2 * Copyright (C) 2005-2007 Brian Paul All Rights Reserved.
3 * Copyright (C) 2008 VMware, Inc. All Rights Reserved.
4 * Copyright © 2010 Intel Corporation
6 * Permission is hereby granted, free of charge, to any person obtaining a
7 * copy of this software and associated documentation files (the "Software"),
8 * to deal in the Software without restriction, including without limitation
9 * the rights to use, copy, modify, merge, publish, distribute, sublicense,
10 * and/or sell copies of the Software, and to permit persons to whom the
11 * Software is furnished to do so, subject to the following conditions:
13 * The above copyright notice and this permission notice (including the next
14 * paragraph) shall be included in all copies or substantial portions of the
17 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
18 * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
19 * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL
20 * THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
21 * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING
22 * FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER
23 * DEALINGS IN THE SOFTWARE.
27 * \file ir_to_mesa.cpp
29 * Translate GLSL IR to Mesa's gl_program representation.
33 #include "main/compiler.h"
35 #include "ir_visitor.h"
36 #include "ir_expression_flattening.h"
37 #include "ir_uniform.h"
38 #include "glsl_types.h"
39 #include "glsl_parser_extras.h"
40 #include "../glsl/program.h"
41 #include "ir_optimization.h"
45 #include "main/mtypes.h"
46 #include "main/shaderobj.h"
47 #include "main/uniforms.h"
48 #include "program/hash_table.h"
51 #include "main/shaderapi.h"
52 #include "program/prog_instruction.h"
53 #include "program/prog_optimize.h"
54 #include "program/prog_print.h"
55 #include "program/program.h"
56 #include "program/prog_parameter.h"
57 #include "program/sampler.h"
60 static int swizzle_for_size(int size
);
68 * This struct is a corresponding struct to Mesa prog_src_register, with
73 src_reg(gl_register_file file
, int index
, const glsl_type
*type
)
77 if (type
&& (type
->is_scalar() || type
->is_vector() || type
->is_matrix()))
78 this->swizzle
= swizzle_for_size(type
->vector_elements
);
80 this->swizzle
= SWIZZLE_XYZW
;
87 this->file
= PROGRAM_UNDEFINED
;
94 explicit src_reg(dst_reg reg
);
96 gl_register_file file
; /**< PROGRAM_* from Mesa */
97 int index
; /**< temporary index, VERT_ATTRIB_*, VARYING_SLOT_*, etc. */
98 GLuint swizzle
; /**< SWIZZLE_XYZWONEZERO swizzles from Mesa. */
99 int negate
; /**< NEGATE_XYZW mask from mesa */
100 /** Register index should be offset by the integer in this reg. */
106 dst_reg(gl_register_file file
, int writemask
)
110 this->writemask
= writemask
;
111 this->cond_mask
= COND_TR
;
112 this->reladdr
= NULL
;
117 this->file
= PROGRAM_UNDEFINED
;
120 this->cond_mask
= COND_TR
;
121 this->reladdr
= NULL
;
124 explicit dst_reg(src_reg reg
);
126 gl_register_file file
; /**< PROGRAM_* from Mesa */
127 int index
; /**< temporary index, VERT_ATTRIB_*, VARYING_SLOT_*, etc. */
128 int writemask
; /**< Bitfield of WRITEMASK_[XYZW] */
130 /** Register index should be offset by the integer in this reg. */
134 } /* anonymous namespace */
136 src_reg::src_reg(dst_reg reg
)
138 this->file
= reg
.file
;
139 this->index
= reg
.index
;
140 this->swizzle
= SWIZZLE_XYZW
;
142 this->reladdr
= reg
.reladdr
;
145 dst_reg::dst_reg(src_reg reg
)
147 this->file
= reg
.file
;
148 this->index
= reg
.index
;
149 this->writemask
= WRITEMASK_XYZW
;
150 this->cond_mask
= COND_TR
;
151 this->reladdr
= reg
.reladdr
;
156 class ir_to_mesa_instruction
: public exec_node
{
158 DECLARE_RALLOC_CXX_OPERATORS(ir_to_mesa_instruction
)
163 /** Pointer to the ir source this tree came from for debugging */
165 GLboolean cond_update
;
167 int sampler
; /**< sampler index */
168 int tex_target
; /**< One of TEXTURE_*_INDEX */
169 GLboolean tex_shadow
;
172 class variable_storage
: public exec_node
{
174 variable_storage(ir_variable
*var
, gl_register_file file
, int index
)
175 : file(file
), index(index
), var(var
)
180 gl_register_file file
;
182 ir_variable
*var
; /* variable that maps to this, if any */
185 class function_entry
: public exec_node
{
187 ir_function_signature
*sig
;
190 * identifier of this function signature used by the program.
192 * At the point that Mesa instructions for function calls are
193 * generated, we don't know the address of the first instruction of
194 * the function body. So we make the BranchTarget that is called a
195 * small integer and rewrite them during set_branchtargets().
200 * Pointer to first instruction of the function body.
202 * Set during function body emits after main() is processed.
204 ir_to_mesa_instruction
*bgn_inst
;
207 * Index of the first instruction of the function body in actual
210 * Set after convertion from ir_to_mesa_instruction to prog_instruction.
214 /** Storage for the return value. */
218 class ir_to_mesa_visitor
: public ir_visitor
{
220 ir_to_mesa_visitor();
221 ~ir_to_mesa_visitor();
223 function_entry
*current_function
;
225 struct gl_context
*ctx
;
226 struct gl_program
*prog
;
227 struct gl_shader_program
*shader_program
;
228 struct gl_shader_compiler_options
*options
;
232 variable_storage
*find_variable_storage(const ir_variable
*var
);
234 src_reg
get_temp(const glsl_type
*type
);
235 void reladdr_to_temp(ir_instruction
*ir
, src_reg
*reg
, int *num_reladdr
);
237 src_reg
src_reg_for_float(float val
);
240 * \name Visit methods
242 * As typical for the visitor pattern, there must be one \c visit method for
243 * each concrete subclass of \c ir_instruction. Virtual base classes within
244 * the hierarchy should not have \c visit methods.
247 virtual void visit(ir_variable
*);
248 virtual void visit(ir_loop
*);
249 virtual void visit(ir_loop_jump
*);
250 virtual void visit(ir_function_signature
*);
251 virtual void visit(ir_function
*);
252 virtual void visit(ir_expression
*);
253 virtual void visit(ir_swizzle
*);
254 virtual void visit(ir_dereference_variable
*);
255 virtual void visit(ir_dereference_array
*);
256 virtual void visit(ir_dereference_record
*);
257 virtual void visit(ir_assignment
*);
258 virtual void visit(ir_constant
*);
259 virtual void visit(ir_call
*);
260 virtual void visit(ir_return
*);
261 virtual void visit(ir_discard
*);
262 virtual void visit(ir_texture
*);
263 virtual void visit(ir_if
*);
264 virtual void visit(ir_emit_vertex
*);
265 virtual void visit(ir_end_primitive
*);
270 /** List of variable_storage */
273 /** List of function_entry */
274 exec_list function_signatures
;
275 int next_signature_id
;
277 /** List of ir_to_mesa_instruction */
278 exec_list instructions
;
280 ir_to_mesa_instruction
*emit(ir_instruction
*ir
, enum prog_opcode op
);
282 ir_to_mesa_instruction
*emit(ir_instruction
*ir
, enum prog_opcode op
,
283 dst_reg dst
, src_reg src0
);
285 ir_to_mesa_instruction
*emit(ir_instruction
*ir
, enum prog_opcode op
,
286 dst_reg dst
, src_reg src0
, src_reg src1
);
288 ir_to_mesa_instruction
*emit(ir_instruction
*ir
, enum prog_opcode op
,
290 src_reg src0
, src_reg src1
, src_reg src2
);
293 * Emit the correct dot-product instruction for the type of arguments
295 ir_to_mesa_instruction
* emit_dp(ir_instruction
*ir
,
301 void emit_scalar(ir_instruction
*ir
, enum prog_opcode op
,
302 dst_reg dst
, src_reg src0
);
304 void emit_scalar(ir_instruction
*ir
, enum prog_opcode op
,
305 dst_reg dst
, src_reg src0
, src_reg src1
);
307 void emit_scs(ir_instruction
*ir
, enum prog_opcode op
,
308 dst_reg dst
, const src_reg
&src
);
310 bool try_emit_mad(ir_expression
*ir
,
312 bool try_emit_mad_for_and_not(ir_expression
*ir
,
314 bool try_emit_sat(ir_expression
*ir
);
316 void emit_swz(ir_expression
*ir
);
318 bool process_move_condition(ir_rvalue
*ir
);
320 void copy_propagate(void);
325 } /* anonymous namespace */
327 static src_reg undef_src
= src_reg(PROGRAM_UNDEFINED
, 0, NULL
);
329 static dst_reg undef_dst
= dst_reg(PROGRAM_UNDEFINED
, SWIZZLE_NOOP
);
331 static dst_reg address_reg
= dst_reg(PROGRAM_ADDRESS
, WRITEMASK_X
);
334 swizzle_for_size(int size
)
336 static const int size_swizzles
[4] = {
337 MAKE_SWIZZLE4(SWIZZLE_X
, SWIZZLE_X
, SWIZZLE_X
, SWIZZLE_X
),
338 MAKE_SWIZZLE4(SWIZZLE_X
, SWIZZLE_Y
, SWIZZLE_Y
, SWIZZLE_Y
),
339 MAKE_SWIZZLE4(SWIZZLE_X
, SWIZZLE_Y
, SWIZZLE_Z
, SWIZZLE_Z
),
340 MAKE_SWIZZLE4(SWIZZLE_X
, SWIZZLE_Y
, SWIZZLE_Z
, SWIZZLE_W
),
343 assert((size
>= 1) && (size
<= 4));
344 return size_swizzles
[size
- 1];
347 ir_to_mesa_instruction
*
348 ir_to_mesa_visitor::emit(ir_instruction
*ir
, enum prog_opcode op
,
350 src_reg src0
, src_reg src1
, src_reg src2
)
352 ir_to_mesa_instruction
*inst
= new(mem_ctx
) ir_to_mesa_instruction();
355 /* If we have to do relative addressing, we want to load the ARL
356 * reg directly for one of the regs, and preload the other reladdr
357 * sources into temps.
359 num_reladdr
+= dst
.reladdr
!= NULL
;
360 num_reladdr
+= src0
.reladdr
!= NULL
;
361 num_reladdr
+= src1
.reladdr
!= NULL
;
362 num_reladdr
+= src2
.reladdr
!= NULL
;
364 reladdr_to_temp(ir
, &src2
, &num_reladdr
);
365 reladdr_to_temp(ir
, &src1
, &num_reladdr
);
366 reladdr_to_temp(ir
, &src0
, &num_reladdr
);
369 emit(ir
, OPCODE_ARL
, address_reg
, *dst
.reladdr
);
372 assert(num_reladdr
== 0);
381 this->instructions
.push_tail(inst
);
387 ir_to_mesa_instruction
*
388 ir_to_mesa_visitor::emit(ir_instruction
*ir
, enum prog_opcode op
,
389 dst_reg dst
, src_reg src0
, src_reg src1
)
391 return emit(ir
, op
, dst
, src0
, src1
, undef_src
);
394 ir_to_mesa_instruction
*
395 ir_to_mesa_visitor::emit(ir_instruction
*ir
, enum prog_opcode op
,
396 dst_reg dst
, src_reg src0
)
398 assert(dst
.writemask
!= 0);
399 return emit(ir
, op
, dst
, src0
, undef_src
, undef_src
);
402 ir_to_mesa_instruction
*
403 ir_to_mesa_visitor::emit(ir_instruction
*ir
, enum prog_opcode op
)
405 return emit(ir
, op
, undef_dst
, undef_src
, undef_src
, undef_src
);
408 ir_to_mesa_instruction
*
409 ir_to_mesa_visitor::emit_dp(ir_instruction
*ir
,
410 dst_reg dst
, src_reg src0
, src_reg src1
,
413 static const gl_inst_opcode dot_opcodes
[] = {
414 OPCODE_DP2
, OPCODE_DP3
, OPCODE_DP4
417 return emit(ir
, dot_opcodes
[elements
- 2], dst
, src0
, src1
);
421 * Emits Mesa scalar opcodes to produce unique answers across channels.
423 * Some Mesa opcodes are scalar-only, like ARB_fp/vp. The src X
424 * channel determines the result across all channels. So to do a vec4
425 * of this operation, we want to emit a scalar per source channel used
426 * to produce dest channels.
