2 * Mesa 3-D graphics library
4 * Copyright (C) 2005-2008 Brian Paul All Rights Reserved.
5 * Copyright (C) 2008 VMware, Inc. All Rights Reserved.
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 shall be included
15 * in all copies or substantial portions of the Software.
17 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS
18 * OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
19 * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL
20 * BRIAN PAUL BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN
21 * AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN
22 * CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE SOFTWARE.
27 * Emit program instructions (PI code) from IR trees.
34 *** To emit GPU instructions, we basically just do an in-order traversal
39 #include "main/imports.h"
40 #include "main/context.h"
41 #include "main/macros.h"
42 #include "shader/program.h"
43 #include "shader/prog_instruction.h"
44 #include "shader/prog_parameter.h"
45 #include "shader/prog_print.h"
46 #include "slang_builtin.h"
47 #include "slang_emit.h"
48 #include "slang_mem.h"
51 #define PEEPHOLE_OPTIMIZATIONS 1
59 struct gl_program
*prog
;
60 struct gl_program
**Subroutines
;
61 GLuint NumSubroutines
;
63 GLuint MaxInstructions
; /**< size of prog->Instructions[] buffer */
65 /* code-gen options */
66 GLboolean EmitHighLevelInstructions
;
67 GLboolean EmitCondCodes
;
68 GLboolean EmitComments
;
69 GLboolean EmitBeginEndSub
; /* XXX TEMPORARY */
74 static struct gl_program
*
75 new_subroutine(slang_emit_info
*emitInfo
, GLuint
*id
)
77 GET_CURRENT_CONTEXT(ctx
);
78 const GLuint n
= emitInfo
->NumSubroutines
;
80 emitInfo
->Subroutines
= (struct gl_program
**)
81 _mesa_realloc(emitInfo
->Subroutines
,
82 n
* sizeof(struct gl_program
),
83 (n
+ 1) * sizeof(struct gl_program
));
84 emitInfo
->Subroutines
[n
] = ctx
->Driver
.NewProgram(ctx
, emitInfo
->prog
->Target
, 0);
85 emitInfo
->Subroutines
[n
]->Parameters
= emitInfo
->prog
->Parameters
;
86 emitInfo
->NumSubroutines
++;
88 return emitInfo
->Subroutines
[n
];
93 * Convert a writemask to a swizzle. Used for testing cond codes because
94 * we only want to test the cond code component(s) that was set by the
95 * previous instruction.
98 writemask_to_swizzle(GLuint writemask
)
100 if (writemask
== WRITEMASK_X
)
102 if (writemask
== WRITEMASK_Y
)
104 if (writemask
== WRITEMASK_Z
)
106 if (writemask
== WRITEMASK_W
)
108 return SWIZZLE_XYZW
; /* shouldn't be hit */
113 * Convert a swizzle mask to a writemask.
114 * Note that the slang_ir_storage->Swizzle field can represent either a
115 * swizzle mask or a writemask, depending on how it's used. For example,
116 * when we parse "direction.yz" alone, we don't know whether .yz is a
117 * writemask or a swizzle. In this case, we encode ".yz" in store->Swizzle
118 * as a swizzle mask (.yz?? actually). Later, if direction.yz is used as
119 * an R-value, we use store->Swizzle as-is. Otherwise, if direction.yz is
120 * used as an L-value, we convert it to a writemask.
123 swizzle_to_writemask(GLuint swizzle
)
125 GLuint i
, writemask
= 0x0;
126 for (i
= 0; i
< 4; i
++) {
127 GLuint swz
= GET_SWZ(swizzle
, i
);
128 if (swz
<= SWIZZLE_W
) {
129 writemask
|= (1 << swz
);
137 * Swizzle a swizzle (function composition).
138 * That is, return swz2(swz1), or said another way: swz1.szw2
139 * Example: swizzle_swizzle(".zwxx", ".xxyw") yields ".zzwx"
142 _slang_swizzle_swizzle(GLuint swz1
, GLuint swz2
)
145 for (i
= 0; i
< 4; i
++) {
146 GLuint c
= GET_SWZ(swz2
, i
);
148 s
[i
] = GET_SWZ(swz1
, c
);
152 swz
= MAKE_SWIZZLE4(s
[0], s
[1], s
[2], s
[3]);
158 * Return the default swizzle mask for accessing a variable of the
159 * given size (in floats). If size = 1, comp is used to identify
160 * which component [0..3] of the register holds the variable.
163 _slang_var_swizzle(GLint size
, GLint comp
)
167 return MAKE_SWIZZLE4(comp
, comp
, comp
, comp
);
169 return MAKE_SWIZZLE4(SWIZZLE_X
, SWIZZLE_Y
, SWIZZLE_NIL
, SWIZZLE_NIL
);
171 return MAKE_SWIZZLE4(SWIZZLE_X
, SWIZZLE_Y
, SWIZZLE_Z
, SWIZZLE_NIL
);
180 * Allocate storage for the given node (if it hasn't already been allocated).
182 * Typically this is temporary storage for an intermediate result (such as
183 * for a multiply or add, etc).
185 * If n->Store does not exist it will be created and will be of the size
186 * specified by defaultSize.
189 alloc_node_storage(slang_emit_info
*emitInfo
, slang_ir_node
*n
,
194 assert(defaultSize
> 0);
195 n
->Store
= _slang_new_ir_storage(PROGRAM_TEMPORARY
, -1, defaultSize
);
198 /* now allocate actual register(s). I.e. set n->Store->Index >= 0 */
199 if (n
->Store
->Index
< 0) {
200 if (!_slang_alloc_temp(emitInfo
->vt
, n
->Store
)) {
201 slang_info_log_error(emitInfo
->log
,
202 "Ran out of registers, too many temporaries");
203 _slang_free(n
->Store
);
213 * Free temporary storage, if n->Store is, in fact, temp storage.
217 free_node_storage(slang_var_table
*vt
, slang_ir_node
*n
)
219 if (n
->Store
->File
== PROGRAM_TEMPORARY
&&
220 n
->Store
->Index
>= 0 &&
221 n
->Opcode
!= IR_SWIZZLE
) {
222 if (_slang_is_temp(vt
, n
->Store
)) {
223 _slang_free_temp(vt
, n
->Store
);
224 n
->Store
->Index
= -1;
225 n
->Store
= NULL
; /* XXX this may not be needed */
232 * Helper function to allocate a short-term temporary.
233 * Free it with _slang_free_temp().
236 alloc_local_temp(slang_emit_info
*emitInfo
, slang_ir_storage
*temp
, GLint size
)
240 _mesa_bzero(temp
, sizeof(*temp
));
242 temp
->File
= PROGRAM_TEMPORARY
;
244 return _slang_alloc_temp(emitInfo
->vt
, temp
);
249 * Remove any SWIZZLE_NIL terms from given swizzle mask.
250 * For a swizzle like .z??? generate .zzzz (replicate single component).
251 * Else, for .wx?? generate .wxzw (insert default component for the position).
254 fix_swizzle(GLuint swizzle
)
256 GLuint c0
= GET_SWZ(swizzle
, 0),
257 c1
= GET_SWZ(swizzle
, 1),
258 c2
= GET_SWZ(swizzle
, 2),
259 c3
= GET_SWZ(swizzle
, 3);
260 if (c1
== SWIZZLE_NIL
&& c2
== SWIZZLE_NIL
&& c3
== SWIZZLE_NIL
) {
261 /* smear first component across all positions */
265 /* insert default swizzle components */
266 if (c0
== SWIZZLE_NIL
)
268 if (c1
== SWIZZLE_NIL
)
270 if (c2
== SWIZZLE_NIL
)
272 if (c3
== SWIZZLE_NIL
)
275 return MAKE_SWIZZLE4(c0
, c1
, c2
, c3
);
281 * Convert IR storage to an instruction dst register.
284 storage_to_dst_reg(struct prog_dst_register
*dst
, const slang_ir_storage
*st
)
286 const GLboolean relAddr
= st
->RelAddr
;
287 const GLint size
= st
->Size
;
288 GLint index
= st
->Index
;
289 GLuint swizzle
= st
->Swizzle
;
292 /* if this is storage relative to some parent storage, walk up the tree */
295 assert(st
->Index
>= 0);
297 swizzle
= _slang_swizzle_swizzle(st
->Swizzle
, swizzle
);
300 assert(st
->File
!= PROGRAM_UNDEFINED
);
301 dst
->File
= st
->File
;
309 if (swizzle
!= SWIZZLE_XYZW
) {
310 dst
->WriteMask
= swizzle_to_writemask(swizzle
);
316 writemask
= WRITEMASK_X
<< GET_SWZ(st
->Swizzle
, 0);
319 writemask
= WRITEMASK_XY
;
322 writemask
= WRITEMASK_XYZ
;
325 writemask
= WRITEMASK_XYZW
;
328 ; /* error would have been caught above */
330 dst
->WriteMask
= writemask
;
333 dst
->RelAddr
= relAddr
;
338 * Convert IR storage to an instruction src register.
