader_program
, "Allocation of uniform `%s' to target "
2602 "failed (%d vs %d)\n",
2603 uniform
->Name
, index
, parameter_index
);
2608 ralloc_free(sorted_uniforms
);
2612 set_uniform_initializer(struct gl_context
*ctx
, void *mem_ctx
,
2613 struct gl_shader_program
*shader_program
,
2614 const char *name
, const glsl_type
*type
,
2617 if (type
->is_record()) {
2618 ir_constant
*field_constant
;
2620 field_constant
= (ir_constant
*)val
->components
.get_head();
2622 for (unsigned int i
= 0; i
< type
->length
; i
++) {
2623 const glsl_type
*field_type
= type
->fields
.structure
[i
].type
;
2624 const char *field_name
= ralloc_asprintf(mem_ctx
, "%s.%s", name
,
2625 type
->fields
.structure
[i
].name
);
2626 set_uniform_initializer(ctx
, mem_ctx
, shader_program
, field_name
,
2627 field_type
, field_constant
);
2628 field_constant
= (ir_constant
*)field_constant
->next
;
2633 int loc
= _mesa_get_uniform_location(ctx
, shader_program
, name
);
2636 fail_link(shader_program
,
2637 "Couldn't find uniform for initializer %s\n", name
);
2641 for (unsigned int i
= 0; i
< (type
->is_array() ? type
->length
: 1); i
++) {
2642 ir_constant
*element
;
2643 const glsl_type
*element_type
;
2644 if (type
->is_array()) {
2645 element
= val
->array_elements
[i
];
2646 element_type
= type
->fields
.array
;
2649 element_type
= type
;
2654 if (element_type
->base_type
== GLSL_TYPE_BOOL
) {
2655 int *conv
= ralloc_array(mem_ctx
, int, element_type
->components());
2656 for (unsigned int j
= 0; j
< element_type
->components(); j
++) {
2657 conv
[j
] = element
->value
.b
[j
];
2659 values
= (void *)conv
;
2660 element_type
= glsl_type::get_instance(GLSL_TYPE_INT
,
2661 element_type
->vector_elements
,
2664 values
= &element
->value
;
2667 if (element_type
->is_matrix()) {
2668 _mesa_uniform_matrix(ctx
, shader_program
,
2669 element_type
->matrix_columns
,
2670 element_type
->vector_elements
,
2671 loc
, 1, GL_FALSE
, (GLfloat
*)values
);
2672 loc
+= element_type
->matrix_columns
;
2674 _mesa_uniform(ctx
, shader_program
, loc
, element_type
->matrix_columns
,
2675 values
, element_type
->gl_type
);
2676 loc
+= type_size(element_type
);
2682 set_uniform_initializers(struct gl_context
*ctx
,
2683 struct gl_shader_program
*shader_program
)
2685 void *mem_ctx
= NULL
;
2687 for (unsigned int i
= 0; i
< MESA_SHADER_TYPES
; i
++) {
2688 struct gl_shader
*shader
= shader_program
->_LinkedShaders
[i
];
2693 foreach_iter(exec_list_iterator
, iter
, *shader
->ir
) {
2694 ir_instruction
*ir
= (ir_instruction
*)iter
.get();
2695 ir_variable
*var
= ir
->as_variable();
2697 if (!var
|| var
->mode
!= ir_var_uniform
|| !var
->constant_value
)
2701 mem_ctx
= ralloc_context(NULL
);
2703 set_uniform_initializer(ctx
, mem_ctx
, shader_program
, var
->name
,
2704 var
->type
, var
->constant_value
);
2708 ralloc_free(mem_ctx
);
2712 * Scan/rewrite program to remove reads of custom (output) registers.
2713 * The passed type has to be either PROGRAM_OUTPUT or PROGRAM_VARYING
2714 * (for vertex shaders).
2715 * In GLSL shaders, varying vars can be read and written.
2716 * On some hardware, trying to read an output register causes trouble.
2717 * So, rewrite the program to use a temporary register in this case.
2719 * Based on _mesa_remove_output_reads from programopt.c.
2722 glsl_to_tgsi_visitor::remove_output_reads(gl_register_file type
)
2725 GLint outputMap
[VERT_RESULT_MAX
];
2726 GLint outputTypes
[VERT_RESULT_MAX
];
2727 GLuint numVaryingReads
= 0;
2728 GLboolean usedTemps
[MAX_PROGRAM_TEMPS
];
2729 GLuint firstTemp
= 0;
2731 _mesa_find_used_registers(prog
, PROGRAM_TEMPORARY
,
2732 usedTemps
, MAX_PROGRAM_TEMPS
);
2734 assert(type
== PROGRAM_VARYING
|| type
== PROGRAM_OUTPUT
);
2735 assert(prog
->Target
== GL_VERTEX_PROGRAM_ARB
|| type
!= PROGRAM_VARYING
);
2737 for (i
= 0; i
< VERT_RESULT_MAX
; i
++)
2740 /* look for instructions which read from varying vars */
2741 foreach_iter(exec_list_iterator
, iter
, this->instructions
) {
2742 glsl_to_tgsi_instruction
*inst
= (glsl_to_tgsi_instruction
*)iter
.get();
2743 const GLuint numSrc
= num_inst_src_regs(inst
->op
);
2745 for (j
= 0; j
< numSrc
; j
++) {
2746 if (inst
->src
[j
].file
== type
) {
2747 /* replace the read with a temp reg */
2748 const GLuint var
= inst
->src
[j
].index
;
2749 if (outputMap
[var
] == -1) {
2751 outputMap
[var
] = _mesa_find_free_register(usedTemps
,
2754 outputTypes
[var
] = inst
->src
[j
].type
;
2755 firstTemp
= outputMap
[var
] + 1;
2757 inst
->src
[j
].file
= PROGRAM_TEMPORARY
;
2758 inst
->src
[j
].index
= outputMap
[var
];
2763 if (numVaryingReads
== 0)
2764 return; /* nothing to be done */
2766 /* look for instructions which write to the varying vars identified above */
2767 foreach_iter(exec_list_iterator
, iter
, this->instructions
) {
2768 glsl_to_tgsi_instruction
*inst
= (glsl_to_tgsi_instruction
*)iter
.get();
2769 if (inst
->dst
.file
== type
&& outputMap
[inst
->dst
.index
] >= 0) {
2770 /* change inst to write to the temp reg, instead of the varying */
2771 inst
->dst
.file
= PROGRAM_TEMPORARY
;
2772 inst
->dst
.index
= outputMap
[inst
->dst
.index
];
2776 /* insert new MOV instructions at the end */
2777 for (i
= 0; i
< VERT_RESULT_MAX
; i
++) {
2778 if (outputMap
[i
] >= 0) {
2779 /* MOV VAR[i], TEMP[tmp]; */
2780 st_src_reg src
= st_src_reg(PROGRAM_TEMPORARY
, outputMap
[i
], outputTypes
[i
]);
2781 st_dst_reg dst
= st_dst_reg(type
, WRITEMASK_XYZW
, outputTypes
[i
]);
2783 this->emit(NULL
, TGSI_OPCODE_MOV
, dst
, src
);
2788 /* Replaces all references to a temporary register index with another index. */
2790 glsl_to_tgsi_visitor::rename_temp_register(int index
, int new_index
)
2792 foreach_iter(exec_list_iterator
, iter
, this->instructions
) {
2793 glsl_to_tgsi_instruction
*inst
= (glsl_to_tgsi_instruction
*)iter
.get();
2796 for (j
=0; j
< num_inst_src_regs(inst
->op
); j
++) {
2797 if (inst
->src
[j
].file
== PROGRAM_TEMPORARY
&&
2798 inst
->src
[j
].index
== index
) {
2799 inst
->src
[j
].index
= new_index
;
2803 if (inst
