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
);
315 dst
->WriteMask
= WRITEMASK_X
<< GET_SWZ(st
->Swizzle
, 0);
318 dst
->WriteMask
= WRITEMASK_XY
;
321 dst
->WriteMask
= WRITEMASK_XYZ
;
324 dst
->WriteMask
= WRITEMASK_XYZW
;
327 ; /* error would have been caught above */
331 dst
->RelAddr
= relAddr
;
336 * Convert IR storage to an instruction src register.
339 storage_to_src_reg(struct prog_src_register
*src
, const slang_ir_storage
*st
)
341 const GLboolean relAddr
= st
->RelAddr
;
342 GLint index
= st
->Index
;
343 GLuint swizzle
= st
->Swizzle
;
345 /* if this is storage relative to some parent storage, walk up the tree */
350 /* an error should have been reported already */
353 assert(st
->Index
>= 0);
355 swizzle
= _slang_swizzle_swizzle(fix_swizzle(st
->Swizzle
), swizzle
);
358 assert(st
->File
>= 0);
359 #if 1 /* XXX temporary */
360 if (st
->File
== PROGRAM_UNDEFINED
) {
361 slang_ir_storage
*st0
= (slang_ir_storage
*) st
;
362 st0
->File
= PROGRAM_TEMPORARY
;
365 assert(st
->File
< PROGRAM_UNDEFINED
);
366 src
->File
= st
->File
;
371 swizzle
= fix_swizzle(swizzle
);
372 assert(GET_SWZ(swizzle
, 0) <= SWIZZLE_W
);
373 assert(GET_SWZ(swizzle
, 1) <= SWIZZLE_W
);
374 assert(GET_SWZ(swizzle
, 2) <= SWIZZLE_W
);
375 assert(GET_SWZ(swizzle
, 3) <= SWIZZLE_W
);
376 src
->Swizzle
= swizzle
;
378 src
->RelAddr
= relAddr
;
383 * Setup storage pointing to a scalar constant/literal.
386 constant_to_storage(slang_emit_info
*emitInfo
,
388 slang_ir_storage
*store
)
395 reg
= _mesa_add_unnamed_constant(emitInfo
->prog
->Parameters
,
398 memset(store
, 0, sizeof(*store
));
399 store
->File
= PROGRAM_CONSTANT
;
401 store
->Swizzle
= swizzle
;
406 * Add new instruction at end of given program.
407 * \param prog the program to append instruction onto
408 * \param opcode opcode for the new instruction
409 * \return pointer to the new instruction
411 static struct prog_instruction
*
412 new_instruction(slang_emit_info
*emitInfo
, gl_inst_opcode opcode
)
414 struct gl_program
*prog
= emitInfo
->prog
;
415 struct prog_instruction
*inst
;
418 /* print prev inst */
419 if (prog
->NumInstructions
> 0) {
420 _mesa_print_instruction(prog
->Instructions
+ prog
->NumInstructions
- 1);
423 assert(prog
->NumInstructions
<= emitInfo
->MaxInstructions
);
425 if (prog
->NumInstructions
== emitInfo
->MaxInstructions
) {
426 /* grow the instruction buffer */
427 emitInfo
->MaxInstructions
+= 20;
429 _mesa_realloc_instructions(prog
->Instructions
,
430 prog
->NumInstructions
,
431 emitInfo
->MaxInstructions
);
434 inst
= prog
->Instructions
+ prog
->NumInstructions
;
435 prog
->NumInstructions
++;
436 _mesa_init_instructions(inst
, 1);
437 inst
->Opcode
= opcode
;
438 inst
->BranchTarget
= -1; /* invalid */
440 printf("New inst %d: %p %s\n", prog->NumInstructions-1,(void*)inst,
441 _mesa_opcode_string(inst->Opcode));
447 static struct prog_instruction
*
448 emit_arl_load(slang_emit_info
*emitInfo
,
449 enum register_file file
, GLint index
, GLuint swizzle
)
451 struct prog_instruction
*inst
= new_instruction(emitInfo
, OPCODE_ARL
);
452 inst
->SrcReg
[0].File
= file
;
453 inst
->SrcReg
[0].Index
= index
;
454 inst
->SrcReg
[0].Swizzle
= swizzle
;
455 inst
->DstReg
.File
= PROGRAM_ADDRESS
;
456 inst
->DstReg
.Index
= 0;
457 inst
->DstReg
.WriteMask
= WRITEMASK_X
;
463 * Emit a new instruction with given opcode, operands.
464 * At this point the instruction may have multiple indirect register
465 * loads/stores. We convert those into ARL loads and address-relative
466 * operands. See comments inside.