341 storage_to_src_reg(struct prog_src_register
*src
, const slang_ir_storage
*st
)
343 const GLboolean relAddr
= st
->RelAddr
;
344 GLint index
= st
->Index
;
345 GLuint swizzle
= st
->Swizzle
;
347 /* if this is storage relative to some parent storage, walk up the tree */
351 assert(st
->Index
>= 0);
353 swizzle
= _slang_swizzle_swizzle(fix_swizzle(st
->Swizzle
), swizzle
);
356 assert(st
->File
>= 0);
357 #if 1 /* XXX temporary */
358 if (st
->File
== PROGRAM_UNDEFINED
) {
359 slang_ir_storage
*st0
= (slang_ir_storage
*) st
;
360 st0
->File
= PROGRAM_TEMPORARY
;
363 assert(st
->File
< PROGRAM_UNDEFINED
);
364 src
->File
= st
->File
;
369 swizzle
= fix_swizzle(swizzle
);
370 assert(GET_SWZ(swizzle
, 0) <= SWIZZLE_W
);
371 assert(GET_SWZ(swizzle
, 1) <= SWIZZLE_W
);
372 assert(GET_SWZ(swizzle
, 2) <= SWIZZLE_W
);
373 assert(GET_SWZ(swizzle
, 3) <= SWIZZLE_W
);
374 src
->Swizzle
= swizzle
;
376 src
->RelAddr
= relAddr
;
381 * Setup storage pointing to a scalar constant/literal.
384 constant_to_storage(slang_emit_info
*emitInfo
,
386 slang_ir_storage
*store
)
393 reg
= _mesa_add_unnamed_constant(emitInfo
->prog
->Parameters
,
396 memset(store
, 0, sizeof(*store
));
397 store
->File
= PROGRAM_CONSTANT
;
399 store
->Swizzle
= swizzle
;
404 * Add new instruction at end of given program.
405 * \param prog the program to append instruction onto
406 * \param opcode opcode for the new instruction
407 * \return pointer to the new instruction
409 static struct prog_instruction
*
410 new_instruction(slang_emit_info
*emitInfo
, gl_inst_opcode opcode
)
412 struct gl_program
*prog
= emitInfo
->prog
;
413 struct prog_instruction
*inst
;
416 /* print prev inst */
417 if (prog
->NumInstructions
> 0) {
418 _mesa_print_instruction(prog
->Instructions
+ prog
->NumInstructions
- 1);
421 assert(prog
->NumInstructions
<= emitInfo
->MaxInstructions
);
423 if (prog
->NumInstructions
== emitInfo
->MaxInstructions
) {
424 /* grow the instruction buffer */
425 emitInfo
->MaxInstructions
+= 20;
427 _mesa_realloc_instructions(prog
->Instructions
,
428 prog
->NumInstructions
,
429 emitInfo
->MaxInstructions
);
432 inst
= prog
->Instructions
+ prog
->NumInstructions
;
433 prog
->NumInstructions
++;
434 _mesa_init_instructions(inst
, 1);
435 inst
->Opcode
= opcode
;
436 inst
->BranchTarget
= -1; /* invalid */
438 printf("New inst %d: %p %s\n", prog->NumInstructions-1,(void*)inst,
439 _mesa_opcode_string(inst->Opcode));
445 static struct prog_instruction
*
446 emit_arl_load(slang_emit_info
*emitInfo
,
447 enum register_file file
, GLint index
, GLuint swizzle
)
449 struct prog_instruction
*inst
= new_instruction(emitInfo
, OPCODE_ARL
);
450 inst
->SrcReg
[0].File
= file
;
451 inst
->SrcReg
[0].Index
= index
;
452 inst
->SrcReg
[0].Swizzle
= swizzle
;
453 inst
->DstReg
.File
= PROGRAM_ADDRESS
;
454 inst
->DstReg
.Index
= 0;
455 inst
->DstReg
.WriteMask
= WRITEMASK_X
;
461 * Emit a new instruction with given opcode, operands.
462 * At this point the instruction may have multiple indirect register
463 * loads/stores. We convert those into ARL loads and address-relative
464 * operands. See comments inside.
465 * At some point in the future we could directly emit indirectly addressed
466 * registers in Mesa GPU instructions.
468 static struct prog_instruction
*
469 emit_instruction(slang_emit_info
*emitInfo
,
470 gl_inst_opcode opcode
,
471 const slang_ir_storage
*dst
,
472 const slang_ir_storage
*src0
,
473 const slang_ir_storage
*src1
,
474 const slang_ir_storage
*src2
)
476 struct prog_instruction
*inst
;
477 GLuint numIndirect
= 0;
478 const slang_ir_storage
*src
[3];
479 slang_ir_storage newSrc
[3], newDst
;
483 isTemp
[0] = isTemp
[1] = isTemp
[2] = GL_FALSE
;
489 /* count up how many operands are indirect loads */
490 for (i
= 0; i
< 3; i
++) {
491 if (src
[i
] && src
[i
]->IsIndirect
)
494 if (dst
&& dst
->IsIndirect
)
497 /* Take special steps for indirect register loads.
498 * If we had multiple address registers this would be simpler.
499 * For example, this GLSL code:
500 * x[i] = y[j] + z[k];
501 * would translate into something like:
505 * ADD TEMP[ADDR.x+5], TEMP[ADDR.y+9], TEMP[ADDR.z+4];
506 * But since we currently only have one address register we have to do this:
508 * MOV t1, TEMP[ADDR.x+9];
510 * MOV t2, TEMP[ADDR.x+4];
512 * ADD TEMP[ADDR.x+5], t1, t2;
513 * The code here figures this out...
515 if (numIndirect
> 0) {
516 for (i
= 0; i
< 3; i
++) {
517 if (src
[i
] && src
[i
]->IsIndirect
) {
518 /* load the ARL register with the indirect register */
519 emit_arl_load(emitInfo
,
520 src
[i
]->IndirectFile
,
521 src
[i
]->IndirectIndex
,
522 src
[i
]->IndirectSwizzle
);
524 if (numIndirect
> 1) {
525 /* Need to load src[i] into a temporary register */
526 slang_ir_storage srcRelAddr
;
527 alloc_local_temp(emitInfo
, &newSrc
[i
], src
[i
]->Size
);
530 /* set RelAddr flag on src register */
531 srcRelAddr
= *src
[i
];
532 srcRelAddr
.RelAddr
= GL_TRUE
;
533 srcRelAddr
.IsIndirect
= GL_FALSE
; /* not really needed */
535 /* MOV newSrc, srcRelAddr; */
536 inst
= emit_instruction(emitInfo
,
546 /* just rewrite the src[i] storage to be ARL-relative */
548 newSrc
[i
].RelAddr
= GL_TRUE
;
549 newSrc
[i
].IsIndirect
= GL_FALSE
; /* not really needed */
556 /* Take special steps for indirect dest register write */
557 if (dst
&& dst
->IsIndirect
) {
558 /* load the ARL register with the indirect register */
559 emit_arl_load(emitInfo
,
562 dst
->IndirectSwizzle
);
564 newDst
.RelAddr
= GL_TRUE
;
565 newDst
.IsIndirect
= GL_FALSE
;
569 /* OK, emit the instruction and its dst, src regs */
570 inst
= new_instruction(emitInfo
, opcode
);
575 storage_to_dst_reg(&inst
->DstReg
, dst
);
577 for (i
= 0; i
< 3; i
++) {
579 storage_to_src_reg(&inst
->SrcReg
[i
], src
[i
]);
582 /* Free any temp registers that we allocated above */
583 for (i
= 0; i
< 3; i
++) {
585 _slang_free_temp(emitInfo
->vt
, &newSrc
[i
]);
594 * Put a comment on the given instruction.
597 inst_comment(struct prog_instruction
*inst
, const char *comment
)
600 inst
->Comment
= _mesa_strdup(comment
);
606 * Return pointer to last instruction in program.
608 static struct prog_instruction
*
609 prev_instruction(slang_emit_info
*emitInfo
)
611 struct gl_program
*prog
= emitInfo
->prog
;
612 if (prog
->NumInstructions
== 0)
615 return prog
->Instructions
+ prog
->NumInstructions
- 1;
619 static struct prog_instruction
*
620 emit(slang_emit_info
*emitInfo
, slang_ir_node
*n
);
624 * Return an annotation string for given node's storage.
627 storage_annotation(const slang_ir_node
*n
, const struct gl_program
*prog
)
630 const slang_ir_storage
*st
= n
->Store
;
631 static char s
[100] = "";
634 return _mesa_strdup("");
637 case PROGRAM_CONSTANT
:
638 if (st
->Index
>= 0) {
639 const GLfloat
*val
= prog
->Parameters
->ParameterValues
[st
->Index
];
640 if (st
->Swizzle
== SWIZZLE_NOOP
)
641 sprintf(s
, "{%g, %g, %g, %g}", val
[0], val
[1], val
[2], val
[3]);
643 sprintf(s
, "%g", val
[GET_SWZ(st
->Swizzle
, 0)]);
647 case PROGRAM_TEMPORARY
:
649 sprintf(s
, "%s", (char *) n
->Var
->a_name
);
651 sprintf(s
, "t[%d]", st
->Index
);
653 case PROGRAM_STATE_VAR
:
654 case PROGRAM_UNIFORM
:
655 sprintf(s
, "%s", prog
->Parameters
->Parameters
[st
->Index
].Name
);
657 case PROGRAM_VARYING
:
658 sprintf(s
, "%s", prog
->Varying
->Parameters
[st
->Index
].Name
);
661 sprintf(s
, "input[%d]", st
->Index
);
664 sprintf(s
, "output[%d]", st
->Index
);
669 return _mesa_strdup(s
);
677 * Return an annotation string for an instruction.