429 ir_to_mesa_visitor::emit_scalar(ir_instruction
*ir
, enum prog_opcode op
,
431 src_reg orig_src0
, src_reg orig_src1
)
434 int done_mask
= ~dst
.writemask
;
436 /* Mesa RCP is a scalar operation splatting results to all channels,
437 * like ARB_fp/vp. So emit as many RCPs as necessary to cover our
440 for (i
= 0; i
< 4; i
++) {
441 GLuint this_mask
= (1 << i
);
442 ir_to_mesa_instruction
*inst
;
443 src_reg src0
= orig_src0
;
444 src_reg src1
= orig_src1
;
446 if (done_mask
& this_mask
)
449 GLuint src0_swiz
= GET_SWZ(src0
.swizzle
, i
);
450 GLuint src1_swiz
= GET_SWZ(src1
.swizzle
, i
);
451 for (j
= i
+ 1; j
< 4; j
++) {
452 /* If there is another enabled component in the destination that is
453 * derived from the same inputs, generate its value on this pass as
456 if (!(done_mask
& (1 << j
)) &&
457 GET_SWZ(src0
.swizzle
, j
) == src0_swiz
&&
458 GET_SWZ(src1
.swizzle
, j
) == src1_swiz
) {
459 this_mask
|= (1 << j
);
462 src0
.swizzle
= MAKE_SWIZZLE4(src0_swiz
, src0_swiz
,
463 src0_swiz
, src0_swiz
);
464 src1
.swizzle
= MAKE_SWIZZLE4(src1_swiz
, src1_swiz
,
465 src1_swiz
, src1_swiz
);
467 inst
= emit(ir
, op
, dst
, src0
, src1
);
468 inst
->dst
.writemask
= this_mask
;
469 done_mask
|= this_mask
;
474 ir_to_mesa_visitor::emit_scalar(ir_instruction
*ir
, enum prog_opcode op
,
475 dst_reg dst
, src_reg src0
)
477 src_reg undef
= undef_src
;
479 undef
.swizzle
= SWIZZLE_XXXX
;
481 emit_scalar(ir
, op
, dst
, src0
, undef
);
485 * Emit an OPCODE_SCS instruction
487 * The \c SCS opcode functions a bit differently than the other Mesa (or
488 * ARB_fragment_program) opcodes. Instead of splatting its result across all
489 * four components of the destination, it writes one value to the \c x
490 * component and another value to the \c y component.
492 * \param ir IR instruction being processed
493 * \param op Either \c OPCODE_SIN or \c OPCODE_COS depending on which
495 * \param dst Destination register
496 * \param src Source register
499 ir_to_mesa_visitor::emit_scs(ir_instruction
*ir
, enum prog_opcode op
,
503 /* Vertex programs cannot use the SCS opcode.
505 if (this->prog
->Target
== GL_VERTEX_PROGRAM_ARB
) {
506 emit_scalar(ir
, op
, dst
, src
);
510 const unsigned component
= (op
== OPCODE_SIN
) ? 0 : 1;
511 const unsigned scs_mask
= (1U << component
);
512 int done_mask
= ~dst
.writemask
;
515 assert(op
== OPCODE_SIN
|| op
== OPCODE_COS
);
517 /* If there are compnents in the destination that differ from the component
518 * that will be written by the SCS instrution, we'll need a temporary.
520 if (scs_mask
!= unsigned(dst
.writemask
)) {
521 tmp
= get_temp(glsl_type::vec4_type
);
524 for (unsigned i
= 0; i
< 4; i
++) {
525 unsigned this_mask
= (1U << i
);
528 if ((done_mask
& this_mask
) != 0)
531 /* The source swizzle specified which component of the source generates
532 * sine / cosine for the current component in the destination. The SCS
533 * instruction requires that this value be swizzle to the X component.
534 * Replace the current swizzle with a swizzle that puts the source in
537 unsigned src0_swiz
= GET_SWZ(src
.swizzle
, i
);
539 src0
.swizzle
= MAKE_SWIZZLE4(src0_swiz
, src0_swiz
,
540 src0_swiz
, src0_swiz
);
541 for (unsigned j
= i
+ 1; j
< 4; j
++) {
542 /* If there is another enabled component in the destination that is
543 * derived from the same inputs, generate its value on this pass as
546 if (!(done_mask
& (1 << j
)) &&
547 GET_SWZ(src0
.swizzle
, j
) == src0_swiz
) {
548 this_mask
|= (1 << j
);
552 if (this_mask
!= scs_mask
) {
553 ir_to_mesa_instruction
*inst
;
554 dst_reg tmp_dst
= dst_reg(tmp
);
556 /* Emit the SCS instruction.
558 inst
= emit(ir
, OPCODE_SCS
, tmp_dst
, src0
);
559 inst
->dst
.writemask
= scs_mask
;
561 /* Move the result of the SCS instruction to the desired location in
564 tmp
.swizzle
= MAKE_SWIZZLE4(component
, component
,
565 component
, component
);
566 inst
= emit(ir
, OPCODE_SCS
, dst
, tmp
);
567 inst
->dst
.writemask
= this_mask
;
569 /* Emit the SCS instruction to write directly to the destination.
571 ir_to_mesa_instruction
*inst
= emit(ir
, OPCODE_SCS
, dst
, src0
);
572 inst
->dst
.writemask
= scs_mask
;
575 done_mask
|= this_mask
;
580 ir_to_mesa_visitor::src_reg_for_float(float val
)
582 src_reg
src(PROGRAM_CONSTANT
, -1, NULL
);
584 src
.index
= _mesa_add_unnamed_constant(this->prog
->Parameters
,
585 (const gl_constant_value
*)&val
, 1, &src
.swizzle
);
591 type_size(const struct glsl_type
*type
)
596 switch (type
->base_type
) {
599 case GLSL_TYPE_FLOAT
:
601 if (type
->is_matrix()) {
602 return type
->matrix_columns
;
604 /* Regardless of size of vector, it gets a vec4. This is bad
605 * packing for things like floats, but otherwise arrays become a
606 * mess. Hopefully a later pass over the code can pack scalars
607 * down if appropriate.
611 case GLSL_TYPE_ARRAY
:
612 assert(type
->length
> 0);
613 return type_size(type
->fields
.array
) * type
->length
;
614 case GLSL_TYPE_STRUCT
:
616 for (i
= 0; i
< type
->length
; i
++) {
617 size
+= type_size(type
->fields
.structure
[i
].type
);
620 case GLSL_TYPE_SAMPLER
:
621 case GLSL_TYPE_IMAGE
:
622 /* Samplers take up one slot in UNIFORMS[], but they're baked in
626 case GLSL_TYPE_ATOMIC_UINT
:
628 case GLSL_TYPE_ERROR
:
629 case GLSL_TYPE_INTERFACE
:
630 assert(!"Invalid type in type_size");
638 * In the initial pass of codegen, we assign temporary numbers to
639 * intermediate results. (not SSA -- variable assignments will reuse
640 * storage). Actual register allocation for the Mesa VM occurs in a
641 * pass over the Mesa IR later.
644 ir_to_mesa_visitor::get_temp(const glsl_type
*type
)
648 src
.file
= PROGRAM_TEMPORARY
;
649 src
.index
= next_temp
;
651 next_temp
+= type_size(type
);
653 if (type
->is_array() || type
->is_record()) {
654 src
.swizzle
= SWIZZLE_NOOP
;
656 src
.swizzle
= swizzle_for_size(type
->vector_elements
);
664 ir_to_mesa_visitor::find_variable_storage(const ir_variable
*var
)
666 variable_storage
*entry
;
668 foreach_list(node
, &this->variables
) {
669 entry
= (variable_storage
*) node
;
671 if (entry
->var
== var
)
679 ir_to_mesa_visitor::visit(ir_variable
*ir
)
681 if (strcmp(ir
->name
, "gl_FragCoord") == 0) {
682 struct gl_fragment_program
*fp
= (struct gl_fragment_program
*)this->prog
;
684 fp
->OriginUpperLeft
= ir
->data
.origin_upper_left
;
685 fp
->PixelCenterInteger
= ir
->data
.pixel_center_integer
;
688 if (ir
->data
.mode
== ir_var_uniform
&& strncmp(ir
->name
, "gl_", 3) == 0) {
690 const ir_state_slot
*const slots
= ir
->state_slots
;
691 assert(ir
->state_slots
!= NULL
);
693 /* Check if this statevar's setup in the STATE file exactly
694 * matches how we'll want to reference it as a
695 * struct/array/whatever. If not, then we need to move it into
696 * temporary storage and hope that it'll get copy-propagated
699 for (i
= 0; i
< ir
->num_state_slots
; i
++) {
700 if (slots
[i
].swizzle
!= SWIZZLE_XYZW
) {
705 variable_storage
*storage
;
707 if (i
== ir
->num_state_slots
) {
708 /* We'll set the index later. */
709 storage
= new(mem_ctx
) variable_storage(ir
, PROGRAM_STATE_VAR
, -1);
710 this->variables
.push_tail(storage
);
714 /* The variable_storage constructor allocates slots based on the size
715 * of the type. However, this had better match the number of state
716 * elements that we're going to copy into the new temporary.
718 assert((int) ir
->num_state_slots
== type_size(ir
->type
));
720 storage
= new(mem_ctx
) variable_storage(ir
, PROGRAM_TEMPORARY
,
722 this->variables
.push_tail(storage
);
723 this->next_temp
+= type_size(ir
->type
);
725 dst
= dst_reg(src_reg(PROGRAM_TEMPORARY
, storage
->index
, NULL
));
729 for (unsigned int i
= 0; i
< ir
->num_state_slots
; i
++) {
730 int index
= _mesa_add_state_reference(this->prog
->Parameters
,
731 (gl_state_index
*)slots
[i
].tokens
);
733 if (storage
->file
== PROGRAM_STATE_VAR
) {
734 if (storage
->index
== -1) {
735 storage
->index
= index
;
737 assert(index
== storage
->index
+ (int)i
);
740 src_reg
src(PROGRAM_STATE_VAR
, index
, NULL
);
741 src
.swizzle
= slots
[i
].swizzle
;
742 emit(ir
, OPCODE_MOV
, dst
, src
);
743 /* even a float takes up a whole vec4 reg in a struct/array. */
748 if (storage
->file
== PROGRAM_TEMPORARY
&&
749 dst
.index
!= storage
->index
+ (int) ir
->num_state_slots
) {
750 linker_error(this->shader_program
,
751 "failed to load builtin uniform `%s' "
752 "(%d/%d regs loaded)\n",
753 ir
->name
, dst
.index
- storage
->index
,
754 type_size(ir
->type
));
760 ir_to_mesa_visitor::visit(ir_loop
*ir
)
762 emit(NULL
, OPCODE_BGNLOOP
);
764 visit_exec_list(&ir
->body_instructions
, this);
766 emit(NULL
, OPCODE_ENDLOOP
);
770 ir_to_mesa_visitor::visit(ir_loop_jump
*ir
)
773 case ir_loop_jump::jump_break
:
774 emit(NULL
, OPCODE_BRK
);
776 case ir_loop_jump::jump_continue
:
777 emit(NULL
, OPCODE_CONT
);
784 ir_to_mesa_visitor::visit(ir_function_signature
*ir
)
791 ir_to_mesa_visitor::visit(ir_function
*ir
)
793 /* Ignore function bodies other than main() -- we shouldn't see calls to
794 * them since they should all be inlined before we get to ir_to_mesa.
796 if (strcmp(ir
->name
, "main") == 0) {
797 const ir_function_signature
*sig
;
800 sig
= ir
->matching_signature(NULL
, &empty
);
804 foreach_list(node
, &sig
->body
) {
805 ir_instruction
*ir
= (ir_instruction
*) node
;
813 ir_to_mesa_visitor::try_emit_mad(ir_expression
*ir
, int mul_operand
)
815 int nonmul_operand
= 1 - mul_operand
;
818 ir_expression
*expr
= ir
->operands
[mul_operand
]->as_expression();
819 if (!expr
|| expr
->operation
!= ir_binop_mul
)
822 expr
->operands
[0]->accept(this);
824 expr
->operands
[1]->accept(this);
826 ir
->operands
[nonmul_operand
]->accept(this);
829 this->result
= get_temp(ir
->type
);
830 emit(ir
, OPCODE_MAD
, dst_reg(this->result
), a
, b
, c
);
836 * Emit OPCODE_MAD(a, -b, a) instead of AND(a, NOT(b))
838 * The logic values are 1.0 for true and 0.0 for false. Logical-and is
839 * implemented using multiplication, and logical-or is implemented using
840 * addition. Logical-not can be implemented as (true - x), or (1.0 - x).