->dst
.file
== PROGRAM_TEMPORARY
&& inst
->dst
.index
== index
) {
2804 inst
->dst
.index
= new_index
;
2810 glsl_to_tgsi_visitor::get_first_temp_read(int index
)
2812 int depth
= 0; /* loop depth */
2813 int loop_start
= -1; /* index of the first active BGNLOOP (if any) */
2816 foreach_iter(exec_list_iterator
, iter
, this->instructions
) {
2817 glsl_to_tgsi_instruction
*inst
= (glsl_to_tgsi_instruction
*)iter
.get();
2819 for (j
=0; j
< num_inst_src_regs(inst
->op
); j
++) {
2820 if (inst
->src
[j
].file
== PROGRAM_TEMPORARY
&&
2821 inst
->src
[j
].index
== index
) {
2822 return (depth
== 0) ? i
: loop_start
;
2826 if (inst
->op
== TGSI_OPCODE_BGNLOOP
) {
2829 } else if (inst
->op
== TGSI_OPCODE_ENDLOOP
) {
2842 glsl_to_tgsi_visitor::get_first_temp_write(int index
)
2844 int depth
= 0; /* loop depth */
2845 int loop_start
= -1; /* index of the first active BGNLOOP (if any) */
2848 foreach_iter(exec_list_iterator
, iter
, this->instructions
) {
2849 glsl_to_tgsi_instruction
*inst
= (glsl_to_tgsi_instruction
*)iter
.get();
2851 if (inst
->dst
.file
== PROGRAM_TEMPORARY
&& inst
->dst
.index
== index
) {
2852 return (depth
== 0) ? i
: loop_start
;
2855 if (inst
->op
== TGSI_OPCODE_BGNLOOP
) {
2858 } else if (inst
->op
== TGSI_OPCODE_ENDLOOP
) {
2871 glsl_to_tgsi_visitor::get_last_temp_read(int index
)
2873 int depth
= 0; /* loop depth */
2874 int last
= -1; /* index of last instruction that reads the temporary */
2877 foreach_iter(exec_list_iterator
, iter
, this->instructions
) {
2878 glsl_to_tgsi_instruction
*inst
= (glsl_to_tgsi_instruction
*)iter
.get();
2880 for (j
=0; j
< num_inst_src_regs(inst
->op
); j
++) {
2881 if (inst
->src
[j
].file
== PROGRAM_TEMPORARY
&&
2882 inst
->src
[j
].index
== index
) {
2883 last
= (depth
== 0) ? i
: -2;
2887 if (inst
->op
== TGSI_OPCODE_BGNLOOP
)
2889 else if (inst
->op
== TGSI_OPCODE_ENDLOOP
)
2890 if (--depth
== 0 && last
== -2)
2902 glsl_to_tgsi_visitor::get_last_temp_write(int index
)
2904 int depth
= 0; /* loop depth */
2905 int last
= -1; /* index of last instruction that writes to the temporary */
2908 foreach_iter(exec_list_iterator
, iter
, this->instructions
) {
2909 glsl_to_tgsi_instruction
*inst
= (glsl_to_tgsi_instruction
*)iter
.get();
2911 if (inst
->dst
.file
== PROGRAM_TEMPORARY
&& inst
->dst
.index
== index
)
2912 last
= (depth
== 0) ? i
: -2;
2914 if (inst
->op
== TGSI_OPCODE_BGNLOOP
)
2916 else if (inst
->op
== TGSI_OPCODE_ENDLOOP
)
2917 if (--depth
== 0 && last
== -2)
2929 * On a basic block basis, tracks available PROGRAM_TEMPORARY register
2930 * channels for copy propagation and updates following instructions to
2931 * use the original versions.
2933 * The glsl_to_tgsi_visitor lazily produces code assuming that this pass
2934 * will occur. As an example, a TXP production before this pass:
2936 * 0: MOV TEMP[1], INPUT[4].xyyy;
2937 * 1: MOV TEMP[1].w, INPUT[4].wwww;
2938 * 2: TXP TEMP[2], TEMP[1], texture[0], 2D;
2942 * 0: MOV TEMP[1], INPUT[4].xyyy;
2943 * 1: MOV TEMP[1].w, INPUT[4].wwww;
2944 * 2: TXP TEMP[2], INPUT[4].xyyw, texture[0], 2D;
2946 * which allows for dead code elimination on TEMP[1]'s writes.
2949 glsl_to_tgsi_visitor::copy_propagate(void)
2951 glsl_to_tgsi_instruction
**acp
= rzalloc_array(mem_ctx
,
2952 glsl_to_tgsi_instruction
*,
2953 this->next_temp
* 4);
2954 int *acp_level
= rzalloc_array(mem_ctx
, int, this->next_temp
* 4);
2957 foreach_iter(exec_list_iterator
, iter
, this->instructions
) {
2958 glsl_to_tgsi_instruction
*inst
= (glsl_to_tgsi_instruction
*)iter
.get();
2960 assert(inst
->dst
.file
!= PROGRAM_TEMPORARY
2961 || inst
->dst
.index
< this->next_temp
);
2963 /* First, do any copy propagation possible into the src regs. */
2964 for (int r
= 0; r
< 3; r
++) {
2965 glsl_to_tgsi_instruction
*first
= NULL
;
2967 int acp_base
= inst
->src
[r
].index
* 4;
2969 if (inst
->src
[r
].file
!= PROGRAM_TEMPORARY
||
2970 inst
->src
[r
].reladdr
)
2973 /* See if we can find entries in the ACP consisting of MOVs
2974 * from the same src register for all the swizzled channels
2975 * of this src register reference.
2977 for (int i
= 0; i
< 4; i
++) {
2978 int src_chan
= GET_SWZ(inst
->src
[r
].swizzle
, i
);
2979 glsl_to_tgsi_instruction
*copy_chan
= acp
[acp_base
+ src_chan
];
2986 assert(acp_level
[acp_base
+ src_chan
] <= level
);
2991 if (first
->src
[0].file
!= copy_chan
->src
[0].file
||
2992 first
->src
[0].index
!= copy_chan
->src
[0].index
) {
3000 /* We've now validated that we can copy-propagate to
3001 * replace this src register reference. Do it.
3003 inst
->src
[r
].file
= first
->src
[0].file
;
3004 inst
->src
[r
].index
= first
->src
[0].index
;
3007 for (int i
= 0; i
< 4; i
++) {
3008 int src_chan
= GET_SWZ(inst
->src
[r
].swizzle
, i
);
3009 glsl_to_tgsi_instruction
*copy_inst
= acp
[acp_base
+ src_chan
];
3010 swizzle
|= (GET_SWZ(copy_inst
->src
[0].swizzle
, src_chan
) <<
3013 inst
->src
[r
].swizzle
= swizzle
;
3018 case TGSI_OPCODE_BGNLOOP
:
3019 case TGSI_OPCODE_ENDLOOP
:
3020 /* End of a basic block, clear the ACP entirely. */
3021 memset(acp
, 0, sizeof(*acp
) * this->next_temp
* 4);
3024 case TGSI_OPCODE_IF
:
3028 case TGSI_OPCODE_ENDIF
:
3029 case TGSI_OPCODE_ELSE
:
3030 /* Clear all channels written inside the block from the ACP, but
3031 * leaving those that were not touched.
3033 for (int r
= 0; r
< this->next_temp
; r
++) {
3034 for (int c
= 0; c
< 4; c
++) {
3035 if (!acp
[4 * r
+ c
])
3038 if (acp_level
[4 * r
+ c
] >= level
)
3039 acp
[4 * r
+ c
] = NULL
;
3042 if (inst
->op
== TGSI_OPCODE_ENDIF
)
3047 /* Continuing the block, clear any written channels from
3050 if (inst
->dst
.file
== PROGRAM_TEMPORARY
&& inst
->dst
.reladdr
) {
3051 /* Any temporary might be written, so no copy propagation
3052 * across this instruction.
3054 memset(acp
, 0, sizeof(*acp
) * this->next_temp
* 4);
3055 } else if (inst
->dst
.file
== PROGRAM_OUTPUT
&&
3056 inst
->dst
.reladdr
) {
3057 /* Any output might be written, so no copy propagation
3058 * from outputs across this instruction.