680 instruction_annotation(gl_inst_opcode opcode
, char *dstAnnot
,
681 char *srcAnnot0
, char *srcAnnot1
, char *srcAnnot2
)
684 const char *operator;
689 len
+= strlen(dstAnnot
);
691 dstAnnot
= _mesa_strdup("");
694 len
+= strlen(srcAnnot0
);
696 srcAnnot0
= _mesa_strdup("");
699 len
+= strlen(srcAnnot1
);
701 srcAnnot1
= _mesa_strdup("");
704 len
+= strlen(srcAnnot2
);
706 srcAnnot2
= _mesa_strdup("");
740 s
= (char *) malloc(len
);
741 sprintf(s
, "%s = %s %s %s %s", dstAnnot
,
742 srcAnnot0
, operator, srcAnnot1
, srcAnnot2
);
743 assert(_mesa_strlen(s
) < len
);
758 * Emit an instruction that's just a comment.
760 static struct prog_instruction
*
761 emit_comment(slang_emit_info
*emitInfo
, const char *comment
)
763 struct prog_instruction
*inst
= new_instruction(emitInfo
, OPCODE_NOP
);
764 inst_comment(inst
, comment
);
770 * Generate code for a simple arithmetic instruction.
771 * Either 1, 2 or 3 operands.
773 static struct prog_instruction
*
774 emit_arith(slang_emit_info
*emitInfo
, slang_ir_node
*n
)
776 const slang_ir_info
*info
= _slang_ir_info(n
->Opcode
);
777 struct prog_instruction
*inst
;
781 assert(info
->InstOpcode
!= OPCODE_NOP
);
783 #if PEEPHOLE_OPTIMIZATIONS
784 /* Look for MAD opportunity */
785 if (info
->NumParams
== 2 &&
786 n
->Opcode
== IR_ADD
&& n
->Children
[0]->Opcode
== IR_MUL
) {
787 /* found pattern IR_ADD(IR_MUL(A, B), C) */
788 emit(emitInfo
, n
->Children
[0]->Children
[0]); /* A */
789 emit(emitInfo
, n
->Children
[0]->Children
[1]); /* B */
790 emit(emitInfo
, n
->Children
[1]); /* C */
791 alloc_node_storage(emitInfo
, n
, -1); /* dest */
793 inst
= emit_instruction(emitInfo
,
796 n
->Children
[0]->Children
[0]->Store
,
797 n
->Children
[0]->Children
[1]->Store
,
798 n
->Children
[1]->Store
);
800 free_node_storage(emitInfo
->vt
, n
->Children
[0]->Children
[0]);
801 free_node_storage(emitInfo
->vt
, n
->Children
[0]->Children
[1]);
802 free_node_storage(emitInfo
->vt
, n
->Children
[1]);
806 if (info
->NumParams
== 2 &&
807 n
->Opcode
== IR_ADD
&& n
->Children
[1]->Opcode
== IR_MUL
) {
808 /* found pattern IR_ADD(A, IR_MUL(B, C)) */
809 emit(emitInfo
, n
->Children
[0]); /* A */
810 emit(emitInfo
, n
->Children
[1]->Children
[0]); /* B */
811 emit(emitInfo
, n
->Children
[1]->Children
[1]); /* C */
812 alloc_node_storage(emitInfo
, n
, -1); /* dest */
814 inst
= emit_instruction(emitInfo
,
817 n
->Children
[1]->Children
[0]->Store
,
818 n
->Children
[1]->Children
[1]->Store
,
819 n
->Children
[0]->Store
);
821 free_node_storage(emitInfo
->vt
, n
->Children
[1]->Children
[0]);
822 free_node_storage(emitInfo
->vt
, n
->Children
[1]->Children
[1]);
823 free_node_storage(emitInfo
->vt
, n
->Children
[0]);
828 /* gen code for children, may involve temp allocation */
829 for (i
= 0; i
< info
->NumParams
; i
++) {
830 emit(emitInfo
, n
->Children
[i
]);
831 if (!n
->Children
[i
] || !n
->Children
[i
]->Store
) {
838 alloc_node_storage(emitInfo
, n
, -1);
840 inst
= emit_instruction(emitInfo
,
843 (info
->NumParams
> 0 ? n
->Children
[0]->Store
: NULL
),
844 (info
->NumParams
> 1 ? n
->Children
[1]->Store
: NULL
),
845 (info
->NumParams
> 2 ? n
->Children
[2]->Store
: NULL
)
849 for (i
= 0; i
< info
->NumParams
; i
++)
850 free_node_storage(emitInfo
->vt
, n
->Children
[i
]);
857 * Emit code for == and != operators. These could normally be handled
858 * by emit_arith() except we need to be able to handle structure comparisons.
860 static struct prog_instruction
*
861 emit_compare(slang_emit_info
*emitInfo
, slang_ir_node
*n
)
863 struct prog_instruction
*inst
= NULL
;
866 assert(n
->Opcode
== IR_EQUAL
|| n
->Opcode
== IR_NOTEQUAL
);
868 /* gen code for children */
869 emit(emitInfo
, n
->Children
[0]);
870 emit(emitInfo
, n
->Children
[1]);
872 if (n
->Children
[0]->Store
->Size
!= n
->Children
[1]->Store
->Size
) {
873 slang_info_log_error(emitInfo
->log
, "invalid operands to == or !=");
877 /* final result is 1 bool */
878 if (!alloc_node_storage(emitInfo
, n
, 1))
881 size
= n
->Children
[0]->Store
->Size
;
884 gl_inst_opcode opcode
= n
->Opcode
== IR_EQUAL
? OPCODE_SEQ
: OPCODE_SNE
;
885 inst
= emit_instruction(emitInfo
,
888 n
->Children
[0]->Store
,
889 n
->Children
[1]->Store
,
892 else if (size
<= 4) {
893 /* compare two vectors.
894 * Unfortunately, there's no instruction to compare vectors and
895 * return a scalar result. Do it with some compare and dot product
899 gl_inst_opcode dotOp
;
900 slang_ir_storage tempStore
;
902 if (!alloc_local_temp(emitInfo
, &tempStore
, 4)) {
909 swizzle
= SWIZZLE_XYZW
;
911 else if (size
== 3) {
913 swizzle
= SWIZZLE_XYZW
;
917 dotOp
= OPCODE_DP3
; /* XXX use OPCODE_DP2 eventually */
918 swizzle
= MAKE_SWIZZLE4(SWIZZLE_X
, SWIZZLE_Y
, SWIZZLE_Y
, SWIZZLE_Y
);
921 /* Compute inequality (temp = (A != B)) */
922 inst
= emit_instruction(emitInfo
,
925 n
->Children
[0]->Store
,
926 n
->Children
[1]->Store
,
928 inst_comment(inst
, "Compare values");
930 /* Compute val = DOT(temp, temp) (reduction) */
931 inst
= emit_instruction(emitInfo
,
937 inst
->SrcReg
[0].Swizzle
= inst
->SrcReg
[1].Swizzle
= swizzle
; /*override*/
938 inst_comment(inst
, "Reduce vec to bool");
940 _slang_free_temp(emitInfo
->vt
, &tempStore
); /* free temp */
942 if (n
->Opcode
== IR_EQUAL
) {
943 /* compute val = !val.x with SEQ val, val, 0; */
944 slang_ir_storage zero
;
945 constant_to_storage(emitInfo
, 0.0, &zero
);
946 inst
= emit_instruction(emitInfo
,
952 inst_comment(inst
, "Invert true/false");
956 /* size > 4, struct or array compare.
957 * XXX this won't work reliably for structs with padding!!
959 GLint i
, num
= (n
->Children
[0]->Store
->Size
+ 3) / 4;
960 slang_ir_storage accTemp
, sneTemp
;
962 if (!alloc_local_temp(emitInfo
, &accTemp
, 4))
965 if (!alloc_local_temp(emitInfo
, &sneTemp
, 4))
968 for (i
= 0; i
< num
; i
++) {
969 slang_ir_storage srcStore0
= *n
->Children
[0]->Store
;
970 slang_ir_storage srcStore1
= *n
->Children
[1]->Store
;
971 srcStore0
.Index
+= i
;
972 srcStore1
.Index
+= i
;
975 /* SNE accTemp, left[i], right[i] */
976 inst
= emit_instruction(emitInfo
, OPCODE_SNE
,
981 inst_comment(inst
, "Begin struct/array comparison");
984 /* SNE sneTemp, left[i], right[i] */
985 inst
= emit_instruction(emitInfo
, OPCODE_SNE
,
990 /* ADD accTemp, accTemp, sneTemp; # like logical-OR */
991 inst
= emit_instruction(emitInfo
, OPCODE_ADD
,
999 /* compute accTemp.x || accTemp.y || accTemp.z || accTemp.w with DOT4 */
1000 inst
= emit_instruction(emitInfo
, OPCODE_DP4
,
1005 inst_comment(inst
, "End struct/array comparison");
1007 if (n
->Opcode
== IR_EQUAL
) {
1008 /* compute tmp.x = !tmp.x via tmp.x = (tmp.x == 0) */
1009 slang_ir_storage zero
;
1010 constant_to_storage(emitInfo
, 0.0, &zero
);
1011 inst
= emit_instruction(emitInfo
, OPCODE_SEQ
,
1012 n
->Store
, /* dest */
1016 inst_comment(inst
, "Invert true/false");
1019 _slang_free_temp(emitInfo
->vt
, &accTemp
);
1020 _slang_free_temp(emitInfo
->vt
, &sneTemp
);
1024 free_node_storage(emitInfo
->vt
, n
->Children
[0]);
1025 free_node_storage(emitInfo
->vt
, n
->Children
[1]);
1033 * Generate code for an IR_CLAMP instruction.