841 * As result, the logical expression (a & !b) can be rewritten as:
845 * - (a * 1) - (a * b)
849 * This final expression can be implemented as a single MAD(a, -b, a)
853 ir_to_mesa_visitor::try_emit_mad_for_and_not(ir_expression
*ir
, int try_operand
)
855 const int other_operand
= 1 - try_operand
;
858 ir_expression
*expr
= ir
->operands
[try_operand
]->as_expression();
859 if (!expr
|| expr
->operation
!= ir_unop_logic_not
)
862 ir
->operands
[other_operand
]->accept(this);
864 expr
->operands
[0]->accept(this);
867 b
.negate
= ~b
.negate
;
869 this->result
= get_temp(ir
->type
);
870 emit(ir
, OPCODE_MAD
, dst_reg(this->result
), a
, b
, a
);
876 ir_to_mesa_visitor::try_emit_sat(ir_expression
*ir
)
878 /* Saturates were only introduced to vertex programs in
879 * NV_vertex_program3, so don't give them to drivers in the VP.
881 if (this->prog
->Target
== GL_VERTEX_PROGRAM_ARB
)
884 ir_rvalue
*sat_src
= ir
->as_rvalue_to_saturate();
888 sat_src
->accept(this);
889 src_reg src
= this->result
;
891 /* If we generated an expression instruction into a temporary in
892 * processing the saturate's operand, apply the saturate to that
893 * instruction. Otherwise, generate a MOV to do the saturate.
895 * Note that we have to be careful to only do this optimization if
896 * the instruction in question was what generated src->result. For
897 * example, ir_dereference_array might generate a MUL instruction
898 * to create the reladdr, and return us a src reg using that
899 * reladdr. That MUL result is not the value we're trying to
902 ir_expression
*sat_src_expr
= sat_src
->as_expression();
903 ir_to_mesa_instruction
*new_inst
;
904 new_inst
= (ir_to_mesa_instruction
*)this->instructions
.get_tail();
905 if (sat_src_expr
&& (sat_src_expr
->operation
== ir_binop_mul
||
906 sat_src_expr
->operation
== ir_binop_add
||
907 sat_src_expr
->operation
== ir_binop_dot
)) {
908 new_inst
->saturate
= true;
910 this->result
= get_temp(ir
->type
);
911 ir_to_mesa_instruction
*inst
;
912 inst
= emit(ir
, OPCODE_MOV
, dst_reg(this->result
), src
);
913 inst
->saturate
= true;
920 ir_to_mesa_visitor::reladdr_to_temp(ir_instruction
*ir
,
921 src_reg
*reg
, int *num_reladdr
)
926 emit(ir
, OPCODE_ARL
, address_reg
, *reg
->reladdr
);
928 if (*num_reladdr
!= 1) {
929 src_reg temp
= get_temp(glsl_type::vec4_type
);
931 emit(ir
, OPCODE_MOV
, dst_reg(temp
), *reg
);
939 ir_to_mesa_visitor::emit_swz(ir_expression
*ir
)
941 /* Assume that the vector operator is in a form compatible with OPCODE_SWZ.
942 * This means that each of the operands is either an immediate value of -1,
943 * 0, or 1, or is a component from one source register (possibly with
946 uint8_t components
[4] = { 0 };
947 bool negate
[4] = { false };
948 ir_variable
*var
= NULL
;
950 for (unsigned i
= 0; i
< ir
->type
->vector_elements
; i
++) {
951 ir_rvalue
*op
= ir
->operands
[i
];
953 assert(op
->type
->is_scalar());
956 switch (op
->ir_type
) {
957 case ir_type_constant
: {
959 assert(op
->type
->is_scalar());
961 const ir_constant
*const c
= op
->as_constant();
963 components
[i
] = SWIZZLE_ONE
;
964 } else if (c
->is_zero()) {
965 components
[i
] = SWIZZLE_ZERO
;
966 } else if (c
->is_negative_one()) {
967 components
[i
] = SWIZZLE_ONE
;
970 assert(!"SWZ constant must be 0.0 or 1.0.");
977 case ir_type_dereference_variable
: {
978 ir_dereference_variable
*const deref
=
979 (ir_dereference_variable
*) op
;
981 assert((var
== NULL
) || (deref
->var
== var
));
982 components
[i
] = SWIZZLE_X
;
988 case ir_type_expression
: {
989 ir_expression
*const expr
= (ir_expression
*) op
;
991 assert(expr
->operation
== ir_unop_neg
);
994 op
= expr
->operands
[0];
998 case ir_type_swizzle
: {
999 ir_swizzle
*const swiz
= (ir_swizzle
*) op
;
1001 components
[i
] = swiz
->mask
.x
;
1007 assert(!"Should not get here.");
1013 assert(var
!= NULL
);
1015 ir_dereference_variable
*const deref
=
1016 new(mem_ctx
) ir_dereference_variable(var
);
1018 this->result
.file
= PROGRAM_UNDEFINED
;
1019 deref
->accept(this);
1020 if (this->result
.file
== PROGRAM_UNDEFINED
) {
1021 printf("Failed to get tree for expression operand:\n");
1030 src
.swizzle
= MAKE_SWIZZLE4(components
[0],
1034 src
.negate
= ((unsigned(negate
[0]) << 0)
1035 | (unsigned(negate
[1]) << 1)
1036 | (unsigned(negate
[2]) << 2)
1037 | (unsigned(negate
[3]) << 3));
1039 /* Storage for our result. Ideally for an assignment we'd be using the
1040 * actual storage for the result here, instead.
1042 const src_reg result_src
= get_temp(ir
->type
);
1043 dst_reg result_dst
= dst_reg(result_src
);
1045 /* Limit writes to the channels that will be used by result_src later.
1046 * This does limit this temp's use as a temporary for multi-instruction
1049 result_dst
.writemask
= (1 << ir
->type
->vector_elements
) - 1;
1051 emit(ir
, OPCODE_SWZ
, result_dst
, src
);
1052 this->result
= result_src
;
1056 ir_to_mesa_visitor::visit(ir_expression
*ir
)
1058 unsigned int operand
;
1059 src_reg op
[Elements(ir
->operands
)];
1063 /* Quick peephole: Emit OPCODE_MAD(a, b, c) instead of ADD(MUL(a, b), c)
1065 if (ir
->operation
== ir_binop_add
) {
1066 if (try_emit_mad(ir
, 1))
1068 if (try_emit_mad(ir
, 0))
1072 /* Quick peephole: Emit OPCODE_MAD(-a, -b, a) instead of AND(a, NOT(b))
1074 if (ir
->operation
== ir_binop_logic_and
) {
1075 if (try_emit_mad_for_and_not(ir
, 1))
1077 if (try_emit_mad_for_and_not(ir
, 0))
1081 if (try_emit_sat(ir
))
1084 if (ir
->operation
== ir_quadop_vector
) {
1089 for (operand
= 0; operand
< ir
->get_num_operands(); operand
++) {
1090 this->result
.file
= PROGRAM_UNDEFINED
;
1091 ir
->operands
[operand
]->accept(this);
1092 if (this->result
.file
== PROGRAM_UNDEFINED
) {
1093 printf("Failed to get tree for expression operand:\n");
1094 ir
->operands
[operand
]->print();
1098 op
[operand
] = this->result
;
1100 /* Matrix expression operands should have been broken down to vector
1101 * operations already.
1103 assert(!ir
->operands
[operand
]->type
->is_matrix());
1106 int vector_elements
= ir
->operands
[0]->type
->vector_elements
;
1107 if (ir
->operands
[1]) {
1108 vector_elements
= MAX2(vector_elements
,
1109 ir
->operands
[1]->type
->vector_elements
);
1112 this->result
.file
= PROGRAM_UNDEFINED
;
1114 /* Storage for our result. Ideally for an assignment we'd be using
1115 * the actual storage for the result here, instead.
1117 result_src
= get_temp(ir
->type
);
1118 /* convenience for the emit functions below. */
1119 result_dst
= dst_reg(result_src
);
1120 /* Limit writes to the channels that will be used by result_src later.
1121 * This does limit this temp's use as a temporary for multi-instruction
1124 result_dst
.writemask
= (1 << ir
->type
->vector_elements
) - 1;
1126 switch (ir
->operation
) {
1127 case ir_unop_logic_not
:
1128 /* Previously 'SEQ dst, src, 0.0' was used for this. However, many
1129 * older GPUs implement SEQ using multiple instructions (i915 uses two
1130 * SGE instructions and a MUL instruction). Since our logic values are
1131 * 0.0 and 1.0, 1-x also implements !x.
1133 op
[0].negate
= ~op
[0].negate
;
1134 emit(ir
, OPCODE_ADD
, result_dst
, op
[0], src_reg_for_float(1.0));
1137 op
[0].negate
= ~op
[0].negate
;
1141 emit(ir
, OPCODE_ABS
, result_dst
, op
[0]);
1144 emit(ir
, OPCODE_SSG
, result_dst
, op
[0]);
1147 emit_scalar(ir
, OPCODE_RCP
, result_dst
, op
[0]);
1151 emit_scalar(ir
, OPCODE_EX2
, result_dst
, op
[0]);
1155 assert(!"not reached: should be handled by ir_explog_to_explog2");
1158 emit_scalar(ir
, OPCODE_LG2
, result_dst
, op
[0]);
1161 emit_scalar(ir
, OPCODE_SIN
, result_dst
, op
[0]);
1164 emit_scalar(ir
, OPCODE_COS
, result_dst
, op
[0]);
1166 case ir_unop_sin_reduced
:
1167 emit_scs(ir
, OPCODE_SIN
, result_dst
, op
[0]);
1169 case ir_unop_cos_reduced
:
1170 emit_scs(ir
, OPCODE_COS
, result_dst
, op
[0]);
1174 emit(ir
, OPCODE_DDX
, result_dst
, op
[0]);
1177 emit(ir
, OPCODE_DDY
, result_dst
, op
[0]);
1180 case ir_unop_noise
: {
1181 const enum prog_opcode opcode
=
1182 prog_opcode(OPCODE_NOISE1
1183 + (ir
->operands
[0]->type
->vector_elements
) - 1);
1184 assert((opcode
>= OPCODE_NOISE1
) && (opcode
<= OPCODE_NOISE4
));
1186 emit(ir
, opcode
, result_dst
, op
[0]);
1191 emit(ir
, OPCODE_ADD
, result_dst
, op
[0], op
[1]);
1194 emit(ir
, OPCODE_SUB
, result_dst
, op
[0], op
[1]);
1198 emit(ir
, OPCODE_MUL
, result_dst
, op
[0], op
[1]);
1201 assert(!"not reached: should be handled by ir_div_to_mul_rcp");
1204 /* Floating point should be lowered by MOD_TO_FRACT in the compiler. */
1205 assert(ir
->type
->is_integer());
1206 emit(ir
, OPCODE_MUL
, result_dst
, op
[0], op
[1]);
1210 emit(ir
, OPCODE_SLT
, result_dst
, op
[0], op
[1]);
1212 case ir_binop_greater
:
1213 emit(ir
, OPCODE_SGT
, result_dst
, op
[0], op
[1]);
1215 case ir_binop_lequal
:
1216 emit(ir
, OPCODE_SLE
, result_dst
, op
[0], op
[1]);
1218 case ir_binop_gequal
:
1219 emit(ir
, OPCODE_SGE
, result_dst
, op
[0], op
[1]);
1221 case ir_binop_equal
:
1222 emit(ir
, OPCODE_SEQ
, result_dst
, op
[0], op
[1]);
1224 case ir_binop_nequal
:
1225 emit(ir
, OPCODE_SNE
, result_dst
, op
[0], op
[1]);
1227 case ir_binop_all_equal
:
1228 /* "==" operator producing a scalar boolean. */
1229 if (ir
->operands
[0]->type
->is_vector() ||
1230 ir
->operands
[1]->type
->is_vector()) {
1231 src_reg temp
= get_temp(glsl_type::vec4_type
);
1232 emit(ir
, OPCODE_SNE
, dst_reg(temp
), op
[0], op
[1]);
1234 /* After the dot-product, the value will be an integer on the
1235 * range [0,4]. Zero becomes 1.0, and positive values become zero.