3060 for (int r
= 0; r
< this->next_temp
; r
++) {
3061 for (int c
= 0; c
< 4; c
++) {
3062 if (!acp
[4 * r
+ c
])
3065 if (acp
[4 * r
+ c
]->src
[0].file
== PROGRAM_OUTPUT
)
3066 acp
[4 * r
+ c
] = NULL
;
3069 } else if (inst
->dst
.file
== PROGRAM_TEMPORARY
||
3070 inst
->dst
.file
== PROGRAM_OUTPUT
) {
3071 /* Clear where it's used as dst. */
3072 if (inst
->dst
.file
== PROGRAM_TEMPORARY
) {
3073 for (int c
= 0; c
< 4; c
++) {
3074 if (inst
->dst
.writemask
& (1 << c
)) {
3075 acp
[4 * inst
->dst
.index
+ c
] = NULL
;
3080 /* Clear where it's used as src. */
3081 for (int r
= 0; r
< this->next_temp
; r
++) {
3082 for (int c
= 0; c
< 4; c
++) {
3083 if (!acp
[4 * r
+ c
])
3086 int src_chan
= GET_SWZ(acp
[4 * r
+ c
]->src
[0].swizzle
, c
);
3088 if (acp
[4 * r
+ c
]->src
[0].file
== inst
->dst
.file
&&
3089 acp
[4 * r
+ c
]->src
[0].index
== inst
->dst
.index
&&
3090 inst
->dst
.writemask
& (1 << src_chan
))
3092 acp
[4 * r
+ c
] = NULL
;
3100 /* If this is a copy, add it to the ACP. */
3101 if (inst
->op
== TGSI_OPCODE_MOV
&&
3102 inst
->dst
.file
== PROGRAM_TEMPORARY
&&
3103 !inst
->dst
.reladdr
&&
3105 !inst
->src
[0].reladdr
&&
3106 !inst
->src
[0].negate
) {
3107 for (int i
= 0; i
< 4; i
++) {
3108 if (inst
->dst
.writemask
& (1 << i
)) {
3109 acp
[4 * inst
->dst
.index
+ i
] = inst
;
3110 acp_level
[4 * inst
->dst
.index
+ i
] = level
;
3116 ralloc_free(acp_level
);
3121 * Tracks available PROGRAM_TEMPORARY registers for dead code elimination.
3123 * The glsl_to_tgsi_visitor lazily produces code assuming that this pass
3124 * will occur. As an example, a TXP production after copy propagation but
3127 * 0: MOV TEMP[1], INPUT[4].xyyy;
3128 * 1: MOV TEMP[1].w, INPUT[4].wwww;
3129 * 2: TXP TEMP[2], INPUT[4].xyyw, texture[0], 2D;
3131 * and after this pass:
3133 * 0: TXP TEMP[2], INPUT[4].xyyw, texture[0], 2D;
3135 * FIXME: assumes that all functions are inlined (no support for BGNSUB/ENDSUB)
3136 * FIXME: doesn't eliminate all dead code inside of loops; it steps around them
3139 glsl_to_tgsi_visitor::eliminate_dead_code(void)
3143 for (i
=0; i
< this->next_temp
; i
++) {
3144 int last_read
= get_last_temp_read(i
);
3147 foreach_iter(exec_list_iterator
, iter
, this->instructions
) {
3148 glsl_to_tgsi_instruction
*inst
= (glsl_to_tgsi_instruction
*)iter
.get();
3150 if (inst
->dst
.file
== PROGRAM_TEMPORARY
&& inst
->dst
.index
== i
&&
3162 /* Merges temporary registers together where possible to reduce the number of
3163 * registers needed to run a program.
3165 * Produces optimal code only after copy propagation and dead code elimination
3168 glsl_to_tgsi_visitor::merge_registers(void)
3170 int *last_reads
= rzalloc_array(mem_ctx
, int, this->next_temp
);
3171 int *first_writes
= rzalloc_array(mem_ctx
, int, this->next_temp
);
3174 /* Read the indices of the last read and first write to each temp register
3175 * into an array so that we don't have to traverse the instruction list as
3177 for (i
=0; i
< this->next_temp
; i
++) {
3178 last_reads
[i
] = get_last_temp_read(i
);
3179 first_writes
[i
] = get_first_temp_write(i
);
3182 /* Start looking for registers with non-overlapping usages that can be
3183 * merged together. */
3184 for (i
=0; i
< this->next_temp
; i
++) {
3185 /* Don't touch unused registers. */
3186 if (last_reads
[i
] < 0 || first_writes
[i
] < 0) continue;
3188 for (j
=0; j
< this->next_temp
; j
++) {
3189 /* Don't touch unused registers. */
3190 if (last_reads
[j
] < 0 || first_writes
[j
] < 0) continue;
3192 /* We can merge the two registers if the first write to j is after or
3193 * in the same instruction as the last read from i. Note that the
3194 * register at index i will always be used earlier or at the same time
3195 * as the register at index j. */
3196 if (first_writes
[i
] <= first_writes
[j
] &&
3197 last_reads
[i
] <= first_writes
[j
])
3199 rename_temp_register(j
, i
); /* Replace all references to j with i.*/
3201 /* Update the first_writes and last_reads arrays with the new
3202 * values for the merged register index, and mark the newly unused
3203 * register index as such. */
3204 last_reads
[i
] = last_reads
[j
];
3205 first_writes
[j
] = -1;
3211 ralloc_free(last_reads
);
3212 ralloc_free(first_writes
);
3215 /* Reassign indices to temporary registers by reusing unused indices created
3216 * by optimization passes. */
3218 glsl_to_tgsi_visitor::renumber_registers(void)
3223 for (i
=0; i
< this->next_temp
; i
++) {
3224 if (get_first_temp_read(i
) < 0) continue;
3226 rename_temp_register(i
, new_index
);
3230 this->next_temp
= new_index
;
3233 /* ------------------------- TGSI conversion stuff -------------------------- */
3235 unsigned branch_target
;
3240 * Intermediate state used during shader translation.
3242 struct st_translate
{
3243 struct ureg_program
*ureg
;
3245 struct ureg_dst temps
[MAX_PROGRAM_TEMPS
];
3246 struct ureg_src
*constants
;
3247 struct ureg_dst outputs
[PIPE_MAX_SHADER_OUTPUTS
];
3248 struct ureg_src inputs
[PIPE_MAX_SHADER_INPUTS
];
3249 struct ureg_dst address
[1];
3250 struct ureg_src samplers
[PIPE_MAX_SAMPLERS
];
3251 struct ureg_src systemValues
[SYSTEM_VALUE_MAX
];
3253 /* Extra info for handling point size clamping in vertex shader */
3254 struct ureg_dst pointSizeResult
; /**< Actual point size output register */
3255 struct ureg_src pointSizeConst
; /**< Point size range constant register */
3256 GLint pointSizeOutIndex
; /**< Temp point size output register */
3257 GLboolean prevInstWrotePointSize
;
3259 const GLuint
*inputMapping
;
3260 const GLuint
*outputMapping
;
3262 /* For every instruction that contains a label (eg CALL), keep
3263 * details so that we can go back afterwards and emit the correct
3264 * tgsi instruction number for each label.
3266 struct label
*labels
;
3267 unsigned labels_size
;
3268 unsigned labels_count
;
3270 /* Keep a record of the tgsi instruction number that each mesa
3271 * instruction starts at, will be used to fix up labels after
3276 unsigned insn_count
;
3278 unsigned procType
; /**< TGSI_PROCESSOR_VERTEX/FRAGMENT */
3283 /** Map Mesa's SYSTEM_VALUE_x to TGSI_SEMANTIC_x */
3284 static unsigned mesa_sysval_to_semantic
[SYSTEM_VALUE_MAX
] = {
3286 TGSI_SEMANTIC_INSTANCEID
3290 * Make note of a branch to a label in the TGSI code.
3291 * After we've emitted all instructions, we'll go over the list
3292 * of labels built here and patch the TGSI code with the actual
3293 * location of each label.