1035 static struct prog_instruction
*
1036 emit_clamp(slang_emit_info
*emitInfo
, slang_ir_node
*n
)
1038 struct prog_instruction
*inst
;
1039 slang_ir_node tmpNode
;
1041 assert(n
->Opcode
== IR_CLAMP
);
1047 inst
= emit(emitInfo
, n
->Children
[0]);
1049 /* If lower limit == 0.0 and upper limit == 1.0,
1050 * set prev instruction's SaturateMode field to SATURATE_ZERO_ONE.
1052 * emit OPCODE_MIN, OPCODE_MAX sequence.
1055 /* XXX this isn't quite finished yet */
1056 if (n
->Children
[1]->Opcode
== IR_FLOAT
&&
1057 n
->Children
[1]->Value
[0] == 0.0 &&
1058 n
->Children
[1]->Value
[1] == 0.0 &&
1059 n
->Children
[1]->Value
[2] == 0.0 &&
1060 n
->Children
[1]->Value
[3] == 0.0 &&
1061 n
->Children
[2]->Opcode
== IR_FLOAT
&&
1062 n
->Children
[2]->Value
[0] == 1.0 &&
1063 n
->Children
[2]->Value
[1] == 1.0 &&
1064 n
->Children
[2]->Value
[2] == 1.0 &&
1065 n
->Children
[2]->Value
[3] == 1.0) {
1067 inst
= prev_instruction(prog
);
1069 if (inst
&& inst
->Opcode
!= OPCODE_NOP
) {
1070 /* and prev instruction's DstReg matches n->Children[0]->Store */
1071 inst
->SaturateMode
= SATURATE_ZERO_ONE
;
1072 n
->Store
= n
->Children
[0]->Store
;
1078 if (!alloc_node_storage(emitInfo
, n
, n
->Children
[0]->Store
->Size
))
1081 emit(emitInfo
, n
->Children
[1]);
1082 emit(emitInfo
, n
->Children
[2]);
1084 /* Some GPUs don't allow reading from output registers. So if the
1085 * dest for this clamp() is an output reg, we can't use that reg for
1086 * the intermediate result. Use a temp register instead.
1088 _mesa_bzero(&tmpNode
, sizeof(tmpNode
));
1089 alloc_node_storage(emitInfo
, &tmpNode
, n
->Store
->Size
);
1091 /* tmp = max(ch[0], ch[1]) */
1092 inst
= emit_instruction(emitInfo
, OPCODE_MAX
,
1093 tmpNode
.Store
, /* dest */
1094 n
->Children
[0]->Store
,
1095 n
->Children
[1]->Store
,
1098 /* n->dest = min(tmp, ch[2]) */
1099 inst
= emit_instruction(emitInfo
, OPCODE_MIN
,
1100 n
->Store
, /* dest */
1102 n
->Children
[2]->Store
,
1105 free_node_storage(emitInfo
->vt
, &tmpNode
);
1111 static struct prog_instruction
*
1112 emit_negation(slang_emit_info
*emitInfo
, slang_ir_node
*n
)
1114 /* Implement as MOV dst, -src; */
1115 /* XXX we could look at the previous instruction and in some circumstances
1116 * modify it to accomplish the negation.
1118 struct prog_instruction
*inst
;
1120 emit(emitInfo
, n
->Children
[0]);
1122 if (!alloc_node_storage(emitInfo
, n
, n
->Children
[0]->Store
->Size
))
1125 inst
= emit_instruction(emitInfo
,
1127 n
->Store
, /* dest */
1128 n
->Children
[0]->Store
,
1131 inst
->SrcReg
[0].NegateBase
= NEGATE_XYZW
;
1136 static struct prog_instruction
*
1137 emit_label(slang_emit_info
*emitInfo
, const slang_ir_node
*n
)
1141 /* XXX this fails in loop tail code - investigate someday */
1142 assert(_slang_label_get_location(n
->Label
) < 0);
1143 _slang_label_set_location(n
->Label
, emitInfo
->prog
->NumInstructions
,
1146 if (_slang_label_get_location(n
->Label
) < 0)
1147 _slang_label_set_location(n
->Label
, emitInfo
->prog
->NumInstructions
,
1155 * Emit code for a function call.
1156 * Note that for each time a function is called, we emit the function's
1157 * body code again because the set of available registers may be different.
1159 static struct prog_instruction
*
1160 emit_fcall(slang_emit_info
*emitInfo
, slang_ir_node
*n
)
1162 struct gl_program
*progSave
;
1163 struct prog_instruction
*inst
;
1164 GLuint subroutineId
;
1167 assert(n
->Opcode
== IR_CALL
);
1170 /* save/push cur program */
1171 maxInstSave
= emitInfo
->MaxInstructions
;
1172 progSave
= emitInfo
->prog
;
1174 emitInfo
->prog
= new_subroutine(emitInfo
, &subroutineId
);
1175 emitInfo
->MaxInstructions
= emitInfo
->prog
->NumInstructions
;
1177 _slang_label_set_location(n
->Label
, emitInfo
->prog
->NumInstructions
,
1180 if (emitInfo
->EmitBeginEndSub
) {
1181 /* BGNSUB isn't a real instruction.
1182 * We require a label (i.e. "foobar:") though, if we're going to
1183 * print the program in the NV format. The BNGSUB instruction is
1184 * really just a NOP to attach the label to.
1186 inst
= new_instruction(emitInfo
, OPCODE_BGNSUB
);
1187 inst_comment(inst
, n
->Label
->Name
);
1190 /* body of function: */
1191 emit(emitInfo
, n
->Children
[0]);
1192 n
->Store
= n
->Children
[0]->Store
;
1194 /* add RET instruction now, if needed */
1195 inst
= prev_instruction(emitInfo
);
1196 if (inst
&& inst
->Opcode
!= OPCODE_RET
) {
1197 inst
= new_instruction(emitInfo
, OPCODE_RET
);
1200 if (emitInfo
->EmitBeginEndSub
) {
1201 inst
= new_instruction(emitInfo
, OPCODE_ENDSUB
);
1202 inst_comment(inst
, n
->Label
->Name
);
1205 /* pop/restore cur program */
1206 emitInfo
->prog
= progSave
;
1207 emitInfo
->MaxInstructions
= maxInstSave
;
1209 /* emit the function call */
1210 inst
= new_instruction(emitInfo
, OPCODE_CAL
);
1211 /* The branch target is just the subroutine number (changed later) */
1212 inst
->BranchTarget
= subroutineId
;
1213 inst_comment(inst
, n
->Label
->Name
);
1214 assert(inst
->BranchTarget
>= 0);
1221 * Emit code for a 'return' statement.
1223 static struct prog_instruction
*
1224 emit_return(slang_emit_info
*emitInfo
, slang_ir_node
*n
)
1226 struct prog_instruction
*inst
;
1228 assert(n
->Opcode
== IR_RETURN
);
1230 inst
= new_instruction(emitInfo
, OPCODE_RET
);
1231 inst
->DstReg
.CondMask
= COND_TR
; /* always return */
1236 static struct prog_instruction
*
1237 emit_kill(slang_emit_info
*emitInfo
)
1239 struct gl_fragment_program
*fp
;
1240 struct prog_instruction
*inst
;
1241 /* NV-KILL - discard fragment depending on condition code.
1242 * Note that ARB-KILL depends on sign of vector operand.