1237 emit_dp(ir
, result_dst
, temp
, temp
, vector_elements
);
1239 /* Negating the result of the dot-product gives values on the range
1240 * [-4, 0]. Zero becomes 1.0, and negative values become zero. This
1241 * achieved using SGE.
1243 src_reg sge_src
= result_src
;
1244 sge_src
.negate
= ~sge_src
.negate
;
1245 emit(ir
, OPCODE_SGE
, result_dst
, sge_src
, src_reg_for_float(0.0));
1247 emit(ir
, OPCODE_SEQ
, result_dst
, op
[0], op
[1]);
1250 case ir_binop_any_nequal
:
1251 /* "!=" operator producing a scalar boolean. */
1252 if (ir
->operands
[0]->type
->is_vector() ||
1253 ir
->operands
[1]->type
->is_vector()) {
1254 src_reg temp
= get_temp(glsl_type::vec4_type
);
1255 emit(ir
, OPCODE_SNE
, dst_reg(temp
), op
[0], op
[1]);
1257 /* After the dot-product, the value will be an integer on the
1258 * range [0,4]. Zero stays zero, and positive values become 1.0.
1260 ir_to_mesa_instruction
*const dp
=
1261 emit_dp(ir
, result_dst
, temp
, temp
, vector_elements
);
1262 if (this->prog
->Target
== GL_FRAGMENT_PROGRAM_ARB
) {
1263 /* The clamping to [0,1] can be done for free in the fragment
1264 * shader with a saturate.
1266 dp
->saturate
= true;
1268 /* Negating the result of the dot-product gives values on the range
1269 * [-4, 0]. Zero stays zero, and negative values become 1.0. This
1270 * achieved using SLT.
1272 src_reg slt_src
= result_src
;
1273 slt_src
.negate
= ~slt_src
.negate
;
1274 emit(ir
, OPCODE_SLT
, result_dst
, slt_src
, src_reg_for_float(0.0));
1277 emit(ir
, OPCODE_SNE
, result_dst
, op
[0], op
[1]);
1282 assert(ir
->operands
[0]->type
->is_vector());
1284 /* After the dot-product, the value will be an integer on the
1285 * range [0,4]. Zero stays zero, and positive values become 1.0.
1287 ir_to_mesa_instruction
*const dp
=
1288 emit_dp(ir
, result_dst
, op
[0], op
[0],
1289 ir
->operands
[0]->type
->vector_elements
);
1290 if (this->prog
->Target
== GL_FRAGMENT_PROGRAM_ARB
) {
1291 /* The clamping to [0,1] can be done for free in the fragment
1292 * shader with a saturate.
1294 dp
->saturate
= true;
1296 /* Negating the result of the dot-product gives values on the range
1297 * [-4, 0]. Zero stays zero, and negative values become 1.0. This
1298 * is achieved using SLT.
1300 src_reg slt_src
= result_src
;
1301 slt_src
.negate
= ~slt_src
.negate
;
1302 emit(ir
, OPCODE_SLT
, result_dst
, slt_src
, src_reg_for_float(0.0));
1307 case ir_binop_logic_xor
:
1308 emit(ir
, OPCODE_SNE
, result_dst
, op
[0], op
[1]);
1311 case ir_binop_logic_or
: {
1312 /* After the addition, the value will be an integer on the
1313 * range [0,2]. Zero stays zero, and positive values become 1.0.
1315 ir_to_mesa_instruction
*add
=
1316 emit(ir
, OPCODE_ADD
, result_dst
, op
[0], op
[1]);
1317 if (this->prog
->Target
== GL_FRAGMENT_PROGRAM_ARB
) {
1318 /* The clamping to [0,1] can be done for free in the fragment
1319 * shader with a saturate.
1321 add
->saturate
= true;
1323 /* Negating the result of the addition gives values on the range
1324 * [-2, 0]. Zero stays zero, and negative values become 1.0. This
1325 * is achieved using SLT.
1327 src_reg slt_src
= result_src
;
1328 slt_src
.negate
= ~slt_src
.negate
;
1329 emit(ir
, OPCODE_SLT
, result_dst
, slt_src
, src_reg_for_float(0.0));
1334 case ir_binop_logic_and
:
1335 /* the bool args are stored as float 0.0 or 1.0, so "mul" gives us "and". */
1336 emit(ir
, OPCODE_MUL
, result_dst
, op
[0], op
[1]);
1340 assert(ir
->operands
[0]->type
->is_vector());
1341 assert(ir
->operands
[0]->type
== ir
->operands
[1]->type
);
1342 emit_dp(ir
, result_dst
, op
[0], op
[1],
1343 ir
->operands
[0]->type
->vector_elements
);
1347 /* sqrt(x) = x * rsq(x). */
1348 emit_scalar(ir
, OPCODE_RSQ
, result_dst
, op
[0]);
1349 emit(ir
, OPCODE_MUL
, result_dst
, result_src
, op
[0]);
1350 /* For incoming channels <= 0, set the result to 0. */
1351 op
[0].negate
= ~op
[0].negate
;
1352 emit(ir
, OPCODE_CMP
, result_dst
,
1353 op
[0], result_src
, src_reg_for_float(0.0));
1356 emit_scalar(ir
, OPCODE_RSQ
, result_dst
, op
[0]);
1364 /* Mesa IR lacks types, ints are stored as truncated floats. */
1369 emit(ir
, OPCODE_TRUNC
, result_dst
, op
[0]);
1373 emit(ir
, OPCODE_SNE
, result_dst
,
1374 op
[0], src_reg_for_float(0.0));
1376 case ir_unop_bitcast_f2i
: // Ignore these 4, they can't happen here anyway
1377 case ir_unop_bitcast_f2u
:
1378 case ir_unop_bitcast_i2f
:
1379 case ir_unop_bitcast_u2f
:
1382 emit(ir
, OPCODE_TRUNC
, result_dst
, op
[0]);
1385 op
[0].negate
= ~op
[0].negate
;
1386 emit(ir
, OPCODE_FLR
, result_dst
, op
[0]);
1387 result_src
.negate
= ~result_src
.negate
;
1390 emit(ir
, OPCODE_FLR
, result_dst
, op
[0]);
1393 emit(ir
, OPCODE_FRC
, result_dst
, op
[0]);
1395 case ir_unop_pack_snorm_2x16
:
1396 case ir_unop_pack_snorm_4x8
:
1397 case ir_unop_pack_unorm_2x16
:
1398 case ir_unop_pack_unorm_4x8
:
1399 case ir_unop_pack_half_2x16
:
1400 case ir_unop_unpack_snorm_2x16
:
1401 case ir_unop_unpack_snorm_4x8
:
1402 case ir_unop_unpack_unorm_2x16
:
1403 case ir_unop_unpack_unorm_4x8
:
1404 case ir_unop_unpack_half_2x16
:
1405 case ir_unop_unpack_half_2x16_split_x
:
1406 case ir_unop_unpack_half_2x16_split_y
:
1407 case ir_binop_pack_half_2x16_split
:
1408 case ir_unop_bitfield_reverse
:
1409 case ir_unop_bit_count
:
1410 case ir_unop_find_msb
:
1411 case ir_unop_find_lsb
:
1412 assert(!"not supported");
1415 emit(ir
, OPCODE_MIN
, result_dst
, op
[0], op
[1]);
1418 emit(ir
, OPCODE_MAX
, result_dst
, op
[0], op
[1]);
1421 emit_scalar(ir
, OPCODE_POW
, result_dst
, op
[0], op
[1]);
1424 /* GLSL 1.30 integer ops are unsupported in Mesa IR, but since
1425 * hardware backends have no way to avoid Mesa IR generation
1426 * even if they don't use it, we need to emit "something" and
1429 case ir_binop_lshift
:
1430 case ir_binop_rshift
:
1431 case ir_binop_bit_and
:
1432 case ir_binop_bit_xor
:
1433 case ir_binop_bit_or
:
1434 emit(ir
, OPCODE_ADD
, result_dst
, op
[0], op
[1]);
1437 case ir_unop_bit_not
:
1438 case ir_unop_round_even
:
1439 emit(ir
, OPCODE_MOV
, result_dst
, op
[0]);
1442 case ir_binop_ubo_load
:
1443 assert(!"not supported");
1447 /* ir_triop_lrp operands are (x, y, a) while
1448 * OPCODE_LRP operands are (a, y, x) to match ARB_fragment_program.
1450 emit(ir
, OPCODE_LRP
, result_dst
, op
[2], op
[1], op
[0]);
1453 case ir_binop_vector_extract
:
1457 case ir_triop_bitfield_extract
:
1458 case ir_triop_vector_insert
:
1459 case ir_quadop_bitfield_insert
:
1460 case ir_binop_ldexp
:
1462 case ir_binop_carry
:
1463 case ir_binop_borrow
:
1464 case ir_binop_imul_high
:
1465 assert(!"not supported");
1468 case ir_quadop_vector
:
1469 /* This operation should have already been handled.
1471 assert(!"Should not get here.");
1475 this->result
= result_src
;
1480 ir_to_mesa_visitor::visit(ir_swizzle
*ir
)
1486 /* Note that this is only swizzles in expressions, not those on the left
1487 * hand side of an assignment, which do write masking. See ir_assignment
1491 ir
->val
->accept(this);
1493 assert(src
.file
!= PROGRAM_UNDEFINED
);
1495 for (i
= 0; i
< 4; i
++) {
1496 if (i
< ir
->type
->vector_elements
) {
1499 swizzle
[i
] = GET_SWZ(src
.swizzle
, ir
->mask
.x
);
1502 swizzle
[i
] = GET_SWZ(src
.swizzle
, ir
->mask
.y
);
1505 swizzle
[i
] = GET_SWZ(src
.swizzle
, ir
->mask
.z
);
1508 swizzle
[i
] = GET_SWZ(src
.swizzle
, ir
->mask
.w
);
1512 /* If the type is smaller than a vec4, replicate the last
1515 swizzle
[i
] = swizzle
[ir
->type
->vector_elements
- 1];
1519 src
.swizzle
= MAKE_SWIZZLE4(swizzle
[0], swizzle
[1], swizzle
[2], swizzle
[3]);
1525 ir_to_mesa_visitor::visit(ir_dereference_variable
*ir
)
1527 variable_storage
*entry
= find_variable_storage(ir
->var
);
1528 ir_variable
*var
= ir
->var
;
1531 switch (var
->data
.mode
) {
1532 case ir_var_uniform
:
1533 entry
= new(mem_ctx
) variable_storage(var
, PROGRAM_UNIFORM
,
1534 var
->data
.location
);
1535 this->variables
.push_tail(entry
);
1537 case ir_var_shader_in
:
1538 /* The linker assigns locations for varyings and attributes,
1539 * including deprecated builtins (like gl_Color),
1540 * user-assigned generic attributes (glBindVertexLocation),
1541 * and user-defined varyings.
1543 assert(var
->data
.location
!= -1);
1544 entry
= new(mem_ctx
) variable_storage(var
,
1546 var
->data
.location
);
1548 case ir_var_shader_out
:
1549 assert(var
->data
.location
!= -1);
1550 entry
= new(mem_ctx
) variable_storage(var
,
1552 var
->data
.location
);
1554 case ir_var_system_value
:
1555 entry
= new(mem_ctx
) variable_storage(var
,
1556 PROGRAM_SYSTEM_VALUE
,
1557 var
->data
.location
);
1560 case ir_var_temporary
:
1561 entry
= new(mem_ctx
) variable_storage(var
, PROGRAM_TEMPORARY
,
1563 this->variables
.push_tail(entry
);
1565 next_temp
+= type_size(var
->type
);
1570 printf("Failed to make storage for %s\n", var
->name
);
1575 this->result
= src_reg(entry
->file
, entry
->index
, var
->type
);
1579 ir_to_mesa_visitor::visit(ir_dereference_array
*ir
)
1583 int element_size
= type_size(ir
->type
);
1585 index
= ir
->array_index
->constant_expression_value();
1587 ir
->array
->accept(this);
1591 src
.index
+= index
->value
.i
[0] * element_size
;
1593 /* Variable index array dereference. It eats the "vec4" of the
1594 * base of the array and an index that offsets the Mesa register
1597 ir
->array_index
->accept(this);
1601 if (element_size
== 1) {
1602 index_reg
= this->result
;
1604 index_reg
= get_temp(glsl_type::float_type
);
1606 emit(ir
, OPCODE_MUL
, dst_reg(index_reg
),
1607 this->result
, src_reg_for_float(element_size
));
1610 /* If there was already a relative address register involved, add the
1611 * new and the old together to get the new offset.