3295 static unsigned *get_label( struct st_translate
*t
,
3296 unsigned branch_target
)
3300 if (t
->labels_count
+ 1 >= t
->labels_size
) {
3301 t
->labels_size
= 1 << (util_logbase2(t
->labels_size
) + 1);
3302 t
->labels
= (struct label
*)realloc(t
->labels
,
3303 t
->labels_size
* sizeof t
->labels
[0]);
3304 if (t
->labels
== NULL
) {
3305 static unsigned dummy
;
3311 i
= t
->labels_count
++;
3312 t
->labels
[i
].branch_target
= branch_target
;
3313 return &t
->labels
[i
].token
;
3317 * Called prior to emitting the TGSI code for each Mesa instruction.
3318 * Allocate additional space for instructions if needed.
3319 * Update the insn[] array so the next Mesa instruction points to
3320 * the next TGSI instruction.
3322 static void set_insn_start( struct st_translate
*t
,
3325 if (t
->insn_count
+ 1 >= t
->insn_size
) {
3326 t
->insn_size
= 1 << (util_logbase2(t
->insn_size
) + 1);
3327 t
->insn
= (unsigned *)realloc(t
->insn
, t
->insn_size
* sizeof t
->insn
[0]);
3328 if (t
->insn
== NULL
) {
3334 t
->insn
[t
->insn_count
++] = start
;
3338 * Map a Mesa dst register to a TGSI ureg_dst register.
3340 static struct ureg_dst
3341 dst_register( struct st_translate
*t
,
3342 gl_register_file file
,
3346 case PROGRAM_UNDEFINED
:
3347 return ureg_dst_undef();
3349 case PROGRAM_TEMPORARY
:
3350 if (ureg_dst_is_undef(t
->temps
[index
]))
3351 t
->temps
[index
] = ureg_DECL_temporary( t
->ureg
);
3353 return t
->temps
[index
];
3355 case PROGRAM_OUTPUT
:
3356 if (t
->procType
== TGSI_PROCESSOR_VERTEX
&& index
== VERT_RESULT_PSIZ
)
3357 t
->prevInstWrotePointSize
= GL_TRUE
;
3359 if (t
->procType
== TGSI_PROCESSOR_VERTEX
)
3360 assert(index
< VERT_RESULT_MAX
);
3361 else if (t
->procType
== TGSI_PROCESSOR_FRAGMENT
)
3362 assert(index
< FRAG_RESULT_MAX
);
3364 assert(index
< GEOM_RESULT_MAX
);
3366 assert(t
->outputMapping
[index
] < Elements(t
->outputs
));
3368 return t
->outputs
[t
->outputMapping
[index
]];
3370 case PROGRAM_ADDRESS
:
3371 return t
->address
[index
];
3375 return ureg_dst_undef();
3380 * Map a Mesa src register to a TGSI ureg_src register.
3382 static struct ureg_src
3383 src_register( struct st_translate
*t
,
3384 gl_register_file file
,
3388 case PROGRAM_UNDEFINED
:
3389 return ureg_src_undef();
3391 case PROGRAM_TEMPORARY
:
3393 assert(index
< Elements(t
->temps
));
3394 if (ureg_dst_is_undef(t
->temps
[index
]))
3395 t
->temps
[index
] = ureg_DECL_temporary( t
->ureg
);
3396 return ureg_src(t
->temps
[index
]);
3398 case PROGRAM_NAMED_PARAM
:
3399 case PROGRAM_ENV_PARAM
:
3400 case PROGRAM_LOCAL_PARAM
:
3401 case PROGRAM_UNIFORM
:
3403 return t
->constants
[index
];
3404 case PROGRAM_STATE_VAR
:
3405 case PROGRAM_CONSTANT
: /* ie, immediate */
3407 return ureg_DECL_constant( t
->ureg
, 0 );
3409 return t
->constants
[index
];
3412 assert(t
->inputMapping
[index
] < Elements(t
->inputs
));
3413 return t
->inputs
[t
->inputMapping
[index
]];
3415 case PROGRAM_OUTPUT
:
3416 assert(t
->outputMapping
[index
] < Elements(t
->outputs
));
3417 return ureg_src(t
->outputs
[t
->outputMapping
[index
]]); /* not needed? */
3419 case PROGRAM_ADDRESS
:
3420 return ureg_src(t
->address
[index
]);
3422 case PROGRAM_SYSTEM_VALUE
:
3423 assert(index
< Elements(t
->systemValues
));
3424 return t
->systemValues
[index
];
3428 return ureg_src_undef();
3433 * Create a TGSI ureg_dst register from an st_dst_reg.
3435 static struct ureg_dst
3436 translate_dst( struct st_translate
*t
,
3437 const st_dst_reg
*dst_reg
,
3440 struct ureg_dst dst
= dst_register( t
,
3444 dst
= ureg_writemask( dst
,
3445 dst_reg
->writemask
);
3448 dst
= ureg_saturate( dst
);
3450 if (dst_reg
->reladdr
!= NULL
)
3451 dst
= ureg_dst_indirect( dst
, ureg_src(t
->address
[0]) );
3457 * Create a TGSI ureg_src register from an st_src_reg.
3459 static struct ureg_src
3460 translate_src( struct st_translate
*t
,
3461 const st_src_reg
*src_reg
)
3463 struct ureg_src src
= src_register( t
, src_reg
->file
, src_reg
->index
);
3465 src
= ureg_swizzle( src
,
3466 GET_SWZ( src_reg
->swizzle
, 0 ) & 0x3,
3467 GET_SWZ( src_reg
->swizzle
, 1 ) & 0x3,
3468 GET_SWZ( src_reg
->swizzle
, 2 ) & 0x3,
3469 GET_SWZ( src_reg
->swizzle
, 3 ) & 0x3);
3471 if ((src_reg
->negate
& 0xf) == NEGATE_XYZW
)
3472 src
= ureg_negate(src
);
3474 if (src_reg
->reladdr
!= NULL
) {
3475 /* Normally ureg_src_indirect() would be used here, but a stupid compiler
3476 * bug in g++ makes ureg_src_indirect (an inline C function) erroneously
3477 * set the bit for src.Negate. So we have to do the operation manually
3478 * here to work around the compiler's problems. */
3479 /*src = ureg_src_indirect(src, ureg_src(t->address[0]));*/
3480 struct ureg_src addr
= ureg_src(t
->address
[0]);
3482 src
.IndirectFile
= addr
.File
;
3483 src
.IndirectIndex
= addr
.Index
;
3484 src
.IndirectSwizzle
= addr
.SwizzleX
;
3486 if (src_reg
->file
!= PROGRAM_INPUT
&&
3487 src_reg
->file
!= PROGRAM_OUTPUT
) {
3488 /* If src_reg->index was negative, it was set to zero in
3489 * src_register(). Reassign it now. But don't do this
3490 * for input/output regs since they get remapped while
3491 * const buffers don't.
3493 src
.Index
= src_reg
->index
;
3501 compile_tgsi_instruction(struct st_translate
*t
,
3502 const struct glsl_to_tgsi_instruction
*inst
)
3504 struct ureg_program
*ureg
= t
->ureg
;
3506 struct ureg_dst dst
[1];
3507 struct ureg_src src
[4];
3511 num_dst
= num_inst_dst_regs( inst
->op
);
3512 num_src
= num_inst_src_regs( inst
->op
);
3515 dst
[0] = translate_dst( t
,
3519 for (i
= 0; i
< num_src
; i
++)
3520 src
[i
] = translate_src( t
, &inst
->src
[i
] );
3522 switch( inst
->op
) {
3523 case TGSI_OPCODE_BGNLOOP
:
3524 case TGSI_OPCODE_CAL
:
3525 case TGSI_OPCODE_ELSE
:
3526 case TGSI_OPCODE_ENDLOOP
:
3527 case TGSI_OPCODE_IF
:
3528 debug_assert(num_dst
== 0);
3529 ureg_label_insn( ureg
,
3533 inst
->op
== TGSI_OPCODE_CAL
? inst
->function
->sig_id
: 0 ));
3536 case TGSI_OPCODE_TEX
:
3537 case TGSI_OPCODE_TXB
:
3538 case TGSI_OPCODE_TXD
:
3539 case TGSI_OPCODE_TXL
:
3540 case TGSI_OPCODE_TXP
:
3541 src
[num_src
++] = t
->samplers
[inst
->sampler
];
3542 ureg_tex_insn( ureg
,
3545 translate_texture_target( inst
->tex_target
,
3550 case TGSI_OPCODE_SCS
:
3551 dst
[0] = ureg_writemask(dst
[0], TGSI_WRITEMASK_XY
);
3558 case TGSI_OPCODE_XPD
:
3559 dst
[0] = ureg_writemask(dst
[0], TGSI_WRITEMASK_XYZ
);
3576 * Emit the TGSI instructions to adjust the WPOS pixel center convention
3577 * Basically, add (adjX, adjY) to the fragment position.