1244 inst
= new_instruction(emitInfo
, OPCODE_KIL_NV
);
1245 inst
->DstReg
.CondMask
= COND_TR
; /* always kill */
1247 assert(emitInfo
->prog
->Target
== GL_FRAGMENT_PROGRAM_ARB
);
1248 fp
= (struct gl_fragment_program
*) emitInfo
->prog
;
1249 fp
->UsesKill
= GL_TRUE
;
1255 static struct prog_instruction
*
1256 emit_tex(slang_emit_info
*emitInfo
, slang_ir_node
*n
)
1258 struct prog_instruction
*inst
;
1259 gl_inst_opcode opcode
;
1261 if (n
->Opcode
== IR_TEX
) {
1262 opcode
= OPCODE_TEX
;
1264 else if (n
->Opcode
== IR_TEXB
) {
1265 opcode
= OPCODE_TXB
;
1268 assert(n
->Opcode
== IR_TEXP
);
1269 opcode
= OPCODE_TXP
;
1272 /* emit code for the texcoord operand */
1273 (void) emit(emitInfo
, n
->Children
[1]);
1275 /* alloc storage for result of texture fetch */
1276 if (!alloc_node_storage(emitInfo
, n
, 4))
1279 /* emit TEX instruction; Child[1] is the texcoord */
1280 inst
= emit_instruction(emitInfo
,
1283 n
->Children
[1]->Store
,
1287 /* Child[0] is the sampler (a uniform which'll indicate the texture unit) */
1288 assert(n
->Children
[0]->Store
);
1289 /* Store->Index is the sampler index */
1290 assert(n
->Children
[0]->Store
->Index
>= 0);
1291 /* Store->Size is the texture target */
1292 assert(n
->Children
[0]->Store
->Size
>= TEXTURE_1D_INDEX
);
1293 assert(n
->Children
[0]->Store
->Size
<= TEXTURE_RECT_INDEX
);
1295 inst
->TexSrcTarget
= n
->Children
[0]->Store
->Size
;
1296 inst
->TexSrcUnit
= n
->Children
[0]->Store
->Index
; /* i.e. uniform's index */
1305 static struct prog_instruction
*
1306 emit_copy(slang_emit_info
*emitInfo
, slang_ir_node
*n
)
1308 struct prog_instruction
*inst
;
1310 assert(n
->Opcode
== IR_COPY
);
1313 emit(emitInfo
, n
->Children
[0]);
1314 if (!n
->Children
[0]->Store
|| n
->Children
[0]->Store
->Index
< 0) {
1315 /* an error should have been already recorded */
1320 assert(n
->Children
[1]);
1321 inst
= emit(emitInfo
, n
->Children
[1]);
1323 if (!n
->Children
[1]->Store
|| n
->Children
[1]->Store
->Index
< 0) {
1324 if (!emitInfo
->log
->text
) {
1325 slang_info_log_error(emitInfo
->log
, "invalid assignment");
1330 assert(n
->Children
[1]->Store
->Index
>= 0);
1332 /*assert(n->Children[0]->Store->Size == n->Children[1]->Store->Size);*/
1334 n
->Store
= n
->Children
[0]->Store
;
1336 if (n
->Store
->File
== PROGRAM_SAMPLER
) {
1337 /* no code generated for sampler assignments,
1338 * just copy the sampler index at compile time.
1340 n
->Store
->Index
= n
->Children
[1]->Store
->Index
;
1344 #if PEEPHOLE_OPTIMIZATIONS
1346 _slang_is_temp(emitInfo
->vt
, n
->Children
[1]->Store
) &&
1347 (inst
->DstReg
.File
== n
->Children
[1]->Store
->File
) &&
1348 (inst
->DstReg
.Index
== n
->Children
[1]->Store
->Index
) &&
1349 !n
->Children
[0]->Store
->IsIndirect
&&
1350 n
->Children
[0]->Store
->Size
<= 4) {
1351 /* Peephole optimization:
1352 * The Right-Hand-Side has its results in a temporary place.
1353 * Modify the RHS (and the prev instruction) to store its results
1354 * in the destination specified by n->Children[0].
1355 * Then, this MOVE is a no-op.
1362 if (n
->Children
[1]->Opcode
!= IR_SWIZZLE
)
1363 _slang_free_temp(emitInfo
->vt
, n
->Children
[1]->Store
);
1364 *n
->Children
[1]->Store
= *n
->Children
[0]->Store
;
1366 /* fixup the previous instruction (which stored the RHS result) */
1367 assert(n
->Children
[0]->Store
->Index
>= 0);
1369 storage_to_dst_reg(&inst
->DstReg
, n
->Children
[0]->Store
);
1375 if (n
->Children
[0]->Store
->Size
> 4) {
1376 /* move matrix/struct etc (block of registers) */
1377 slang_ir_storage dstStore
= *n
->Children
[0]->Store
;
1378 slang_ir_storage srcStore
= *n
->Children
[1]->Store
;
1379 GLint size
= srcStore
.Size
;
1380 ASSERT(n
->Children
[1]->Store
->Swizzle
== SWIZZLE_NOOP
);
1384 inst
= emit_instruction(emitInfo
, OPCODE_MOV
,
1389 inst_comment(inst
, "IR_COPY block");
1396 /* single register move */
1397 char *srcAnnot
, *dstAnnot
;
1398 assert(n
->Children
[0]->Store
->Index
>= 0);
1399 inst
= emit_instruction(emitInfo
, OPCODE_MOV
,
1400 n
->Children
[0]->Store
, /* dest */
1401 n
->Children
[1]->Store
,
1404 dstAnnot
= storage_annotation(n
->Children
[0], emitInfo
->prog
);
1405 srcAnnot
= storage_annotation(n
->Children
[1], emitInfo
->prog
);
1406 inst
->Comment
= instruction_annotation(inst
->Opcode
, dstAnnot
,
1407 srcAnnot
, NULL
, NULL
);
1409 free_node_storage(emitInfo
->vt
, n
->Children
[1]);
1416 * An IR_COND node wraps a boolean expression which is used by an
1417 * IF or WHILE test. This is where we'll set condition codes, if needed.
1419 static struct prog_instruction
*
1420 emit_cond(slang_emit_info
*emitInfo
, slang_ir_node
*n
)
1422 struct prog_instruction
*inst
;
1424 assert(n
->Opcode
== IR_COND
);
1426 if (!n
->Children
[0])
1429 /* emit code for the expression */
1430 inst
= emit(emitInfo
, n
->Children
[0]);
1432 if (!n
->Children
[0]->Store
) {
1433 /* error recovery */
1437 assert(n
->Children
[0]->Store
);
1438 /*assert(n->Children[0]->Store->Size == 1);*/
1440 if (emitInfo
->EmitCondCodes
) {
1442 n
->Children
[0]->Store
&&
1443 inst
->DstReg
.File
== n
->Children
[0]->Store
->File
&&
1444 inst
->DstReg
.Index
== n
->Children
[0]->Store
->Index
) {
1445 /* The previous instruction wrote to the register who's value
1446 * we're testing. Just fix that instruction so that the
1447 * condition codes are computed.
1449 inst
->CondUpdate
= GL_TRUE
;
1450 n
->Store
= n
->Children
[0]->Store
;
1454 /* This'll happen for things like "if (i) ..." where no code
1455 * is normally generated for the expression "i".
1456 * Generate a move instruction just to set condition codes.
1458 if (!alloc_node_storage(emitInfo
, n
, 1))
1460 inst
= emit_instruction(emitInfo
, OPCODE_MOV
,
1461 n
->Store
, /* dest */
1462 n
->Children
[0]->Store
,
1465 inst
->CondUpdate
= GL_TRUE
;
1466 inst_comment(inst
, "COND expr");
1467 _slang_free_temp(emitInfo
->vt
, n
->Store
);
1472 /* No-op: the boolean result of the expression is in a regular reg */
1473 n
->Store
= n
->Children
[0]->Store
;
1482 static struct prog_instruction
*
1483 emit_not(slang_emit_info
*emitInfo
, slang_ir_node
*n
)
1485 static const struct {
1486 gl_inst_opcode op
, opNot
;
1488 { OPCODE_SLT
, OPCODE_SGE
},
1489 { OPCODE_SLE
, OPCODE_SGT
},
1490 { OPCODE_SGT
, OPCODE_SLE
},
1491 { OPCODE_SGE
, OPCODE_SLT
},
1492 { OPCODE_SEQ
, OPCODE_SNE
},
1493 { OPCODE_SNE
, OPCODE_SEQ
},
1496 struct prog_instruction
*inst
;
1497 slang_ir_storage zero
;
1501 inst
= emit(emitInfo
, n
->Children
[0]);
1503 #if PEEPHOLE_OPTIMIZATIONS
1505 /* if the prev instruction was a comparison instruction, invert it */
1506 for (i
= 0; operators
[i
].op
; i
++) {
1507 if (inst
->Opcode
== operators
[i
].op
) {
1508 inst
->Opcode
= operators
[i
].opNot
;
1509 n
->Store
= n
->Children
[0]->Store
;
1516 /* else, invert using SEQ (v = v == 0) */
1517 if (!alloc_node_storage(emitInfo
, n
, n
->Children
[0]->Store
->Size
))
1520 constant_to_storage(emitInfo
, 0.0, &zero
);
1521 inst
= emit_instruction(emitInfo
,
1524 n
->Children
[0]->Store
,
1527 inst_comment(inst
, "NOT");
1529 free_node_storage(emitInfo
->vt
, n
->Children
[0]);
1535 static struct prog_instruction
*
1536 emit_if(slang_emit_info
*emitInfo
, slang_ir_node
*n
)
1538 struct gl_program
*prog
= emitInfo
->prog
;
1539 GLuint ifInstLoc
, elseInstLoc
= 0;
1540 GLuint condWritemask
= 0;
1542 /* emit condition expression code */
1544 struct prog_instruction
*inst
;
1545 inst
= emit(emitInfo
, n
->Children
[0]);
1546 if (emitInfo
->EmitCondCodes
) {
1548 /* error recovery */
1551 condWritemask
= inst
->DstReg
.WriteMask
;
1555 if (!n
->Children
[0]->Store
)
1559 assert(n
->Children
[0]->Store
->Size
== 1); /* a bool! */
1562 ifInstLoc
= prog
->NumInstructions
;
1563 if (emitInfo
->EmitHighLevelInstructions
) {
1564 if (emitInfo
->EmitCondCodes
) {
1565 /* IF condcode THEN ... */
1566 struct prog_instruction
*ifInst
;
1567 ifInst
= new_instruction(emitInfo
, OPCODE_IF
);
1568 ifInst
->DstReg
.CondMask
= COND_NE
; /* if cond is non-zero */
1569 /* only test the cond code (1 of 4) that was updated by the
1570 * previous instruction.