1613 if (src
.reladdr
!= NULL
) {
1614 src_reg accum_reg
= get_temp(glsl_type::float_type
);
1616 emit(ir
, OPCODE_ADD
, dst_reg(accum_reg
),
1617 index_reg
, *src
.reladdr
);
1619 index_reg
= accum_reg
;
1622 src
.reladdr
= ralloc(mem_ctx
, src_reg
);
1623 memcpy(src
.reladdr
, &index_reg
, sizeof(index_reg
));
1626 /* If the type is smaller than a vec4, replicate the last channel out. */
1627 if (ir
->type
->is_scalar() || ir
->type
->is_vector())
1628 src
.swizzle
= swizzle_for_size(ir
->type
->vector_elements
);
1630 src
.swizzle
= SWIZZLE_NOOP
;
1636 ir_to_mesa_visitor::visit(ir_dereference_record
*ir
)
1639 const glsl_type
*struct_type
= ir
->record
->type
;
1642 ir
->record
->accept(this);
1644 for (i
= 0; i
< struct_type
->length
; i
++) {
1645 if (strcmp(struct_type
->fields
.structure
[i
].name
, ir
->field
) == 0)
1647 offset
+= type_size(struct_type
->fields
.structure
[i
].type
);
1650 /* If the type is smaller than a vec4, replicate the last channel out. */
1651 if (ir
->type
->is_scalar() || ir
->type
->is_vector())
1652 this->result
.swizzle
= swizzle_for_size(ir
->type
->vector_elements
);
1654 this->result
.swizzle
= SWIZZLE_NOOP
;
1656 this->result
.index
+= offset
;
1660 * We want to be careful in assignment setup to hit the actual storage
1661 * instead of potentially using a temporary like we might with the
1662 * ir_dereference handler.
1665 get_assignment_lhs(ir_dereference
*ir
, ir_to_mesa_visitor
*v
)
1667 /* The LHS must be a dereference. If the LHS is a variable indexed array
1668 * access of a vector, it must be separated into a series conditional moves
1669 * before reaching this point (see ir_vec_index_to_cond_assign).
1671 assert(ir
->as_dereference());
1672 ir_dereference_array
*deref_array
= ir
->as_dereference_array();
1674 assert(!deref_array
->array
->type
->is_vector());
1677 /* Use the rvalue deref handler for the most part. We'll ignore
1678 * swizzles in it and write swizzles using writemask, though.
1681 return dst_reg(v
->result
);
1685 * Process the condition of a conditional assignment
1687 * Examines the condition of a conditional assignment to generate the optimal
1688 * first operand of a \c CMP instruction. If the condition is a relational
1689 * operator with 0 (e.g., \c ir_binop_less), the value being compared will be
1690 * used as the source for the \c CMP instruction. Otherwise the comparison
1691 * is processed to a boolean result, and the boolean result is used as the
1692 * operand to the CMP instruction.
1695 ir_to_mesa_visitor::process_move_condition(ir_rvalue
*ir
)
1697 ir_rvalue
*src_ir
= ir
;
1699 bool switch_order
= false;
1701 ir_expression
*const expr
= ir
->as_expression();
1702 if ((expr
!= NULL
) && (expr
->get_num_operands() == 2)) {
1703 bool zero_on_left
= false;
1705 if (expr
->operands
[0]->is_zero()) {
1706 src_ir
= expr
->operands
[1];
1707 zero_on_left
= true;
1708 } else if (expr
->operands
[1]->is_zero()) {
1709 src_ir
= expr
->operands
[0];
1710 zero_on_left
= false;
1714 * (a < 0) T F F ( a < 0) T F F
1715 * (0 < a) F F T (-a < 0) F F T
1716 * (a <= 0) T T F (-a < 0) F F T (swap order of other operands)
1717 * (0 <= a) F T T ( a < 0) T F F (swap order of other operands)
1718 * (a > 0) F F T (-a < 0) F F T
1719 * (0 > a) T F F ( a < 0) T F F
1720 * (a >= 0) F T T ( a < 0) T F F (swap order of other operands)
1721 * (0 >= a) T T F (-a < 0) F F T (swap order of other operands)
1723 * Note that exchanging the order of 0 and 'a' in the comparison simply
1724 * means that the value of 'a' should be negated.
1727 switch (expr
->operation
) {
1729 switch_order
= false;
1730 negate
= zero_on_left
;
1733 case ir_binop_greater
:
1734 switch_order
= false;
1735 negate
= !zero_on_left
;
1738 case ir_binop_lequal
:
1739 switch_order
= true;
1740 negate
= !zero_on_left
;
1743 case ir_binop_gequal
:
1744 switch_order
= true;
1745 negate
= zero_on_left
;
1749 /* This isn't the right kind of comparison afterall, so make sure
1750 * the whole condition is visited.
1758 src_ir
->accept(this);
1760 /* We use the OPCODE_CMP (a < 0 ? b : c) for conditional moves, and the
1761 * condition we produced is 0.0 or 1.0. By flipping the sign, we can
1762 * choose which value OPCODE_CMP produces without an extra instruction
1763 * computing the condition.
1766 this->result
.negate
= ~this->result
.negate
;
1768 return switch_order
;
1772 ir_to_mesa_visitor::visit(ir_assignment
*ir
)
1778 ir
->rhs
->accept(this);
1781 l
= get_assignment_lhs(ir
->lhs
, this);
1783 /* FINISHME: This should really set to the correct maximal writemask for each
1784 * FINISHME: component written (in the loops below). This case can only
1785 * FINISHME: occur for matrices, arrays, and structures.
1787 if (ir
->write_mask
== 0) {
1788 assert(!ir
->lhs
->type
->is_scalar() && !ir
->lhs
->type
->is_vector());
1789 l
.writemask
= WRITEMASK_XYZW
;
1790 } else if (ir
->lhs
->type
->is_scalar()) {
1791 /* FINISHME: This hack makes writing to gl_FragDepth, which lives in the
1792 * FINISHME: W component of fragment shader output zero, work correctly.
1794 l
.writemask
= WRITEMASK_XYZW
;
1797 int first_enabled_chan
= 0;
1800 assert(ir
->lhs
->type
->is_vector());
1801 l
.writemask
= ir
->write_mask
;
1803 for (int i
= 0; i
< 4; i
++) {
1804 if (l
.writemask
& (1 << i
)) {
1805 first_enabled_chan
= GET_SWZ(r
.swizzle
, i
);
1810 /* Swizzle a small RHS vector into the channels being written.
1812 * glsl ir treats write_mask as dictating how many channels are
1813 * present on the RHS while Mesa IR treats write_mask as just
1814 * showing which channels of the vec4 RHS get written.
1816 for (int i
= 0; i
< 4; i
++) {
1817 if (l
.writemask
& (1 << i
))
1818 swizzles
[i
] = GET_SWZ(r
.swizzle
, rhs_chan
++);
1820 swizzles
[i
] = first_enabled_chan
;
1822 r
.swizzle
= MAKE_SWIZZLE4(swizzles
[0], swizzles
[1],
1823 swizzles
[2], swizzles
[3]);
1826 assert(l
.file
!= PROGRAM_UNDEFINED
);
1827 assert(r
.file
!= PROGRAM_UNDEFINED
);
1829 if (ir
->condition
) {
1830 const bool switch_order
= this->process_move_condition(ir
->condition
);
1831 src_reg condition
= this->result
;
1833 for (i
= 0; i
< type_size(ir
->lhs
->type
); i
++) {
1835 emit(ir
, OPCODE_CMP
, l
, condition
, src_reg(l
), r
);
1837 emit(ir
, OPCODE_CMP
, l
, condition
, r
, src_reg(l
));
1844 for (i
= 0; i
< type_size(ir
->lhs
->type
); i
++) {
1845 emit(ir
, OPCODE_MOV
, l
, r
);
1854 ir_to_mesa_visitor::visit(ir_constant
*ir
)
1857 GLfloat stack_vals
[4] = { 0 };
1858 GLfloat
*values
= stack_vals
;
1861 /* Unfortunately, 4 floats is all we can get into
1862 * _mesa_add_unnamed_constant. So, make a temp to store an
1863 * aggregate constant and move each constant value into it. If we
1864 * get lucky, copy propagation will eliminate the extra moves.
1867 if (ir
->type
->base_type
== GLSL_TYPE_STRUCT
) {
1868 src_reg temp_base
= get_temp(ir
->type
);
1869 dst_reg temp
= dst_reg(temp_base
);
1871 foreach_list(node
, &ir
->components
) {
1872 ir_constant
*field_value
= (ir_constant
*) node
;
1873 int size
= type_size(field_value
->type
);
1877 field_value
->accept(this);
1880 for (i
= 0; i
< (unsigned int)size
; i
++) {
1881 emit(ir
, OPCODE_MOV
, temp
, src
);
1887 this->result
= temp_base
;
1891 if (ir
->type
->is_array()) {
1892 src_reg temp_base
= get_temp(ir
->type
);
1893 dst_reg temp
= dst_reg(temp_base
);
1894 int size
= type_size(ir
->type
->fields
.array
);
1898 for (i
= 0; i
< ir
->type
->length
; i
++) {
1899 ir
->array_elements
[i
]->accept(this);
1901 for (int j
= 0; j
< size
; j
++) {
1902 emit(ir
, OPCODE_MOV
, temp
, src
);
1908 this->result
= temp_base
;
1912 if (ir
->type
->is_matrix()) {
1913 src_reg mat
= get_temp(ir
->type
);
1914 dst_reg mat_column
= dst_reg(mat
);
1916 for (i
= 0; i
< ir
->type
->matrix_columns
; i
++) {
1917 assert(ir
->type
->base_type
== GLSL_TYPE_FLOAT
);
1918 values
= &ir
->value
.f
[i
* ir
->type
->vector_elements
];
1920 src
= src_reg(PROGRAM_CONSTANT
, -1, NULL
);
1921 src
.index
= _mesa_add_unnamed_constant(this->prog
->Parameters
,
1922 (gl_constant_value
*) values
,
1923 ir
->type
->vector_elements
,
1925 emit(ir
, OPCODE_MOV
, mat_column
, src
);
1934 src
.file
= PROGRAM_CONSTANT
;
1935 switch (ir
->type
->base_type
) {
1936 case GLSL_TYPE_FLOAT
:
1937 values
= &ir
->value
.f
[0];
1939 case GLSL_TYPE_UINT
:
1940 for (i
= 0; i
< ir
->type
->vector_elements
; i
++) {
1941 values
[i
] = ir
->value
.u
[i
];
1945 for (i
= 0; i
< ir
->type
->vector_elements
; i
++) {
1946 values
[i
] = ir
->value
.i
[i
];
1949 case GLSL_TYPE_BOOL
:
1950 for (i
= 0; i
< ir
->type
->vector_elements
; i
++) {
1951 values
[i
] = ir
->value
.b
[i
];
1955 assert(!"Non-float/uint/int/bool constant");
1958 this->result
= src_reg(PROGRAM_CONSTANT
, -1, ir
->type
);
1959 this->result
.index
= _mesa_add_unnamed_constant(this->prog
->Parameters
,
1960 (gl_constant_value
*) values
,
1961 ir
->type
->vector_elements
,
1962 &this->result
.swizzle
);
1966 ir_to_mesa_visitor::visit(ir_call
*ir
)
1968 assert(!"ir_to_mesa: All function calls should have been inlined by now.");
1972 ir_to_mesa_visitor::visit(ir_texture
*ir
)
1974 src_reg result_src
, coord
, lod_info
, projector
, dx
, dy
;
1975 dst_reg result_dst
, coord_dst
;
1976 ir_to_mesa_instruction
*inst
= NULL
;
1977 prog_opcode opcode
= OPCODE_NOP
;
1979 if (ir
->op
== ir_txs
)
1980 this->result
= src_reg_for_float(0.0);
1982 ir
->coordinate
->accept(this);
1984 /* Put our coords in a temp. We'll need to modify them for shadow,
1985 * projection, or LOD, so the only case we'd use it as is is if
1986 * we're doing plain old texturing. Mesa IR optimization should
1987 * handle cleaning up our mess in that case.
1989 coord
= get_temp(glsl_type::vec4_type
);
1990 coord_dst
= dst_reg(coord
);
1991 emit(ir
, OPCODE_MOV
, coord_dst
, this->result
);
1993 if (ir
->projector
) {
1994 ir
->projector
->accept(this);
1995 projector
= this->result
;
1998 /* Storage for our result. Ideally for an assignment we'd be using
1999 * the actual storage for the result here, instead.