3580 emit_adjusted_wpos( struct st_translate
*t
,
3581 const struct gl_program
*program
,
3582 GLfloat adjX
, GLfloat adjY
)
3584 struct ureg_program
*ureg
= t
->ureg
;
3585 struct ureg_dst wpos_temp
= ureg_DECL_temporary(ureg
);
3586 struct ureg_src wpos_input
= t
->inputs
[t
->inputMapping
[FRAG_ATTRIB_WPOS
]];
3588 /* Note that we bias X and Y and pass Z and W through unchanged.
3589 * The shader might also use gl_FragCoord.w and .z.
3591 ureg_ADD(ureg
, wpos_temp
, wpos_input
,
3592 ureg_imm4f(ureg
, adjX
, adjY
, 0.0f
, 0.0f
));
3594 t
->inputs
[t
->inputMapping
[FRAG_ATTRIB_WPOS
]] = ureg_src(wpos_temp
);
3599 * Emit the TGSI instructions for inverting the WPOS y coordinate.
3600 * This code is unavoidable because it also depends on whether
3601 * a FBO is bound (STATE_FB_WPOS_Y_TRANSFORM).
3604 emit_wpos_inversion( struct st_translate
*t
,
3605 const struct gl_program
*program
,
3608 struct ureg_program
*ureg
= t
->ureg
;
3610 /* Fragment program uses fragment position input.
3611 * Need to replace instances of INPUT[WPOS] with temp T
3612 * where T = INPUT[WPOS] by y is inverted.
3614 static const gl_state_index wposTransformState
[STATE_LENGTH
]
3615 = { STATE_INTERNAL
, STATE_FB_WPOS_Y_TRANSFORM
,
3616 (gl_state_index
)0, (gl_state_index
)0, (gl_state_index
)0 };
3618 /* XXX: note we are modifying the incoming shader here! Need to
3619 * do this before emitting the constant decls below, or this
3622 unsigned wposTransConst
= _mesa_add_state_reference(program
->Parameters
,
3623 wposTransformState
);
3625 struct ureg_src wpostrans
= ureg_DECL_constant( ureg
, wposTransConst
);
3626 struct ureg_dst wpos_temp
;
3627 struct ureg_src wpos_input
= t
->inputs
[t
->inputMapping
[FRAG_ATTRIB_WPOS
]];
3629 /* MOV wpos_temp, input[wpos]
3631 if (wpos_input
.File
== TGSI_FILE_TEMPORARY
)
3632 wpos_temp
= ureg_dst(wpos_input
);
3634 wpos_temp
= ureg_DECL_temporary( ureg
);
3635 ureg_MOV( ureg
, wpos_temp
, wpos_input
);
3639 /* MAD wpos_temp.y, wpos_input, wpostrans.xxxx, wpostrans.yyyy
3642 ureg_writemask(wpos_temp
, TGSI_WRITEMASK_Y
),
3644 ureg_scalar(wpostrans
, 0),
3645 ureg_scalar(wpostrans
, 1));
3647 /* MAD wpos_temp.y, wpos_input, wpostrans.zzzz, wpostrans.wwww
3650 ureg_writemask(wpos_temp
, TGSI_WRITEMASK_Y
),
3652 ureg_scalar(wpostrans
, 2),
3653 ureg_scalar(wpostrans
, 3));
3656 /* Use wpos_temp as position input from here on:
3658 t
->inputs
[t
->inputMapping
[FRAG_ATTRIB_WPOS
]] = ureg_src(wpos_temp
);
3663 * Emit fragment position/ooordinate code.
3666 emit_wpos(struct st_context
*st
,
3667 struct st_translate
*t
,
3668 const struct gl_program
*program
,
3669 struct ureg_program
*ureg
)
3671 const struct gl_fragment_program
*fp
=
3672 (const struct gl_fragment_program
*) program
;
3673 struct pipe_screen
*pscreen
= st
->pipe
->screen
;
3674 boolean invert
= FALSE
;
3676 if (fp
->OriginUpperLeft
) {
3677 /* Fragment shader wants origin in upper-left */
3678 if (pscreen
->get_param(pscreen
, PIPE_CAP_TGSI_FS_COORD_ORIGIN_UPPER_LEFT
)) {
3679 /* the driver supports upper-left origin */
3681 else if (pscreen
->get_param(pscreen
, PIPE_CAP_TGSI_FS_COORD_ORIGIN_LOWER_LEFT
)) {
3682 /* the driver supports lower-left origin, need to invert Y */
3683 ureg_property_fs_coord_origin(ureg
, TGSI_FS_COORD_ORIGIN_LOWER_LEFT
);
3690 /* Fragment shader wants origin in lower-left */
3691 if (pscreen
->get_param(pscreen
, PIPE_CAP_TGSI_FS_COORD_ORIGIN_LOWER_LEFT
))
3692 /* the driver supports lower-left origin */
3693 ureg_property_fs_coord_origin(ureg
, TGSI_FS_COORD_ORIGIN_LOWER_LEFT
);
3694 else if (pscreen
->get_param(pscreen
, PIPE_CAP_TGSI_FS_COORD_ORIGIN_UPPER_LEFT
))
3695 /* the driver supports upper-left origin, need to invert Y */
3701 if (fp
->PixelCenterInteger
) {
3702 /* Fragment shader wants pixel center integer */
3703 if (pscreen
->get_param(pscreen
, PIPE_CAP_TGSI_FS_COORD_PIXEL_CENTER_INTEGER
))
3704 /* the driver supports pixel center integer */
3705 ureg_property_fs_coord_pixel_center(ureg
, TGSI_FS_COORD_PIXEL_CENTER_INTEGER
);
3706 else if (pscreen
->get_param(pscreen
, PIPE_CAP_TGSI_FS_COORD_PIXEL_CENTER_HALF_INTEGER
))
3707 /* the driver supports pixel center half integer, need to bias X,Y */
3708 emit_adjusted_wpos(t
, program
, 0.5f
, invert
? 0.5f
: -0.5f
);
3713 /* Fragment shader wants pixel center half integer */
3714 if (pscreen
->get_param(pscreen
, PIPE_CAP_TGSI_FS_COORD_PIXEL_CENTER_HALF_INTEGER
)) {
3715 /* the driver supports pixel center half integer */
3717 else if (pscreen
->get_param(pscreen
, PIPE_CAP_TGSI_FS_COORD_PIXEL_CENTER_INTEGER
)) {
3718 /* the driver supports pixel center integer, need to bias X,Y */
3719 ureg_property_fs_coord_pixel_center(ureg
, TGSI_FS_COORD_PIXEL_CENTER_INTEGER
);
3720 emit_adjusted_wpos(t
, program
, 0.5f
, invert
? -0.5f
: 0.5f
);
3726 /* we invert after adjustment so that we avoid the MOV to temporary,
3727 * and reuse the adjustment ADD instead */
3728 emit_wpos_inversion(t
, program
, invert
);
3732 * Translate intermediate IR (glsl_to_tgsi_instruction) to TGSI format.