1572 ifInst
->DstReg
.CondSwizzle
= writemask_to_swizzle(condWritemask
);
1575 /* IF src[0] THEN ... */
1576 emit_instruction(emitInfo
, OPCODE_IF
,
1578 n
->Children
[0]->Store
, /* op0 */
1584 /* conditional jump to else, or endif */
1585 struct prog_instruction
*ifInst
= new_instruction(emitInfo
, OPCODE_BRA
);
1586 ifInst
->DstReg
.CondMask
= COND_EQ
; /* BRA if cond is zero */
1587 inst_comment(ifInst
, "if zero");
1588 ifInst
->DstReg
.CondSwizzle
= writemask_to_swizzle(condWritemask
);
1592 emit(emitInfo
, n
->Children
[1]);
1594 if (n
->Children
[2]) {
1595 /* have else body */
1596 elseInstLoc
= prog
->NumInstructions
;
1597 if (emitInfo
->EmitHighLevelInstructions
) {
1598 (void) new_instruction(emitInfo
, OPCODE_ELSE
);
1601 /* jump to endif instruction */
1602 struct prog_instruction
*inst
;
1603 inst
= new_instruction(emitInfo
, OPCODE_BRA
);
1604 inst_comment(inst
, "else");
1605 inst
->DstReg
.CondMask
= COND_TR
; /* always branch */
1607 prog
->Instructions
[ifInstLoc
].BranchTarget
= prog
->NumInstructions
;
1608 emit(emitInfo
, n
->Children
[2]);
1612 prog
->Instructions
[ifInstLoc
].BranchTarget
= prog
->NumInstructions
;
1615 if (emitInfo
->EmitHighLevelInstructions
) {
1616 (void) new_instruction(emitInfo
, OPCODE_ENDIF
);
1619 if (n
->Children
[2]) {
1620 prog
->Instructions
[elseInstLoc
].BranchTarget
= prog
->NumInstructions
;
1626 static struct prog_instruction
*
1627 emit_loop(slang_emit_info
*emitInfo
, slang_ir_node
*n
)
1629 struct gl_program
*prog
= emitInfo
->prog
;
1630 struct prog_instruction
*endInst
;
1631 GLuint beginInstLoc
, tailInstLoc
, endInstLoc
;
1634 /* emit OPCODE_BGNLOOP */
1635 beginInstLoc
= prog
->NumInstructions
;
1636 if (emitInfo
->EmitHighLevelInstructions
) {
1637 (void) new_instruction(emitInfo
, OPCODE_BGNLOOP
);
1641 emit(emitInfo
, n
->Children
[0]);
1644 tailInstLoc
= prog
->NumInstructions
;
1645 if (n
->Children
[1]) {
1646 if (emitInfo
->EmitComments
)
1647 emit_comment(emitInfo
, "Loop tail code:");
1648 emit(emitInfo
, n
->Children
[1]);
1651 endInstLoc
= prog
->NumInstructions
;
1652 if (emitInfo
->EmitHighLevelInstructions
) {
1653 /* emit OPCODE_ENDLOOP */
1654 endInst
= new_instruction(emitInfo
, OPCODE_ENDLOOP
);
1657 /* emit unconditional BRA-nch */
1658 endInst
= new_instruction(emitInfo
, OPCODE_BRA
);
1659 endInst
->DstReg
.CondMask
= COND_TR
; /* always true */
1661 /* ENDLOOP's BranchTarget points to the BGNLOOP inst */
1662 endInst
->BranchTarget
= beginInstLoc
;
1664 if (emitInfo
->EmitHighLevelInstructions
) {
1665 /* BGNLOOP's BranchTarget points to the ENDLOOP inst */
1666 prog
->Instructions
[beginInstLoc
].BranchTarget
= prog
->NumInstructions
-1;
1669 /* Done emitting loop code. Now walk over the loop's linked list of
1670 * BREAK and CONT nodes, filling in their BranchTarget fields (which
1671 * will point to the ENDLOOP+1 or BGNLOOP instructions, respectively).
1673 for (ir
= n
->List
; ir
; ir
= ir
->List
) {
1674 struct prog_instruction
*inst
= prog
->Instructions
+ ir
->InstLocation
;
1675 assert(inst
->BranchTarget
< 0);
1676 if (ir
->Opcode
== IR_BREAK
||
1677 ir
->Opcode
== IR_BREAK_IF_TRUE
) {
1678 assert(inst
->Opcode
== OPCODE_BRK
||
1679 inst
->Opcode
== OPCODE_BRA
);
1680 /* go to instruction after end of loop */
1681 inst
->BranchTarget
= endInstLoc
+ 1;
1684 assert(ir
->Opcode
== IR_CONT
||
1685 ir
->Opcode
== IR_CONT_IF_TRUE
);
1686 assert(inst
->Opcode
== OPCODE_CONT
||
1687 inst
->Opcode
== OPCODE_BRA
);
1688 /* go to instruction at tail of loop */
1689 inst
->BranchTarget
= endInstLoc
;
1697 * Unconditional "continue" or "break" statement.
1698 * Either OPCODE_CONT, OPCODE_BRK or OPCODE_BRA will be emitted.
1700 static struct prog_instruction
*
1701 emit_cont_break(slang_emit_info
*emitInfo
, slang_ir_node
*n
)
1703 gl_inst_opcode opcode
;
1704 struct prog_instruction
*inst
;
1706 if (n
->Opcode
== IR_CONT
) {
1707 /* we need to execute the loop's tail code before doing CONT */
1709 assert(n
->Parent
->Opcode
== IR_LOOP
);
1710 if (n
->Parent
->Children
[1]) {
1711 /* emit tail code */
1712 if (emitInfo
->EmitComments
) {
1713 emit_comment(emitInfo
, "continue - tail code:");
1715 emit(emitInfo
, n
->Parent
->Children
[1]);
1719 /* opcode selection */
1720 if (emitInfo
->EmitHighLevelInstructions
) {
1721 opcode
= (n
->Opcode
== IR_CONT
) ? OPCODE_CONT
: OPCODE_BRK
;
1724 opcode
= OPCODE_BRA
;
1726 n
->InstLocation
= emitInfo
->prog
->NumInstructions
;
1727 inst
= new_instruction(emitInfo
, opcode
);
1728 inst
->DstReg
.CondMask
= COND_TR
; /* always true */
1734 * Conditional "continue" or "break" statement.
1735 * Either OPCODE_CONT, OPCODE_BRK or OPCODE_BRA will be emitted.
1737 static struct prog_instruction
*
1738 emit_cont_break_if_true(slang_emit_info
*emitInfo
, slang_ir_node
*n
)
1740 struct prog_instruction
*inst
;
1742 assert(n
->Opcode
== IR_CONT_IF_TRUE
||
1743 n
->Opcode
== IR_BREAK_IF_TRUE
);
1745 /* evaluate condition expr, setting cond codes */
1746 inst
= emit(emitInfo
, n
->Children
[0]);
1747 if (emitInfo
->EmitCondCodes
) {
1749 inst
->CondUpdate
= GL_TRUE
;
1752 n
->InstLocation
= emitInfo
->prog
->NumInstructions
;
1754 /* opcode selection */
1755 if (emitInfo
->EmitHighLevelInstructions
) {
1756 const gl_inst_opcode opcode
1757 = (n
->Opcode
== IR_CONT_IF_TRUE
) ? OPCODE_CONT
: OPCODE_BRK
;
1758 if (emitInfo
->EmitCondCodes
) {
1759 /* Get the writemask from the previous instruction which set
1760 * the condcodes. Use that writemask as the CondSwizzle.
1762 const GLuint condWritemask
= inst
->DstReg
.WriteMask
;
1763 inst
= new_instruction(emitInfo
, opcode
);
1764 inst
->DstReg
.CondMask
= COND_NE
;
1765 inst
->DstReg
.CondSwizzle
= writemask_to_swizzle(condWritemask
);
1774 ifInstLoc
= emitInfo
->prog
->NumInstructions
;
1775 inst
= emit_instruction(emitInfo
, OPCODE_IF
,
1777 n
->Children
[0]->Store
,
1780 n
->InstLocation
= emitInfo
->prog
->NumInstructions
;
1782 inst
= new_instruction(emitInfo
, opcode
);
1783 inst
= new_instruction(emitInfo
, OPCODE_ENDIF
);
1785 emitInfo
->prog
->Instructions
[ifInstLoc
].BranchTarget
1786 = emitInfo
->prog
->NumInstructions
;
1791 const GLuint condWritemask
= inst
->DstReg
.WriteMask
;
1792 assert(emitInfo
->EmitCondCodes
);
1793 inst
= new_instruction(emitInfo
, OPCODE_BRA
);
1794 inst
->DstReg
.CondMask
= COND_NE
;
1795 inst
->DstReg
.CondSwizzle
= writemask_to_swizzle(condWritemask
);
1801 static struct prog_instruction
*
1802 emit_swizzle(slang_emit_info
*emitInfo
, slang_ir_node
*n
)
1804 struct prog_instruction
*inst
;
1806 inst
= emit(emitInfo
, n
->Children
[0]);
1809 assert(n
->Store
->Parent
);
1810 /* Apply this node's swizzle to parent's storage */
1811 GLuint swizzle
= n
->Store
->Swizzle
;
1812 _slang_copy_ir_storage(n
->Store
, n
->Store
->Parent
);
1813 n
->Store
->Swizzle
= _slang_swizzle_swizzle(n
->Store
->Swizzle
, swizzle
);
1814 assert(!n
->Store
->Parent
);
1821 * Dereference array element: element == array[index]
1822 * This basically involves emitting code for computing the array index
1823 * and updating the node/element's storage info.