2001 result_src
= get_temp(glsl_type::vec4_type
);
2002 result_dst
= dst_reg(result_src
);
2007 opcode
= OPCODE_TEX
;
2010 opcode
= OPCODE_TXB
;
2011 ir
->lod_info
.bias
->accept(this);
2012 lod_info
= this->result
;
2015 /* Pretend to be TXL so the sampler, coordinate, lod are available */
2017 opcode
= OPCODE_TXL
;
2018 ir
->lod_info
.lod
->accept(this);
2019 lod_info
= this->result
;
2022 opcode
= OPCODE_TXD
;
2023 ir
->lod_info
.grad
.dPdx
->accept(this);
2025 ir
->lod_info
.grad
.dPdy
->accept(this);
2029 assert(!"Unexpected ir_txf_ms opcode");
2032 assert(!"Unexpected ir_lod opcode");
2035 assert(!"Unexpected ir_tg4 opcode");
2037 case ir_query_levels
:
2038 assert(!"Unexpected ir_query_levels opcode");
2042 const glsl_type
*sampler_type
= ir
->sampler
->type
;
2044 if (ir
->projector
) {
2045 if (opcode
== OPCODE_TEX
) {
2046 /* Slot the projector in as the last component of the coord. */
2047 coord_dst
.writemask
= WRITEMASK_W
;
2048 emit(ir
, OPCODE_MOV
, coord_dst
, projector
);
2049 coord_dst
.writemask
= WRITEMASK_XYZW
;
2050 opcode
= OPCODE_TXP
;
2052 src_reg coord_w
= coord
;
2053 coord_w
.swizzle
= SWIZZLE_WWWW
;
2055 /* For the other TEX opcodes there's no projective version
2056 * since the last slot is taken up by lod info. Do the
2057 * projective divide now.
2059 coord_dst
.writemask
= WRITEMASK_W
;
2060 emit(ir
, OPCODE_RCP
, coord_dst
, projector
);
2062 /* In the case where we have to project the coordinates "by hand,"
2063 * the shadow comparitor value must also be projected.
2065 src_reg tmp_src
= coord
;
2066 if (ir
->shadow_comparitor
) {
2067 /* Slot the shadow value in as the second to last component of the
2070 ir
->shadow_comparitor
->accept(this);
2072 tmp_src
= get_temp(glsl_type::vec4_type
);
2073 dst_reg tmp_dst
= dst_reg(tmp_src
);
2075 /* Projective division not allowed for array samplers. */
2076 assert(!sampler_type
->sampler_array
);
2078 tmp_dst
.writemask
= WRITEMASK_Z
;
2079 emit(ir
, OPCODE_MOV
, tmp_dst
, this->result
);
2081 tmp_dst
.writemask
= WRITEMASK_XY
;
2082 emit(ir
, OPCODE_MOV
, tmp_dst
, coord
);
2085 coord_dst
.writemask
= WRITEMASK_XYZ
;
2086 emit(ir
, OPCODE_MUL
, coord_dst
, tmp_src
, coord_w
);
2088 coord_dst
.writemask
= WRITEMASK_XYZW
;
2089 coord
.swizzle
= SWIZZLE_XYZW
;
2093 /* If projection is done and the opcode is not OPCODE_TXP, then the shadow
2094 * comparitor was put in the correct place (and projected) by the code,
2095 * above, that handles by-hand projection.
2097 if (ir
->shadow_comparitor
&& (!ir
->projector
|| opcode
== OPCODE_TXP
)) {
2098 /* Slot the shadow value in as the second to last component of the
2101 ir
->shadow_comparitor
->accept(this);
2103 /* XXX This will need to be updated for cubemap array samplers. */
2104 if (sampler_type
->sampler_dimensionality
== GLSL_SAMPLER_DIM_2D
&&
2105 sampler_type
->sampler_array
) {
2106 coord_dst
.writemask
= WRITEMASK_W
;
2108 coord_dst
.writemask
= WRITEMASK_Z
;
2111 emit(ir
, OPCODE_MOV
, coord_dst
, this->result
);
2112 coord_dst
.writemask
= WRITEMASK_XYZW
;
2115 if (opcode
== OPCODE_TXL
|| opcode
== OPCODE_TXB
) {
2116 /* Mesa IR stores lod or lod bias in the last channel of the coords. */
2117 coord_dst
.writemask
= WRITEMASK_W
;
2118 emit(ir
, OPCODE_MOV
, coord_dst
, lod_info
);
2119 coord_dst
.writemask
= WRITEMASK_XYZW
;
2122 if (opcode
== OPCODE_TXD
)
2123 inst
= emit(ir
, opcode
, result_dst
, coord
, dx
, dy
);
2125 inst
= emit(ir
, opcode
, result_dst
, coord
);
2127 if (ir
->shadow_comparitor
)
2128 inst
->tex_shadow
= GL_TRUE
;
2130 inst
->sampler
= _mesa_get_sampler_uniform_value(ir
->sampler
,
2131 this->shader_program
,
2134 switch (sampler_type
->sampler_dimensionality
) {
2135 case GLSL_SAMPLER_DIM_1D
:
2136 inst
->tex_target
= (sampler_type
->sampler_array
)
2137 ? TEXTURE_1D_ARRAY_INDEX
: TEXTURE_1D_INDEX
;
2139 case GLSL_SAMPLER_DIM_2D
:
2140 inst
->tex_target
= (sampler_type
->sampler_array
)
2141 ? TEXTURE_2D_ARRAY_INDEX
: TEXTURE_2D_INDEX
;
2143 case GLSL_SAMPLER_DIM_3D
:
2144 inst
->tex_target
= TEXTURE_3D_INDEX
;
2146 case GLSL_SAMPLER_DIM_CUBE
:
2147 inst
->tex_target
= TEXTURE_CUBE_INDEX
;
2149 case GLSL_SAMPLER_DIM_RECT
:
2150 inst
->tex_target
= TEXTURE_RECT_INDEX
;
2152 case GLSL_SAMPLER_DIM_BUF
:
2153 assert(!"FINISHME: Implement ARB_texture_buffer_object");
2155 case GLSL_SAMPLER_DIM_EXTERNAL
:
2156 inst
->tex_target
= TEXTURE_EXTERNAL_INDEX
;
2159 assert(!"Should not get here.");
2162 this->result
= result_src
;
2166 ir_to_mesa_visitor::visit(ir_return
*ir
)
2168 /* Non-void functions should have been inlined. We may still emit RETs
2169 * from main() unless the EmitNoMainReturn option is set.
2171 assert(!ir
->get_value());
2172 emit(ir
, OPCODE_RET
);
2176 ir_to_mesa_visitor::visit(ir_discard
*ir
)
2178 if (ir
->condition
) {
2179 ir
->condition
->accept(this);
2180 this->result
.negate
= ~this->result
.negate
;
2181 emit(ir
, OPCODE_KIL
, undef_dst
, this->result
);
2183 emit(ir
, OPCODE_KIL_NV
);
2188 ir_to_mesa_visitor::visit(ir_if
*ir
)
2190 ir_to_mesa_instruction
*cond_inst
, *if_inst
;
2191 ir_to_mesa_instruction
*prev_inst
;
2193 prev_inst
= (ir_to_mesa_instruction
*)this->instructions
.get_tail();
2195 ir
->condition
->accept(this);
2196 assert(this->result
.file
!= PROGRAM_UNDEFINED
);
2198 if (this->options
->EmitCondCodes
) {
2199 cond_inst
= (ir_to_mesa_instruction
*)this->instructions
.get_tail();
2201 /* See if we actually generated any instruction for generating
2202 * the condition. If not, then cook up a move to a temp so we
2203 * have something to set cond_update on.
2205 if (cond_inst
== prev_inst
) {
2206 src_reg temp
= get_temp(glsl_type::bool_type
);
2207 cond_inst
= emit(ir
->condition
, OPCODE_MOV
, dst_reg(temp
), result
);
2209 cond_inst
->cond_update
= GL_TRUE
;
2211 if_inst
= emit(ir
->condition
, OPCODE_IF
);
2212 if_inst
->dst
.cond_mask
= COND_NE
;
2214 if_inst
= emit(ir
->condition
, OPCODE_IF
, undef_dst
, this->result
);
2217 this->instructions
.push_tail(if_inst
);
2219 visit_exec_list(&ir
->then_instructions
, this);
2221 if (!ir
->else_instructions
.is_empty()) {
2222 emit(ir
->condition
, OPCODE_ELSE
);
2223 visit_exec_list(&ir
->else_instructions
, this);
2226 emit(ir
->condition
, OPCODE_ENDIF
);
2230 ir_to_mesa_visitor::visit(ir_emit_vertex
*ir
)
2232 assert(!"Geometry shaders not supported.");
2236 ir_to_mesa_visitor::visit(ir_end_primitive
*ir
)
2238 assert(!"Geometry shaders not supported.");
2241 ir_to_mesa_visitor::ir_to_mesa_visitor()
2243 result
.file
= PROGRAM_UNDEFINED
;
2245 next_signature_id
= 1;
2246 current_function
= NULL
;
2247 mem_ctx
= ralloc_context(NULL
);
2250 ir_to_mesa_visitor::~ir_to_mesa_visitor()
2252 ralloc_free(mem_ctx
);
2255 static struct prog_src_register
2256 mesa_src_reg_from_ir_src_reg(src_reg reg
)
2258 struct prog_src_register mesa_reg
;
2260 mesa_reg
.File
= reg
.file
;
2261 assert(reg
.index
< (1 << INST_INDEX_BITS
));
2262 mesa_reg
.Index
= reg
.index
;
2263 mesa_reg
.Swizzle
= reg
.swizzle
;
2264 mesa_reg
.RelAddr
= reg
.reladdr
!= NULL
;
2265 mesa_reg
.Negate
= reg
.negate
;
2267 mesa_reg
.HasIndex2
= GL_FALSE
;
2268 mesa_reg
.RelAddr2
= 0;
2269 mesa_reg
.Index2
= 0;
2275 set_branchtargets(ir_to_mesa_visitor
*v
,
2276 struct prog_instruction
*mesa_instructions
,
2277 int num_instructions
)
2279 int if_count
= 0, loop_count
= 0;
2280 int *if_stack
, *loop_stack
;
2281 int if_stack_pos
= 0, loop_stack_pos
= 0;
2284 for (i
= 0; i
< num_instructions
; i
++) {
2285 switch (mesa_instructions
[i
].Opcode
) {
2289 case OPCODE_BGNLOOP
:
2294 mesa_instructions
[i
].BranchTarget
= -1;
2301 if_stack
= rzalloc_array(v
->mem_ctx
, int, if_count
);
2302 loop_stack
= rzalloc_array(v
->mem_ctx
, int, loop_count
);
2304 for (i
= 0; i
< num_instructions
; i
++) {
2305 switch (mesa_instructions
[i
].Opcode
) {
2307 if_stack
[if_stack_pos
] = i
;
2311 mesa_instructions
[if_stack
[if_stack_pos
- 1]].BranchTarget
= i
;
2312 if_stack
[if_stack_pos
- 1] = i
;
2315 mesa_instructions
[if_stack
[if_stack_pos
- 1]].BranchTarget
= i
;
2318 case OPCODE_BGNLOOP
:
2319 loop_stack
[loop_stack_pos
] = i
;
2322 case OPCODE_ENDLOOP
:
2324 /* Rewrite any breaks/conts at this nesting level (haven't
2325 * already had a BranchTarget assigned) to point to the end
2328 for (j
= loop_stack
[loop_stack_pos
]; j
< i
; j
++) {
2329 if (mesa_instructions
[j
].Opcode
== OPCODE_BRK
||
2330 mesa_instructions
[j
].Opcode
== OPCODE_CONT
) {
2331 if (mesa_instructions
[j
].BranchTarget
== -1) {
2332 mesa_instructions
[j
].BranchTarget
= i
;
2336 /* The loop ends point at each other. */
2337 mesa_instructions
[i
].BranchTarget
= loop_stack
[loop_stack_pos
];
2338 mesa_instructions
[loop_stack
[loop_stack_pos
]].BranchTarget
= i
;
2341 foreach_list(n
, &v
->function_signatures
) {
2342 function_entry
*entry
= (function_entry
*) n
;
2344 if (entry
->sig_id
== mesa_instructions
[i
].BranchTarget
) {
2345 mesa_instructions
[i
].BranchTarget
= entry
->inst
;
2357 print_program(struct prog_instruction
*mesa_instructions
,
2358 ir_instruction
**mesa_instruction_annotation
,
2359 int num_instructions
)
2361 ir_instruction
*last_ir
= NULL
;
2365 for (i
= 0; i
< num_instructions
; i
++) {
2366 struct prog_instruction
*mesa_inst
= mesa_instructions
+ i
;
2367 ir_instruction
*ir
= mesa_instruction_annotation
[i
];
2369 fprintf(stdout
, "%3d: ", i
);
2371 if (last_ir
!= ir
&& ir
) {
2374 for (j
= 0; j
< indent
; j
++) {
2375 fprintf(stdout
, " ");
2381 fprintf(stdout
, " "); /* line number spacing. */
2384 indent
= _mesa_fprint_instruction_opt(stdout
, mesa_inst
, indent
,
2385 PROG_PRINT_DEBUG
, NULL
);
2391 class add_uniform_to_shader
: public program_resource_visitor
{
2393 add_uniform_to_shader(struct gl_shader_program
*shader_program
,
2394 struct gl_program_parameter_list
*params
,
2395 gl_shader_stage shader_type
)
2396 : shader_program(shader_program
), params(params
), idx(-1),
2397 shader_type(shader_type
)
2402 void process(ir_variable
*var
)
2405 this->program_resource_visitor::process(var
);
2407 var
->data
.location
= this->idx
;
2411 virtual void visit_field(const glsl_type
*type
, const char *name
,
2414 struct gl_shader_program
*shader_program
;
2415 struct gl_program_parameter_list
*params
;
2417 gl_shader_stage shader_type
;
2420 } /* anonymous namespace */
2423 add_uniform_to_shader::visit_field(const glsl_type
*type
, const char *name
,
2430 if (type
->is_vector() || type
->is_scalar()) {
2431 size
= type
->vector_elements
;
2433 size
= type_size(type
) * 4;
2436 gl_register_file file
;
2437 if (type
->is_sampler() ||
2438 (type
->is_array() && type
->fields
.array
->is_sampler())) {
2439 file
= PROGRAM_SAMPLER
;
2441 file
= PROGRAM_UNIFORM
;
2444 int index
= _mesa_lookup_parameter_index(params
, -1, name
);
2446 index
= _mesa_add_parameter(params
, file
, name
, size
, type
->gl_type
,
2449 /* Sampler uniform values are stored in prog->SamplerUnits,
2450 * and the entry in that array is selected by this index we
2451 * store in ParameterValues[].