3733 * \param program the program to translate
3734 * \param numInputs number of input registers used
3735 * \param inputMapping maps Mesa fragment program inputs to TGSI generic
3737 * \param inputSemanticName the TGSI_SEMANTIC flag for each input
3738 * \param inputSemanticIndex the semantic index (ex: which texcoord) for
3740 * \param interpMode the TGSI_INTERPOLATE_LINEAR/PERSP mode for each input
3741 * \param numOutputs number of output registers used
3742 * \param outputMapping maps Mesa fragment program outputs to TGSI
3744 * \param outputSemanticName the TGSI_SEMANTIC flag for each output
3745 * \param outputSemanticIndex the semantic index (ex: which texcoord) for
3748 * \return PIPE_OK or PIPE_ERROR_OUT_OF_MEMORY
3750 extern "C" enum pipe_error
3751 st_translate_program(
3752 struct gl_context
*ctx
,
3754 struct ureg_program
*ureg
,
3755 glsl_to_tgsi_visitor
*program
,
3756 const struct gl_program
*proginfo
,
3758 const GLuint inputMapping
[],
3759 const ubyte inputSemanticName
[],
3760 const ubyte inputSemanticIndex
[],
3761 const GLuint interpMode
[],
3763 const GLuint outputMapping
[],
3764 const ubyte outputSemanticName
[],
3765 const ubyte outputSemanticIndex
[],
3766 boolean passthrough_edgeflags
)
3768 struct st_translate translate
, *t
;
3770 enum pipe_error ret
= PIPE_OK
;
3772 assert(numInputs
<= Elements(t
->inputs
));
3773 assert(numOutputs
<= Elements(t
->outputs
));
3776 memset(t
, 0, sizeof *t
);
3778 t
->procType
= procType
;
3779 t
->inputMapping
= inputMapping
;
3780 t
->outputMapping
= outputMapping
;
3782 t
->pointSizeOutIndex
= -1;
3783 t
->prevInstWrotePointSize
= GL_FALSE
;
3786 * Declare input attributes.
3788 if (procType
== TGSI_PROCESSOR_FRAGMENT
) {
3789 for (i
= 0; i
< numInputs
; i
++) {
3790 t
->inputs
[i
] = ureg_DECL_fs_input(ureg
,
3791 inputSemanticName
[i
],
3792 inputSemanticIndex
[i
],
3796 if (proginfo
->InputsRead
& FRAG_BIT_WPOS
) {
3797 /* Must do this after setting up t->inputs, and before
3798 * emitting constant references, below:
3800 printf("FRAG_BIT_WPOS\n");
3801 emit_wpos(st_context(ctx
), t
, proginfo
, ureg
);
3804 if (proginfo
->InputsRead
& FRAG_BIT_FACE
) {
3806 printf("FRAG_BIT_FACE\n");
3807 //emit_face_var( t, program );
3811 * Declare output attributes.
3813 for (i
= 0; i
< numOutputs
; i
++) {
3814 switch (outputSemanticName
[i
]) {
3815 case TGSI_SEMANTIC_POSITION
:
3816 t
->outputs
[i
] = ureg_DECL_output( ureg
,
3817 TGSI_SEMANTIC_POSITION
, /* Z / Depth */
3818 outputSemanticIndex
[i
] );
3820 t
->outputs
[i
] = ureg_writemask( t
->outputs
[i
],
3823 case TGSI_SEMANTIC_STENCIL
:
3824 t
->outputs
[i
] = ureg_DECL_output( ureg
,
3825 TGSI_SEMANTIC_STENCIL
, /* Stencil */
3826 outputSemanticIndex
[i
] );
3827 t
->outputs
[i
] = ureg_writemask( t
->outputs
[i
],
3830 case TGSI_SEMANTIC_COLOR
:
3831 t
->outputs
[i
] = ureg_DECL_output( ureg
,
3832 TGSI_SEMANTIC_COLOR
,
3833 outputSemanticIndex
[i
] );
3837 return PIPE_ERROR_BAD_INPUT
;
3841 else if (procType
== TGSI_PROCESSOR_GEOMETRY
) {
3842 for (i
= 0; i
< numInputs
; i
++) {
3843 t
->inputs
[i
] = ureg_DECL_gs_input(ureg
,
3845 inputSemanticName
[i
],
3846 inputSemanticIndex
[i
]);
3849 for (i
= 0; i
< numOutputs
; i
++) {
3850 t
->outputs
[i
] = ureg_DECL_output( ureg
,
3851 outputSemanticName
[i
],
3852 outputSemanticIndex
[i
] );
3856 assert(procType
== TGSI_PROCESSOR_VERTEX
);
3858 for (i
= 0; i
< numInputs
; i
++) {
3859 t
->inputs
[i
] = ureg_DECL_vs_input(ureg
, i
);
3862 for (i
= 0; i
< numOutputs
; i
++) {
3863 t
->outputs
[i
] = ureg_DECL_output( ureg
,
3864 outputSemanticName
[i
],
3865 outputSemanticIndex
[i
] );
3866 if ((outputSemanticName
[i
] == TGSI_SEMANTIC_PSIZE
) && proginfo
->Id
) {
3867 /* Writing to the point size result register requires special
3868 * handling to implement clamping.
3870 static const gl_state_index pointSizeClampState
[STATE_LENGTH
]
3871 = { STATE_INTERNAL
, STATE_POINT_SIZE_IMPL_CLAMP
, (gl_state_index
)0, (gl_state_index
)0, (gl_state_index
)0 };
3872 /* XXX: note we are modifying the incoming shader here! Need to
3873 * do this before emitting the constant decls below, or this
3875 * XXX: depends on "Parameters" field specific to Mesa IR
3877 unsigned pointSizeClampConst
=
3878 _mesa_add_state_reference(proginfo
->Parameters
,
3879 pointSizeClampState
);
3880 struct ureg_dst psizregtemp
= ureg_DECL_temporary( ureg
);
3881 t
->pointSizeConst
= ureg_DECL_constant( ureg
, pointSizeClampConst
);
3882 t
->pointSizeResult
= t
->outputs
[i
];
3883 t
->pointSizeOutIndex
= i
;
3884 t
->outputs
[i
] = psizregtemp
;
3887 /*if (passthrough_edgeflags)
3888 emit_edgeflags( t, program ); */ // TODO: uncomment
3891 /* Declare address register.
3893 if (program
->num_address_regs
> 0) {
3894 debug_assert( program
->num_address_regs
== 1 );
3895 t
->address
[0] = ureg_DECL_address( ureg
);
3898 /* Declare misc input registers
3901 GLbitfield sysInputs
= proginfo
->SystemValuesRead
;
3902 unsigned numSys
= 0;
3903 for (i
= 0; sysInputs
; i
++) {
3904 if (sysInputs
& (1 << i
)) {
3905 unsigned semName
= mesa_sysval_to_semantic
[i
];
3906 t
->systemValues
[i
] = ureg_DECL_system_value(ureg
, numSys
, semName
, 0);
3908 sysInputs
&= ~(1 << i
);
3913 if (program
->indirect_addr_temps
) {
3914 /* If temps are accessed with indirect addressing, declare temporaries
3915 * in sequential order. Else, we declare them on demand elsewhere.
3916 * (Note: the number of temporaries is equal to program->next_temp)
3918 for (i
= 0; i
< (unsigned)program
->next_temp
; i
++) {
3919 /* XXX use TGSI_FILE_TEMPORARY_ARRAY when it's supported by ureg */
3920 t
->temps
[i
] = ureg_DECL_temporary( t
->ureg
);
3924 /* Emit constants and immediates. Mesa uses a single index space
3925 * for these, so we put all the translated regs in t->constants.