1825 static struct prog_instruction
*
1826 emit_array_element(slang_emit_info
*emitInfo
, slang_ir_node
*n
)
1828 slang_ir_storage
*arrayStore
, *indexStore
;
1829 const int elemSize
= n
->Store
->Size
; /* number of floats */
1830 const GLint elemSizeVec
= (elemSize
+ 3) / 4; /* number of vec4 */
1831 struct prog_instruction
*inst
;
1833 assert(n
->Opcode
== IR_ELEMENT
);
1834 assert(elemSize
> 0);
1836 /* special case for built-in state variables, like light state */
1838 slang_ir_storage
*root
= n
->Store
;
1839 assert(!root
->Parent
);
1840 while (root
->Parent
)
1841 root
= root
->Parent
;
1843 if (root
->File
== PROGRAM_STATE_VAR
) {
1844 GLint index
= _slang_alloc_statevar(n
, emitInfo
->prog
->Parameters
);
1845 assert(n
->Store
->Index
== index
);
1850 /* do codegen for array itself */
1851 emit(emitInfo
, n
->Children
[0]);
1852 arrayStore
= n
->Children
[0]->Store
;
1854 /* The initial array element storage is the array's storage,
1855 * then modified below.
1857 _slang_copy_ir_storage(n
->Store
, arrayStore
);
1860 if (n
->Children
[1]->Opcode
== IR_FLOAT
) {
1861 /* Constant array index */
1862 const GLint element
= (GLint
) n
->Children
[1]->Value
[0];
1864 /* this element's storage is the array's storage, plus constant offset */
1865 n
->Store
->Index
+= elemSizeVec
* element
;
1868 /* Variable array index */
1870 /* do codegen for array index expression */
1871 emit(emitInfo
, n
->Children
[1]);
1872 indexStore
= n
->Children
[1]->Store
;
1874 if (indexStore
->IsIndirect
) {
1875 /* need to put the array index into a temporary since we can't
1876 * directly support a[b[i]] constructs.
1880 /*indexStore = tempstore();*/
1885 /* need to multiply array index by array element size */
1886 struct prog_instruction
*inst
;
1887 slang_ir_storage
*indexTemp
;
1888 slang_ir_storage elemSizeStore
;
1890 /* allocate 1 float indexTemp */
1891 indexTemp
= _slang_new_ir_storage(PROGRAM_TEMPORARY
, -1, 1);
1892 _slang_alloc_temp(emitInfo
->vt
, indexTemp
);
1894 /* allocate a constant containing the element size */
1895 constant_to_storage(emitInfo
, (float) elemSizeVec
, &elemSizeStore
);
1897 /* multiply array index by element size */
1898 inst
= emit_instruction(emitInfo
,
1900 indexTemp
, /* dest */
1901 indexStore
, /* the index */
1905 indexStore
= indexTemp
;
1908 if (arrayStore
->IsIndirect
) {
1909 /* ex: in a[i][j], a[i] (the arrayStore) is indirect */
1910 /* Need to add indexStore to arrayStore->Indirect store */
1911 slang_ir_storage indirectArray
;
1912 slang_ir_storage
*indexTemp
;
1914 _slang_init_ir_storage(&indirectArray
,
1915 arrayStore
->IndirectFile
,
1916 arrayStore
->IndirectIndex
,
1918 arrayStore
->IndirectSwizzle
);
1920 /* allocate 1 float indexTemp */
1921 indexTemp
= _slang_new_ir_storage(PROGRAM_TEMPORARY
, -1, 1);
1922 _slang_alloc_temp(emitInfo
->vt
, indexTemp
);
1924 inst
= emit_instruction(emitInfo
,
1926 indexTemp
, /* dest */
1927 indexStore
, /* the index */
1928 &indirectArray
, /* indirect array base */
1931 indexStore
= indexTemp
;
1934 /* update the array element storage info */
1935 n
->Store
->IsIndirect
= GL_TRUE
;
1936 n
->Store
->IndirectFile
= indexStore
->File
;
1937 n
->Store
->IndirectIndex
= indexStore
->Index
;
1938 n
->Store
->IndirectSwizzle
= indexStore
->Swizzle
;
1941 n
->Store
->Size
= elemSize
;
1942 n
->Store
->Swizzle
= _slang_var_swizzle(elemSize
, 0);
1944 return NULL
; /* no instruction */
1949 * Resolve storage for accessing a structure field.
1951 static struct prog_instruction
*
1952 emit_struct_field(slang_emit_info
*emitInfo
, slang_ir_node
*n
)
1954 slang_ir_storage
*root
= n
->Store
;
1955 GLint fieldOffset
, fieldSize
;
1957 assert(n
->Opcode
== IR_FIELD
);
1959 assert(!root
->Parent
);
1960 while (root
->Parent
)
1961 root
= root
->Parent
;
1963 /* If this is the field of a state var, allocate constant/uniform
1964 * storage for it now if we haven't already.
1965 * Note that we allocate storage (uniform/constant slots) for state
1966 * variables here rather than at declaration time so we only allocate
1967 * space for the ones that we actually use!
1969 if (root
->File
== PROGRAM_STATE_VAR
) {
1970 root
->Index
= _slang_alloc_statevar(n
, emitInfo
->prog
->Parameters
);
1971 if (root
->Index
< 0) {
1972 slang_info_log_error(emitInfo
->log
, "Error parsing state variable");
1978 /* do codegen for struct */
1979 emit(emitInfo
, n
->Children
[0]);
1980 assert(n
->Children
[0]->Store
->Index
>= 0);
1983 fieldOffset
= n
->Store
->Index
;
1984 fieldSize
= n
->Store
->Size
;
1986 _slang_copy_ir_storage(n
->Store
, n
->Children
[0]->Store
);
1988 n
->Store
->Index
= n
->Children
[0]->Store
->Index
+ fieldOffset
/ 4;
1990 n->Store->Index += fieldOffset / 4;
1993 switch (fieldSize
) {
1996 GLint swz
= fieldOffset
% 4;
1997 n
->Store
->Swizzle
= MAKE_SWIZZLE4(swz
, swz
, swz
, swz
);
2001 n
->Store
->Swizzle
= MAKE_SWIZZLE4(SWIZZLE_X
, SWIZZLE_Y
,
2002 SWIZZLE_NIL
, SWIZZLE_NIL
);
2005 n
->Store
->Swizzle
= MAKE_SWIZZLE4(SWIZZLE_X
, SWIZZLE_Y
,
2006 SWIZZLE_Z
, SWIZZLE_NIL
);
2009 n
->Store
->Swizzle
= SWIZZLE_XYZW
;
2012 assert(n
->Store
->Index
>= 0);
2014 return NULL
; /* no instruction */
2019 * Emit code for a variable declaration.
2020 * This usually doesn't result in any code generation, but just
2021 * memory allocation.
2023 static struct prog_instruction
*
2024 emit_var_decl(slang_emit_info
*emitInfo
, slang_ir_node
*n
)
2027 assert(n
->Store
->File
!= PROGRAM_UNDEFINED
);
2028 assert(n
->Store
->Size
> 0);
2029 /*assert(n->Store->Index < 0);*/
2031 if (!n
->Var
|| n
->Var
->isTemp
) {
2032 /* a nameless/temporary variable, will be freed after first use */
2034 if (n
->Store
->Index
< 0 && !_slang_alloc_temp(emitInfo
->vt
, n
->Store
)) {
2035 slang_info_log_error(emitInfo
->log
,
2036 "Ran out of registers, too many temporaries");
2041 /* a regular variable */
2042 _slang_add_variable(emitInfo
->vt
, n
->Var
);
2043 if (!_slang_alloc_var(emitInfo
->vt
, n
->Store
)) {
2044 slang_info_log_error(emitInfo
->log
,
2045 "Ran out of registers, too many variables");
2049 printf("IR_VAR_DECL %s %d store %p\n",
2050 (char*) n->Var->a_name, n->Store->Index, (void*) n->Store);
2052 assert(n
->Var
->store
== n
->Store
);
2054 if (emitInfo
->EmitComments
) {
2055 /* emit NOP with comment describing the variable's storage location */
2057 sprintf(s
, "TEMP[%d]%s = variable %s (size %d)",
2059 _mesa_swizzle_string(n
->Store
->Swizzle
, 0, GL_FALSE
),
2060 (n
->Var
? (char *) n
->Var
->a_name
: "anonymous"),
2062 emit_comment(emitInfo
, s
);
2069 * Emit code for a reference to a variable.
2070 * Actually, no code is generated but we may do some memory allocation.