2453 if (file
== PROGRAM_SAMPLER
) {
2456 this->shader_program
->UniformHash
->get(location
,
2457 params
->Parameters
[index
].Name
);
2463 struct gl_uniform_storage
*storage
=
2464 &this->shader_program
->UniformStorage
[location
];
2466 assert(storage
->sampler
[shader_type
].active
);
2468 for (unsigned int j
= 0; j
< size
/ 4; j
++)
2469 params
->ParameterValues
[index
+ j
][0].f
=
2470 storage
->sampler
[shader_type
].index
+ j
;
2474 /* The first part of the uniform that's processed determines the base
2475 * location of the whole uniform (for structures).
2482 * Generate the program parameters list for the user uniforms in a shader
2484 * \param shader_program Linked shader program. This is only used to
2485 * emit possible link errors to the info log.
2486 * \param sh Shader whose uniforms are to be processed.
2487 * \param params Parameter list to be filled in.
2490 _mesa_generate_parameters_list_for_uniforms(struct gl_shader_program
2492 struct gl_shader
*sh
,
2493 struct gl_program_parameter_list
2496 add_uniform_to_shader
add(shader_program
, params
, sh
->Stage
);
2498 foreach_list(node
, sh
->ir
) {
2499 ir_variable
*var
= ((ir_instruction
*) node
)->as_variable();
2501 if ((var
== NULL
) || (var
->data
.mode
!= ir_var_uniform
)
2502 || var
->is_in_uniform_block() || (strncmp(var
->name
, "gl_", 3) == 0))
2510 _mesa_associate_uniform_storage(struct gl_context
*ctx
,
2511 struct gl_shader_program
*shader_program
,
2512 struct gl_program_parameter_list
*params
)
2514 /* After adding each uniform to the parameter list, connect the storage for
2515 * the parameter with the tracking structure used by the API for the
2518 unsigned last_location
= unsigned(~0);
2519 for (unsigned i
= 0; i
< params
->NumParameters
; i
++) {
2520 if (params
->Parameters
[i
].Type
!= PROGRAM_UNIFORM
)
2525 shader_program
->UniformHash
->get(location
, params
->Parameters
[i
].Name
);
2531 if (location
!= last_location
) {
2532 struct gl_uniform_storage
*storage
=
2533 &shader_program
->UniformStorage
[location
];
2534 enum gl_uniform_driver_format format
= uniform_native
;
2536 unsigned columns
= 0;
2537 switch (storage
->type
->base_type
) {
2538 case GLSL_TYPE_UINT
:
2539 assert(ctx
->Const
.NativeIntegers
);
2540 format
= uniform_native
;
2545 (ctx
->Const
.NativeIntegers
) ? uniform_native
: uniform_int_float
;
2548 case GLSL_TYPE_FLOAT
:
2549 format
= uniform_native
;
2550 columns
= storage
->type
->matrix_columns
;
2552 case GLSL_TYPE_BOOL
:
2553 if (ctx
->Const
.NativeIntegers
) {
2554 format
= (ctx
->Const
.UniformBooleanTrue
== 1)
2555 ? uniform_bool_int_0_1
: uniform_bool_int_0_not0
;
2557 format
= uniform_bool_float
;
2561 case GLSL_TYPE_SAMPLER
:
2562 case GLSL_TYPE_IMAGE
:
2563 format
= uniform_native
;
2566 case GLSL_TYPE_ATOMIC_UINT
:
2567 case GLSL_TYPE_ARRAY
:
2568 case GLSL_TYPE_VOID
:
2569 case GLSL_TYPE_STRUCT
:
2570 case GLSL_TYPE_ERROR
:
2571 case GLSL_TYPE_INTERFACE
:
2572 assert(!"Should not get here.");
2576 _mesa_uniform_attach_driver_storage(storage
,
2577 4 * sizeof(float) * columns
,
2580 ¶ms
->ParameterValues
[i
]);
2582 /* After attaching the driver's storage to the uniform, propagate any
2583 * data from the linker's backing store. This will cause values from
2584 * initializers in the source code to be copied over.
2586 _mesa_propagate_uniforms_to_driver_storage(storage
,
2588 MAX2(1, storage
->array_elements
));
2590 last_location
= location
;
2596 * On a basic block basis, tracks available PROGRAM_TEMPORARY register
2597 * channels for copy propagation and updates following instructions to
2598 * use the original versions.
2600 * The ir_to_mesa_visitor lazily produces code assuming that this pass
2601 * will occur. As an example, a TXP production before this pass:
2603 * 0: MOV TEMP[1], INPUT[4].xyyy;
2604 * 1: MOV TEMP[1].w, INPUT[4].wwww;
2605 * 2: TXP TEMP[2], TEMP[1], texture[0], 2D;
2609 * 0: MOV TEMP[1], INPUT[4].xyyy;
2610 * 1: MOV TEMP[1].w, INPUT[4].wwww;
2611 * 2: TXP TEMP[2], INPUT[4].xyyw, texture[0], 2D;
2613 * which allows for dead code elimination on TEMP[1]'s writes.
2616 ir_to_mesa_visitor::copy_propagate(void)
2618 ir_to_mesa_instruction
**acp
= rzalloc_array(mem_ctx
,
2619 ir_to_mesa_instruction
*,
2620 this->next_temp
* 4);
2621 int *acp_level
= rzalloc_array(mem_ctx
, int, this->next_temp
* 4);
2624 foreach_list(node
, &this->instructions
) {
2625 ir_to_mesa_instruction
*inst
= (ir_to_mesa_instruction
*) node
;
2627 assert(inst
->dst
.file
!= PROGRAM_TEMPORARY
2628 || inst
->dst
.index
< this->next_temp
);
2630 /* First, do any copy propagation possible into the src regs. */
2631 for (int r
= 0; r
< 3; r
++) {
2632 ir_to_mesa_instruction
*first
= NULL
;
2634 int acp_base
= inst
->src
[r
].index
* 4;
2636 if (inst
->src
[r
].file
!= PROGRAM_TEMPORARY
||
2637 inst
->src
[r
].reladdr
)
2640 /* See if we can find entries in the ACP consisting of MOVs
2641 * from the same src register for all the swizzled channels
2642 * of this src register reference.
2644 for (int i
= 0; i
< 4; i
++) {
2645 int src_chan
= GET_SWZ(inst
->src
[r
].swizzle
, i
);
2646 ir_to_mesa_instruction
*copy_chan
= acp
[acp_base
+ src_chan
];
2653 assert(acp_level
[acp_base
+ src_chan
] <= level
);
2658 if (first
->src
[0].file
!= copy_chan
->src
[0].file
||
2659 first
->src
[0].index
!= copy_chan
->src
[0].index
) {
2667 /* We've now validated that we can copy-propagate to
2668 * replace this src register reference. Do it.
2670 inst
->src
[r
].file
= first
->src
[0].file
;
2671 inst
->src
[r
].index
= first
->src
[0].index
;
2674 for (int i
= 0; i
< 4; i
++) {
2675 int src_chan
= GET_SWZ(inst
->src
[r
].swizzle
, i
);
2676 ir_to_mesa_instruction
*copy_inst
= acp
[acp_base
+ src_chan
];
2677 swizzle
|= (GET_SWZ(copy_inst
->src
[0].swizzle
, src_chan
) <<
2680 inst
->src
[r
].swizzle
= swizzle
;
2685 case OPCODE_BGNLOOP
:
2686 case OPCODE_ENDLOOP
:
2687 /* End of a basic block, clear the ACP entirely. */
2688 memset(acp
, 0, sizeof(*acp
) * this->next_temp
* 4);
2697 /* Clear all channels written inside the block from the ACP, but
2698 * leaving those that were not touched.
2700 for (int r
= 0; r
< this->next_temp
; r
++) {
2701 for (int c
= 0; c
< 4; c
++) {
2702 if (!acp
[4 * r
+ c
])
2705 if (acp_level
[4 * r
+ c
] >= level
)
2706 acp
[4 * r
+ c
] = NULL
;
2709 if (inst
->op
== OPCODE_ENDIF
)
2714 /* Continuing the block, clear any written channels from
2717 if (inst
->dst
.file
== PROGRAM_TEMPORARY
&& inst
->dst
.reladdr
) {
2718 /* Any temporary might be written, so no copy propagation
2719 * across this instruction.
2721 memset(acp
, 0, sizeof(*acp
) * this->next_temp
* 4);
2722 } else if (inst
->dst
.file
== PROGRAM_OUTPUT
&&
2723 inst
->dst
.reladdr
) {
2724 /* Any output might be written, so no copy propagation
2725 * from outputs across this instruction.
2727 for (int r
= 0; r
< this->next_temp
; r
++) {
2728 for (int c
= 0; c
< 4; c
++) {
2729 if (!acp
[4 * r
+ c
])
2732 if (acp
[4 * r
+ c
]->src
[0].file
== PROGRAM_OUTPUT
)
2733 acp
[4 * r
+ c
] = NULL
;
2736 } else if (inst
->dst
.file
== PROGRAM_TEMPORARY
||
2737 inst
->dst
.file
== PROGRAM_OUTPUT
) {
2738 /* Clear where it's used as dst. */
2739 if (inst
->dst
.file
== PROGRAM_TEMPORARY
) {
2740 for (int c
= 0; c
< 4; c
++) {
2741 if (inst
->dst
.writemask
& (1 << c
)) {
2742 acp
[4 * inst
->dst
.index
+ c
] = NULL
;
2747 /* Clear where it's used as src. */
2748 for (int r
= 0; r
< this->next_temp
; r
++) {
2749 for (int c
= 0; c
< 4; c
++) {
2750 if (!acp
[4 * r
+ c
])
2753 int src_chan
= GET_SWZ(acp
[4 * r
+ c
]->src
[0].swizzle
, c
);
2755 if (acp
[4 * r
+ c
]->src
[0].file
== inst
->dst
.file
&&
2756 acp
[4 * r
+ c
]->src
[0].index
== inst
->dst
.index
&&
2757 inst
->dst
.writemask
& (1 << src_chan
))
2759 acp
[4 * r
+ c
] = NULL
;
2767 /* If this is a copy, add it to the ACP. */
2768 if (inst
->op
== OPCODE_MOV
&&
2769 inst
->dst
.file
== PROGRAM_TEMPORARY
&&
2770 !(inst
->dst
.file
== inst
->src
[0].file
&&
2771 inst
->dst
.index
== inst
->src
[0].index
) &&
2772 !inst
->dst
.reladdr
&&
2774 !inst
->src
[0].reladdr
&&
2775 !inst
->src
[0].negate
) {
2776 for (int i
= 0; i
< 4; i
++) {
2777 if (inst
->dst
.writemask
& (1 << i
)) {
2778 acp
[4 * inst
->dst
.index
+ i
] = inst
;
2779 acp_level
[4 * inst
->dst
.index
+ i
] = level
;
2785 ralloc_free(acp_level
);
2791 * Convert a shader's GLSL IR into a Mesa gl_program.