3926 * XXX: this entire if block depends on proginfo->Parameters from Mesa IR
3928 if (proginfo
->Parameters
) {
3929 t
->constants
= (struct ureg_src
*)CALLOC( proginfo
->Parameters
->NumParameters
* sizeof t
->constants
[0] );
3930 if (t
->constants
== NULL
) {
3931 ret
= PIPE_ERROR_OUT_OF_MEMORY
;
3935 for (i
= 0; i
< proginfo
->Parameters
->NumParameters
; i
++) {
3936 switch (proginfo
->Parameters
->Parameters
[i
].Type
) {
3937 case PROGRAM_ENV_PARAM
:
3938 case PROGRAM_LOCAL_PARAM
:
3939 case PROGRAM_STATE_VAR
:
3940 case PROGRAM_NAMED_PARAM
:
3941 case PROGRAM_UNIFORM
:
3942 t
->constants
[i
] = ureg_DECL_constant( ureg
, i
);
3945 /* Emit immediates only when there's no indirect addressing of
3947 * FIXME: Be smarter and recognize param arrays:
3948 * indirect addressing is only valid within the referenced
3951 case PROGRAM_CONSTANT
:
3952 if (program
->indirect_addr_consts
)
3953 t
->constants
[i
] = ureg_DECL_constant( ureg
, i
);
3955 switch(proginfo
->Parameters
->Parameters
[i
].DataType
)
3961 t
->constants
[i
] = ureg_DECL_immediate(ureg
, (float *)proginfo
->Parameters
->ParameterValues
[i
], 4);
3967 t
->constants
[i
] = ureg_DECL_immediate_int(ureg
, (int *)proginfo
->Parameters
->ParameterValues
[i
], 4);
3969 case GL_UNSIGNED_INT
:
3970 case GL_UNSIGNED_INT_VEC2
:
3971 case GL_UNSIGNED_INT_VEC3
:
3972 case GL_UNSIGNED_INT_VEC4
:
3977 t
->constants
[i
] = ureg_DECL_immediate_uint(ureg
, (unsigned *)proginfo
->Parameters
->ParameterValues
[i
], 4);
3980 assert(!"should not get here");
3989 /* texture samplers */
3990 for (i
= 0; i
< ctx
->Const
.MaxTextureImageUnits
; i
++) {
3991 if (program
->samplers_used
& (1 << i
)) {
3992 t
->samplers
[i
] = ureg_DECL_sampler( ureg
, i
);
3996 /* Emit each instruction in turn:
3998 foreach_iter(exec_list_iterator
, iter
, program
->instructions
) {
3999 set_insn_start( t
, ureg_get_instruction_number( ureg
));
4000 compile_tgsi_instruction( t
, (glsl_to_tgsi_instruction
*)iter
.get() );
4002 if (t
->prevInstWrotePointSize
&& proginfo
->Id
) {
4003 /* The previous instruction wrote to the (fake) vertex point size
4004 * result register. Now we need to clamp that value to the min/max
4005 * point size range, putting the result into the real point size
4007 * Note that we can't do this easily at the end of program due to
4008 * possible early return.
4010 set_insn_start( t
, ureg_get_instruction_number( ureg
));
4012 ureg_writemask(t
->outputs
[t
->pointSizeOutIndex
], WRITEMASK_X
),
4013 ureg_src(t
->outputs
[t
->pointSizeOutIndex
]),
4014 ureg_swizzle(t
->pointSizeConst
, 1,1,1,1));
4015 ureg_MIN( t
->ureg
, ureg_writemask(t
->pointSizeResult
, WRITEMASK_X
),
4016 ureg_src(t
->outputs
[t
->pointSizeOutIndex
]),
4017 ureg_swizzle(t
->pointSizeConst
, 2,2,2,2));
4019 t
->prevInstWrotePointSize
= GL_FALSE
;
4022 /* Fix up all emitted labels:
4024 for (i
= 0; i
< t
->labels_count
; i
++) {
4025 ureg_fixup_label( ureg
,
4027 t
->insn
[t
->labels
[i
].branch_target
] );
4036 debug_printf("%s: translate error flag set\n", __FUNCTION__
);
4041 /* ----------------------------- End TGSI code ------------------------------ */
4044 * Convert a shader's GLSL IR into a Mesa gl_program, although without
4045 * generating Mesa IR.
4047 static struct gl_program
*
4048 get_mesa_program(struct gl_context
*ctx
,
4049 struct gl_shader_program
*shader_program
,
4050 struct gl_shader
*shader
)
4052 glsl_to_tgsi_visitor
* v
= new glsl_to_tgsi_visitor();
4053 struct gl_program
*prog
;
4055 const char *target_string
;
4057 struct gl_shader_compiler_options
*options
=
4058 &ctx
->ShaderCompilerOptions
[_mesa_shader_type_to_index(shader
->Type
)];
4060 switch (shader
->Type
) {
4061 case GL_VERTEX_SHADER
:
4062 target
= GL_VERTEX_PROGRAM_ARB
;
4063 target_string
= "vertex";
4065 case GL_FRAGMENT_SHADER
:
4066 target
= GL_FRAGMENT_PROGRAM_ARB
;
4067 target_string
= "fragment";
4069 case GL_GEOMETRY_SHADER
:
4070 target
= GL_GEOMETRY_PROGRAM_NV
;
4071 target_string
= "geometry";
4074 assert(!"should not be reached");
4078 validate_ir_tree(shader
->ir
);
4080 prog
= ctx
->Driver
.NewProgram(ctx
, target
, shader_program
->Name
);
4083 prog
->Parameters
= _mesa_new_parameter_list();
4084 prog
->Varying
= _mesa_new_parameter_list();
4085 prog
->Attributes
= _mesa_new_parameter_list();
4088 v
->shader_program
= shader_program
;
4089 v
->options
= options
;
4090 v
->glsl_version
= ctx
->Const
.GLSLVersion
;
4092 add_uniforms_to_parameters_list(shader_program
, shader
, prog
);
4094 /* Emit intermediate IR for main(). */
4095 visit_exec_list(shader
->ir
, v
);
4097 /* Now emit bodies for any functions that were used. */
4099 progress
= GL_FALSE
;
4101 foreach_iter(exec_list_iterator
, iter
, v
->function_signatures
) {
4102 function_entry
*entry
= (function_entry
*)iter
.get();
4104 if (!entry
->bgn_inst
) {
4105 v
->current_function
= entry
;
4107 entry
->bgn_inst
= v
->emit(NULL
, TGSI_OPCODE_BGNSUB
);
4108 entry
->bgn_inst
->function
= entry
;
4110 visit_exec_list(&entry
->sig
->body
, v
);
4112 glsl_to_tgsi_instruction
*last
;
4113 last
= (glsl_to_tgsi_instruction
*)v
->instructions
.get_tail();
4114 if (last
->op
!= TGSI_OPCODE_RET
)
4115 v
->emit(NULL
, TGSI_OPCODE_RET
);
4117 glsl_to_tgsi_instruction
*end
;
4118 end
= v
->emit(NULL
, TGSI_OPCODE_ENDSUB
);
4119 end
->function
= entry
;
4127 /* Print out some information (for debugging purposes) used by the
4128 * optimization passes. */
4129 for (i
=0; i
< v
->next_temp
; i
++) {
4130 int fr
= v
->get_first_temp_read(i
);
4131 int fw
= v
->get_first_temp_write(i
);
4132 int lr
= v
->get_last_temp_read(i
);
4133 int lw
= v
->get_last_temp_write(i
);
4135 printf("Temp %d: FR=%3d FW=%3d LR=%3d LW=%3d\n", i
, fr
, fw
, lr
, lw
);
4140 /* Remove reads to output registers, and to varyings in vertex shaders. */
4141 v
->remove_output_reads(PROGRAM_OUTPUT
);
4142 if (target
== GL_VERTEX_PROGRAM_ARB
)
4143 v
->remove_output_reads(PROGRAM_VARYING
);
4145 /* Perform optimizations on the instructions in the glsl_to_tgsi_visitor. */
4146 v
->copy_propagate();
4148 /* FIXME: These passes to optimize temporary registers don't work when there
4149 * is indirect addressing of the temporary register space. We need proper
4150 * array support so that we don't have to give up these passes in every
4151 * shader that uses arrays.