2071 * In particular, state vars (uniforms) are allocated on an as-needed basis.
2073 static struct prog_instruction
*
2074 emit_var_ref(slang_emit_info
*emitInfo
, slang_ir_node
*n
)
2077 assert(n
->Store
->File
!= PROGRAM_UNDEFINED
);
2079 if (n
->Store
->File
== PROGRAM_STATE_VAR
&& n
->Store
->Index
< 0) {
2080 n
->Store
->Index
= _slang_alloc_statevar(n
, emitInfo
->prog
->Parameters
);
2082 else if (n
->Store
->File
== PROGRAM_UNIFORM
) {
2083 /* mark var as used */
2084 _mesa_use_uniform(emitInfo
->prog
->Parameters
, (char *) n
->Var
->a_name
);
2087 if (n
->Store
->Index
< 0) {
2088 /* probably ran out of registers */
2091 assert(n
->Store
->Size
> 0);
2097 static struct prog_instruction
*
2098 emit(slang_emit_info
*emitInfo
, slang_ir_node
*n
)
2100 struct prog_instruction
*inst
;
2104 if (emitInfo
->log
->error_flag
) {
2108 switch (n
->Opcode
) {
2110 /* sequence of two sub-trees */
2111 assert(n
->Children
[0]);
2112 assert(n
->Children
[1]);
2113 emit(emitInfo
, n
->Children
[0]);
2114 if (emitInfo
->log
->error_flag
)
2116 inst
= emit(emitInfo
, n
->Children
[1]);
2120 n
->Store
= n
->Children
[1]->Store
;
2124 /* new variable scope */
2125 _slang_push_var_table(emitInfo
->vt
);
2126 inst
= emit(emitInfo
, n
->Children
[0]);
2127 _slang_pop_var_table(emitInfo
->vt
);
2131 /* Variable declaration - allocate a register for it */
2132 inst
= emit_var_decl(emitInfo
, n
);
2136 /* Reference to a variable
2137 * Storage should have already been resolved/allocated.
2139 return emit_var_ref(emitInfo
, n
);
2142 return emit_array_element(emitInfo
, n
);
2144 return emit_struct_field(emitInfo
, n
);
2146 return emit_swizzle(emitInfo
, n
);
2148 /* Simple arithmetic */
2188 /* trinary operators */
2190 return emit_arith(emitInfo
, n
);
2194 return emit_compare(emitInfo
, n
);
2197 return emit_clamp(emitInfo
, n
);
2201 return emit_tex(emitInfo
, n
);
2203 return emit_negation(emitInfo
, n
);
2205 /* find storage location for this float constant */
2206 n
->Store
->Index
= _mesa_add_unnamed_constant(emitInfo
->prog
->Parameters
,
2209 &n
->Store
->Swizzle
);
2210 if (n
->Store
->Index
< 0) {
2211 slang_info_log_error(emitInfo
->log
, "Ran out of space for constants");
2217 return emit_copy(emitInfo
, n
);
2220 return emit_cond(emitInfo
, n
);
2223 return emit_not(emitInfo
, n
);
2226 return emit_label(emitInfo
, n
);
2229 return emit_kill(emitInfo
);
2232 /* new variable scope for subroutines/function calls */
2233 _slang_push_var_table(emitInfo
->vt
);
2234 inst
= emit_fcall(emitInfo
, n
);
2235 _slang_pop_var_table(emitInfo
->vt
);
2239 return emit_if(emitInfo
, n
);
2242 return emit_loop(emitInfo
, n
);
2243 case IR_BREAK_IF_TRUE
:
2244 case IR_CONT_IF_TRUE
:
2245 return emit_cont_break_if_true(emitInfo
, n
);
2249 return emit_cont_break(emitInfo
, n
);
2252 return new_instruction(emitInfo
, OPCODE_BGNSUB
);
2254 return new_instruction(emitInfo
, OPCODE_ENDSUB
);
2256 return emit_return(emitInfo
, n
);
2262 _mesa_problem(NULL
, "Unexpected IR opcode in emit()\n");
2269 * After code generation, any subroutines will be in separate program
2270 * objects. This function appends all the subroutines onto the main
2271 * program and resolves the linking of all the branch/call instructions.
2272 * XXX this logic should really be part of the linking process...
2275 _slang_resolve_subroutines(slang_emit_info
*emitInfo
)
2277 GET_CURRENT_CONTEXT(ctx
);
2278 struct gl_program
*mainP
= emitInfo
->prog
;
2279 GLuint
*subroutineLoc
, i
, total
;
2282 = (GLuint
*) _mesa_malloc(emitInfo
->NumSubroutines
* sizeof(GLuint
));
2284 /* total number of instructions */
2285 total
= mainP
->NumInstructions
;
2286 for (i
= 0; i
< emitInfo
->NumSubroutines
; i
++) {
2287 subroutineLoc
[i
] = total
;
2288 total
+= emitInfo
->Subroutines
[i
]->NumInstructions
;
2291 /* adjust BranchTargets within the functions */
2292 for (i
= 0; i
< emitInfo
->NumSubroutines
; i
++) {
2293 struct gl_program
*sub
= emitInfo
->Subroutines
[i
];
2295 for (j
= 0; j
< sub
->NumInstructions
; j
++) {
2296 struct prog_instruction
*inst
= sub
->Instructions
+ j
;
2297 if (inst
->Opcode
!= OPCODE_CAL
&& inst
->BranchTarget
>= 0) {
2298 inst
->BranchTarget
+= subroutineLoc
[i
];
2303 /* append subroutines' instructions after main's instructions */
2304 mainP
->Instructions
= _mesa_realloc_instructions(mainP
->Instructions
,
2305 mainP
->NumInstructions
,
2307 mainP
->NumInstructions
= total
;
2308 for (i
= 0; i
< emitInfo
->NumSubroutines
; i
++) {
2309 struct gl_program
*sub
= emitInfo
->Subroutines
[i
];
2310 _mesa_copy_instructions(mainP
->Instructions
+ subroutineLoc
[i
],
2312 sub
->NumInstructions
);
2313 /* delete subroutine code */
2314 sub
->Parameters
= NULL
; /* prevent double-free */
2315 _mesa_reference_program(ctx
, &emitInfo
->Subroutines
[i
], NULL
);
2318 /* free subroutine list */
2319 if (emitInfo
->Subroutines
) {
2320 _mesa_free(emitInfo
->Subroutines
);
2321 emitInfo
->Subroutines
= NULL
;
2323 emitInfo
->NumSubroutines
= 0;
2325 /* Examine CAL instructions.
2326 * At this point, the BranchTarget field of the CAL instruction is
2327 * the number/id of the subroutine to call (an index into the
2328 * emitInfo->Subroutines list).
2329 * Translate that into an actual instruction location now.
2331 for (i
= 0; i
< mainP
->NumInstructions
; i
++) {
2332 struct prog_instruction
*inst
= mainP
->Instructions
+ i
;
2333 if (inst
->Opcode
== OPCODE_CAL
) {
2334 const GLuint f
= inst
->BranchTarget
;
2335 inst
->BranchTarget
= subroutineLoc
[f
];
2339 _mesa_free(subroutineLoc
);
2346 _slang_emit_code(slang_ir_node
*n
, slang_var_table
*vt
,
2347 struct gl_program
*prog
, GLboolean withEnd
,
2348 slang_info_log
*log
)
2350 GET_CURRENT_CONTEXT(ctx
);
2352 slang_emit_info emitInfo
;
2357 emitInfo
.prog
= prog
;
2358 emitInfo
.Subroutines
= NULL
;
2359 emitInfo
.NumSubroutines
= 0;
2360 emitInfo
.MaxInstructions
= prog
->NumInstructions
;
2362 emitInfo
.EmitHighLevelInstructions
= ctx
->Shader
.EmitHighLevelInstructions
;
2363 emitInfo
.EmitCondCodes
= ctx
->Shader
.EmitCondCodes
;
2364 emitInfo
.EmitComments
= ctx
->Shader
.EmitComments
;
2365 emitInfo
.EmitBeginEndSub
= GL_TRUE
;
2367 if (!emitInfo
.EmitCondCodes
) {
2368 emitInfo
.EmitHighLevelInstructions
= GL_TRUE
;
2371 /* Check uniform/constant limits */
2372 if (prog
->Target
== GL_FRAGMENT_PROGRAM_ARB
) {
2373 maxUniforms
= ctx
->Const
.FragmentProgram
.MaxUniformComponents
/ 4;
2376 assert(prog
->Target
== GL_VERTEX_PROGRAM_ARB
);
2377 maxUniforms
= ctx
->Const
.VertexProgram
.MaxUniformComponents
/ 4;
2379 if (prog
->Parameters
->NumParameters
> maxUniforms
) {
2380 slang_info_log_error(log
, "Constant/uniform register limit exceeded");
2384 (void) emit(&emitInfo
, n
);
2386 /* finish up by adding the END opcode to program */
2388 struct prog_instruction
*inst
;
2389 inst
= new_instruction(&emitInfo
, OPCODE_END
);
2392 _slang_resolve_subroutines(&emitInfo
);
2397 printf("*********** End emit code (%u inst):\n", prog
->NumInstructions
);
2398 _mesa_print_program(prog
);
2399 _mesa_print_program_parameters(ctx
,prog
);