2793 static struct gl_program
*
2794 get_mesa_program(struct gl_context
*ctx
,
2795 struct gl_shader_program
*shader_program
,
2796 struct gl_shader
*shader
)
2798 ir_to_mesa_visitor v
;
2799 struct prog_instruction
*mesa_instructions
, *mesa_inst
;
2800 ir_instruction
**mesa_instruction_annotation
;
2802 struct gl_program
*prog
;
2803 GLenum target
= _mesa_shader_stage_to_program(shader
->Stage
);
2804 const char *target_string
= _mesa_shader_stage_to_string(shader
->Stage
);
2805 struct gl_shader_compiler_options
*options
=
2806 &ctx
->ShaderCompilerOptions
[shader
->Stage
];
2808 validate_ir_tree(shader
->ir
);
2810 prog
= ctx
->Driver
.NewProgram(ctx
, target
, shader_program
->Name
);
2813 prog
->Parameters
= _mesa_new_parameter_list();
2816 v
.shader_program
= shader_program
;
2817 v
.options
= options
;
2819 _mesa_generate_parameters_list_for_uniforms(shader_program
, shader
,
2822 /* Emit Mesa IR for main(). */
2823 visit_exec_list(shader
->ir
, &v
);
2824 v
.emit(NULL
, OPCODE_END
);
2826 prog
->NumTemporaries
= v
.next_temp
;
2828 int num_instructions
= 0;
2829 foreach_list(node
, &v
.instructions
) {
2834 (struct prog_instruction
*)calloc(num_instructions
,
2835 sizeof(*mesa_instructions
));
2836 mesa_instruction_annotation
= ralloc_array(v
.mem_ctx
, ir_instruction
*,
2841 /* Convert ir_mesa_instructions into prog_instructions.
2843 mesa_inst
= mesa_instructions
;
2845 foreach_list(node
, &v
.instructions
) {
2846 const ir_to_mesa_instruction
*inst
= (ir_to_mesa_instruction
*) node
;
2848 mesa_inst
->Opcode
= inst
->op
;
2849 mesa_inst
->CondUpdate
= inst
->cond_update
;
2851 mesa_inst
->SaturateMode
= SATURATE_ZERO_ONE
;
2852 mesa_inst
->DstReg
.File
= inst
->dst
.file
;
2853 mesa_inst
->DstReg
.Index
= inst
->dst
.index
;
2854 mesa_inst
->DstReg
.CondMask
= inst
->dst
.cond_mask
;
2855 mesa_inst
->DstReg
.WriteMask
= inst
->dst
.writemask
;
2856 mesa_inst
->DstReg
.RelAddr
= inst
->dst
.reladdr
!= NULL
;
2857 mesa_inst
->SrcReg
[0] = mesa_src_reg_from_ir_src_reg(inst
->src
[0]);
2858 mesa_inst
->SrcReg
[1] = mesa_src_reg_from_ir_src_reg(inst
->src
[1]);
2859 mesa_inst
->SrcReg
[2] = mesa_src_reg_from_ir_src_reg(inst
->src
[2]);
2860 mesa_inst
->TexSrcUnit
= inst
->sampler
;
2861 mesa_inst
->TexSrcTarget
= inst
->tex_target
;
2862 mesa_inst
->TexShadow
= inst
->tex_shadow
;
2863 mesa_instruction_annotation
[i
] = inst
->ir
;
2865 /* Set IndirectRegisterFiles. */
2866 if (mesa_inst
->DstReg
.RelAddr
)
2867 prog
->IndirectRegisterFiles
|= 1 << mesa_inst
->DstReg
.File
;
2869 /* Update program's bitmask of indirectly accessed register files */
2870 for (unsigned src
= 0; src
< 3; src
++)
2871 if (mesa_inst
->SrcReg
[src
].RelAddr
)
2872 prog
->IndirectRegisterFiles
|= 1 << mesa_inst
->SrcReg
[src
].File
;
2874 switch (mesa_inst
->Opcode
) {
2876 if (options
->MaxIfDepth
== 0) {
2877 linker_warning(shader_program
,
2878 "Couldn't flatten if-statement. "
2879 "This will likely result in software "
2880 "rasterization.\n");
2883 case OPCODE_BGNLOOP
:
2884 if (options
->EmitNoLoops
) {
2885 linker_warning(shader_program
,
2886 "Couldn't unroll loop. "
2887 "This will likely result in software "
2888 "rasterization.\n");
2892 if (options
->EmitNoCont
) {
2893 linker_warning(shader_program
,
2894 "Couldn't lower continue-statement. "
2895 "This will likely result in software "
2896 "rasterization.\n");
2900 prog
->NumAddressRegs
= 1;
2909 if (!shader_program
->LinkStatus
)
2913 if (!shader_program
->LinkStatus
) {
2917 set_branchtargets(&v
, mesa_instructions
, num_instructions
);
2919 if (ctx
->_Shader
->Flags
& GLSL_DUMP
) {
2920 fprintf(stderr
, "\n");
2921 fprintf(stderr
, "GLSL IR for linked %s program %d:\n", target_string
,
2922 shader_program
->Name
);
2923 _mesa_print_ir(stderr
, shader
->ir
, NULL
);
2924 fprintf(stderr
, "\n");
2925 fprintf(stderr
, "\n");
2926 fprintf(stderr
, "Mesa IR for linked %s program %d:\n", target_string
,
2927 shader_program
->Name
);
2928 print_program(mesa_instructions
, mesa_instruction_annotation
,
2933 prog
->Instructions
= mesa_instructions
;
2934 prog
->NumInstructions
= num_instructions
;
2936 /* Setting this to NULL prevents a possible double free in the fail_exit
2939 mesa_instructions
= NULL
;
2941 do_set_program_inouts(shader
->ir
, prog
, shader
->Stage
);
2943 prog
->SamplersUsed
= shader
->active_samplers
;
2944 prog
->ShadowSamplers
= shader
->shadow_samplers
;
2945 _mesa_update_shader_textures_used(shader_program
, prog
);
2947 /* Set the gl_FragDepth layout. */
2948 if (target
== GL_FRAGMENT_PROGRAM_ARB
) {
2949 struct gl_fragment_program
*fp
= (struct gl_fragment_program
*)prog
;
2950 fp
->FragDepthLayout
= shader_program
->FragDepthLayout
;
2953 _mesa_reference_program(ctx
, &shader
->Program
, prog
);
2955 if ((ctx
->_Shader
->Flags
& GLSL_NO_OPT
) == 0) {
2956 _mesa_optimize_program(ctx
, prog
);
2959 /* This has to be done last. Any operation that can cause
2960 * prog->ParameterValues to get reallocated (e.g., anything that adds a
2961 * program constant) has to happen before creating this linkage.
2963 _mesa_associate_uniform_storage(ctx
, shader_program
, prog
->Parameters
);
2964 if (!shader_program
->LinkStatus
) {
2971 free(mesa_instructions
);
2972 _mesa_reference_program(ctx
, &shader
->Program
, NULL
);
2980 * Called via ctx->Driver.LinkShader()
2981 * This actually involves converting GLSL IR into Mesa gl_programs with
2982 * code lowering and other optimizations.
2985 _mesa_ir_link_shader(struct gl_context
*ctx
, struct gl_shader_program
*prog
)
2987 assert(prog
->LinkStatus
);
2989 for (unsigned i
= 0; i
< MESA_SHADER_STAGES
; i
++) {
2990 if (prog
->_LinkedShaders
[i
] == NULL
)
2994 exec_list
*ir
= prog
->_LinkedShaders
[i
]->ir
;
2995 const struct gl_shader_compiler_options
*options
=
2996 &ctx
->ShaderCompilerOptions
[prog
->_LinkedShaders
[i
]->Stage
];
3002 do_mat_op_to_vec(ir
);
3003 lower_instructions(ir
, (MOD_TO_FRACT
| DIV_TO_MUL_RCP
| EXP_TO_EXP2
3004 | LOG_TO_LOG2
| INT_DIV_TO_MUL_RCP
3005 | ((options
->EmitNoPow
) ? POW_TO_EXP2
: 0)));
3007 progress
= do_lower_jumps(ir
, true, true, options
->EmitNoMainReturn
, options
->EmitNoCont
, options
->EmitNoLoops
) || progress
;
3009 progress
= do_common_optimization(ir
, true, true,
3010 options
->MaxUnrollIterations
,
3014 progress
= lower_quadop_vector(ir
, true) || progress
;
3016 if (options
->MaxIfDepth
== 0)
3017 progress
= lower_discard(ir
) || progress
;
3019 progress
= lower_if_to_cond_assign(ir
, options
->MaxIfDepth
) || progress
;
3021 if (options
->EmitNoNoise
)
3022 progress
= lower_noise(ir
) || progress
;
3024 /* If there are forms of indirect addressing that the driver
3025 * cannot handle, perform the lowering pass.
3027 if (options
->EmitNoIndirectInput
|| options
->EmitNoIndirectOutput
3028 || options
->EmitNoIndirectTemp
|| options
->EmitNoIndirectUniform
)
3030 lower_variable_index_to_cond_assign(ir
,
3031 options
->EmitNoIndirectInput
,
3032 options
->EmitNoIndirectOutput
,
3033 options
->EmitNoIndirectTemp
,
3034 options
->EmitNoIndirectUniform
)
3037 progress
= do_vec_index_to_cond_assign(ir
) || progress
;
3038 progress
= lower_vector_insert(ir
, true) || progress
;
3041 validate_ir_tree(ir
);
3044 for (unsigned i
= 0; i
< MESA_SHADER_STAGES
; i
++) {
3045 struct gl_program
*linked_prog
;
3047 if (prog
->_LinkedShaders
[i
] == NULL
)
3050 linked_prog
= get_mesa_program(ctx
, prog
, prog
->_LinkedShaders
[i
]);
3053 _mesa_copy_linked_program_data((gl_shader_stage
) i
, prog
, linked_prog
);
3055 _mesa_reference_program(ctx
, &prog
->_LinkedShaders
[i
]->Program
,
3057 if (!ctx
->Driver
.ProgramStringNotify(ctx
,
3058 _mesa_shader_stage_to_program(i
),
3064 _mesa_reference_program(ctx
, &linked_prog
, NULL
);
3067 return prog
->LinkStatus
;
3071 * Link a GLSL shader program. Called via glLinkProgram().
3074 _mesa_glsl_link_shader(struct gl_context
*ctx
, struct gl_shader_program
*prog
)
3078 _mesa_clear_shader_program_data(ctx
, prog
);
3080 prog
->LinkStatus
= GL_TRUE
;
3082 for (i
= 0; i
< prog
->NumShaders
; i
++) {
3083 if (!prog
->Shaders
[i
]->CompileStatus
) {
3084 linker_error(prog
, "linking with uncompiled shader");
3088 if (prog
->LinkStatus
) {
3089 link_shaders(ctx
, prog
);
3092 if (prog
->LinkStatus
) {
3093 if (!ctx
->Driver
.LinkShader(ctx
, prog
)) {
3094 prog
->LinkStatus
= GL_FALSE
;
3098 if (ctx
->_Shader
->Flags
& GLSL_DUMP
) {
3099 if (!prog
->LinkStatus
) {
3100 fprintf(stderr
, "GLSL shader program %d failed to link\n", prog
->Name
);
3103 if (prog
->InfoLog
&& prog
->InfoLog
[0] != 0) {
3104 fprintf(stderr
, "GLSL shader program %d info log:\n", prog
->Name
);
3105 fprintf(stderr
, "%s\n", prog
->InfoLog
);