4153 if (!v
->indirect_addr_temps
) {
4154 v
->merge_registers();
4155 v
->eliminate_dead_code();
4156 v
->renumber_registers();
4159 /* Write the END instruction. */
4160 v
->emit(NULL
, TGSI_OPCODE_END
);
4162 if (ctx
->Shader
.Flags
& GLSL_DUMP
) {
4164 printf("GLSL IR for linked %s program %d:\n", target_string
,
4165 shader_program
->Name
);
4166 _mesa_print_ir(shader
->ir
, NULL
);
4171 prog
->Instructions
= NULL
;
4172 prog
->NumInstructions
= 0;
4174 do_set_program_inouts(shader
->ir
, prog
);
4175 count_resources(v
, prog
);
4177 check_resources(ctx
, shader_program
, v
, prog
);
4179 _mesa_reference_program(ctx
, &shader
->Program
, prog
);
4181 struct st_vertex_program
*stvp
;
4182 struct st_fragment_program
*stfp
;
4183 struct st_geometry_program
*stgp
;
4185 switch (shader
->Type
) {
4186 case GL_VERTEX_SHADER
:
4187 stvp
= (struct st_vertex_program
*)prog
;
4188 stvp
->glsl_to_tgsi
= v
;
4190 case GL_FRAGMENT_SHADER
:
4191 stfp
= (struct st_fragment_program
*)prog
;
4192 stfp
->glsl_to_tgsi
= v
;
4194 case GL_GEOMETRY_SHADER
:
4195 stgp
= (struct st_geometry_program
*)prog
;
4196 stgp
->glsl_to_tgsi
= v
;
4199 assert(!"should not be reached");
4209 st_new_shader(struct gl_context
*ctx
, GLuint name
, GLuint type
)
4211 struct gl_shader
*shader
;
4212 assert(type
== GL_FRAGMENT_SHADER
|| type
== GL_VERTEX_SHADER
||
4213 type
== GL_GEOMETRY_SHADER_ARB
);
4214 shader
= rzalloc(NULL
, struct gl_shader
);
4216 shader
->Type
= type
;
4217 shader
->Name
= name
;
4218 _mesa_init_shader(ctx
, shader
);
4223 struct gl_shader_program
*
4224 st_new_shader_program(struct gl_context
*ctx
, GLuint name
)
4226 struct gl_shader_program
*shProg
;
4227 shProg
= rzalloc(NULL
, struct gl_shader_program
);
4229 shProg
->Name
= name
;
4230 _mesa_init_shader_program(ctx
, shProg
);
4237 * Called via ctx->Driver.LinkShader()
4238 * This actually involves converting GLSL IR into an intermediate TGSI-like IR
4239 * with code lowering and other optimizations.
4242 st_link_shader(struct gl_context
*ctx
, struct gl_shader_program
*prog
)
4244 assert(prog
->LinkStatus
);
4246 for (unsigned i
= 0; i
< MESA_SHADER_TYPES
; i
++) {
4247 if (prog
->_LinkedShaders
[i
] == NULL
)
4251 exec_list
*ir
= prog
->_LinkedShaders
[i
]->ir
;
4252 const struct gl_shader_compiler_options
*options
=
4253 &ctx
->ShaderCompilerOptions
[_mesa_shader_type_to_index(prog
->_LinkedShaders
[i
]->Type
)];
4259 do_mat_op_to_vec(ir
);
4260 lower_instructions(ir
, (MOD_TO_FRACT
| DIV_TO_MUL_RCP
| EXP_TO_EXP2
4262 | ((options
->EmitNoPow
) ? POW_TO_EXP2
: 0)));
4264 progress
= do_lower_jumps(ir
, true, true, options
->EmitNoMainReturn
, options
->EmitNoCont
, options
->EmitNoLoops
) || progress
;
4266 progress
= do_common_optimization(ir
, true, options
->MaxUnrollIterations
) || progress
;
4268 progress
= lower_quadop_vector(ir
, true) || progress
;
4270 if (options
->EmitNoIfs
) {
4271 progress
= lower_discard(ir
) || progress
;
4272 progress
= lower_if_to_cond_assign(ir
) || progress
;
4275 if (options
->EmitNoNoise
)
4276 progress
= lower_noise(ir
) || progress
;
4278 /* If there are forms of indirect addressing that the driver
4279 * cannot handle, perform the lowering pass.
4281 if (options
->EmitNoIndirectInput
|| options
->EmitNoIndirectOutput
4282 || options
->EmitNoIndirectTemp
|| options
->EmitNoIndirectUniform
)
4284 lower_variable_index_to_cond_assign(ir
,
4285 options
->EmitNoIndirectInput
,
4286 options
->EmitNoIndirectOutput
,
4287 options
->EmitNoIndirectTemp
,
4288 options
->EmitNoIndirectUniform
)
4291 progress
= do_vec_index_to_cond_assign(ir
) || progress
;
4294 validate_ir_tree(ir
);
4297 for (unsigned i
= 0; i
< MESA_SHADER_TYPES
; i
++) {
4298 struct gl_program
*linked_prog
;
4300 if (prog
->_LinkedShaders
[i
] == NULL
)
4303 linked_prog
= get_mesa_program(ctx
, prog
, prog
->_LinkedShaders
[i
]);
4308 switch (prog
->_LinkedShaders
[i
]->Type
) {
4309 case GL_VERTEX_SHADER
:
4310 _mesa_reference_vertprog(ctx
, &prog
->VertexProgram
,
4311 (struct gl_vertex_program
*)linked_prog
);
4312 ok
= ctx
->Driver
.ProgramStringNotify(ctx
, GL_VERTEX_PROGRAM_ARB
,
4315 case GL_FRAGMENT_SHADER
:
4316 _mesa_reference_fragprog(ctx
, &prog
->FragmentProgram
,
4317 (struct gl_fragment_program
*)linked_prog
);
4318 ok
= ctx
->Driver
.ProgramStringNotify(ctx
, GL_FRAGMENT_PROGRAM_ARB
,
4321 case GL_GEOMETRY_SHADER
:
4322 _mesa_reference_geomprog(ctx
, &prog
->GeometryProgram
,
4323 (struct gl_geometry_program
*)linked_prog
);
4324 ok
= ctx
->Driver
.ProgramStringNotify(ctx
, GL_GEOMETRY_PROGRAM_NV
,
4333 _mesa_reference_program(ctx
, &linked_prog
, NULL
);
4341 * Link a GLSL shader program. Called via glLinkProgram().
4344 st_glsl_link_shader(struct gl_context
*ctx
, struct gl_shader_program
*prog
)
4348 _mesa_clear_shader_program_data(ctx
, prog
);
4350 prog
->LinkStatus
= GL_TRUE
;
4352 for (i
= 0; i
< prog
->NumShaders
; i
++) {
4353 if (!prog
->Shaders
[i
]->CompileStatus
) {
4354 fail_link(prog
, "linking with uncompiled shader");
4355 prog
->LinkStatus
= GL_FALSE
;
4359 prog
->Varying
= _mesa_new_parameter_list();
4360 _mesa_reference_vertprog(ctx
, &prog
->VertexProgram
, NULL
);
4361 _mesa_reference_fragprog(ctx
, &prog
->FragmentProgram
, NULL
);
4362 _mesa_reference_geomprog(ctx
, &prog
->GeometryProgram
, NULL
);
4364 if (prog
->LinkStatus
) {
4365 link_shaders(ctx
, prog
);
4368 if (prog
->LinkStatus
) {
4369 if (!ctx
->Driver
.LinkShader(ctx
, prog
)) {
4370 prog
->LinkStatus
= GL_FALSE
;
4374 set_uniform_initializers(ctx
, prog
);
4376 if (ctx
->Shader
.Flags
& GLSL_DUMP
) {
4377 if (!prog
->LinkStatus
) {
4378 printf("GLSL shader program %d failed to link\n", prog
->Name
);
4381 if (prog
->InfoLog
&& prog
->InfoLog
[0] != 0) {
4382 printf("GLSL shader program %d info log:\n", prog
->Name
);
4383 printf("%s\n", prog
->InfoLog
);