467 * At some point in the future we could directly emit indirectly addressed
468 * registers in Mesa GPU instructions.
470 static struct prog_instruction
*
471 emit_instruction(slang_emit_info
*emitInfo
,
472 gl_inst_opcode opcode
,
473 const slang_ir_storage
*dst
,
474 const slang_ir_storage
*src0
,
475 const slang_ir_storage
*src1
,
476 const slang_ir_storage
*src2
)
478 struct prog_instruction
*inst
;
479 GLuint numIndirect
= 0;
480 const slang_ir_storage
*src
[3];
481 slang_ir_storage newSrc
[3], newDst
;
485 isTemp
[0] = isTemp
[1] = isTemp
[2] = GL_FALSE
;
491 /* count up how many operands are indirect loads */
492 for (i
= 0; i
< 3; i
++) {
493 if (src
[i
] && src
[i
]->IsIndirect
)
496 if (dst
&& dst
->IsIndirect
)
499 /* Take special steps for indirect register loads.
500 * If we had multiple address registers this would be simpler.
501 * For example, this GLSL code:
502 * x[i] = y[j] + z[k];
503 * would translate into something like:
507 * ADD TEMP[ADDR.x+5], TEMP[ADDR.y+9], TEMP[ADDR.z+4];
508 * But since we currently only have one address register we have to do this:
510 * MOV t1, TEMP[ADDR.x+9];
512 * MOV t2, TEMP[ADDR.x+4];
514 * ADD TEMP[ADDR.x+5], t1, t2;
515 * The code here figures this out...
517 if (numIndirect
> 0) {
518 for (i
= 0; i
< 3; i
++) {
519 if (src
[i
] && src
[i
]->IsIndirect
) {
520 /* load the ARL register with the indirect register */
521 emit_arl_load(emitInfo
,
522 src
[i
]->IndirectFile
,
523 src
[i
]->IndirectIndex
,
524 src
[i
]->IndirectSwizzle
);
526 if (numIndirect
> 1) {
527 /* Need to load src[i] into a temporary register */
528 slang_ir_storage srcRelAddr
;
529 alloc_local_temp(emitInfo
, &newSrc
[i
], src
[i
]->Size
);
532 /* set RelAddr flag on src register */
533 srcRelAddr
= *src
[i
];
534 srcRelAddr
.RelAddr
= GL_TRUE
;
535 srcRelAddr
.IsIndirect
= GL_FALSE
; /* not really needed */
537 /* MOV newSrc, srcRelAddr; */
538 inst
= emit_instruction(emitInfo
,
548 /* just rewrite the src[i] storage to be ARL-relative */
550 newSrc
[i
].RelAddr
= GL_TRUE
;
551 newSrc
[i
].IsIndirect
= GL_FALSE
; /* not really needed */
558 /* Take special steps for indirect dest register write */
559 if (dst
&& dst
->IsIndirect
) {
560 /* load the ARL register with the indirect register */
561 emit_arl_load(emitInfo
,
564 dst
->IndirectSwizzle
);
566 newDst
.RelAddr
= GL_TRUE
;
567 newDst
.IsIndirect
= GL_FALSE
;
571 /* OK, emit the instruction and its dst, src regs */
572 inst
= new_instruction(emitInfo
, opcode
);
577 storage_to_dst_reg(&inst
->DstReg
, dst
);
579 for (i
= 0; i
< 3; i
++) {
581 storage_to_src_reg(&inst
->SrcReg
[i
], src
[i
]);
584 /* Free any temp registers that we allocated above */
585 for (i
= 0; i
< 3; i
++) {
587 _slang_free_temp(emitInfo
->vt
, &newSrc
[i
]);
596 * Put a comment on the given instruction.
599 inst_comment(struct prog_instruction
*inst
, const char *comment
)
602 inst
->Comment
= _mesa_strdup(comment
);
608 * Return pointer to last instruction in program.
610 static struct prog_instruction
*
611 prev_instruction(slang_emit_info
*emitInfo
)
613 struct gl_program
*prog
= emitInfo
->prog
;
614 if (prog
->NumInstructions
== 0)
617 return prog
->Instructions
+ prog
->NumInstructions
- 1;
621 static struct prog_instruction
*
622 emit(slang_emit_info
*emitInfo
, slang_ir_node
*n
);
626 * Return an annotation string for given node's storage.
629 storage_annotation(const slang_ir_node
*n
, const struct gl_program
*prog
)
632 const slang_ir_storage
*st
= n
->Store
;
633 static char s
[100] = "";
636 return _mesa_strdup("");
639 case PROGRAM_CONSTANT
:
640 if (st
->Index
>= 0) {
641 const GLfloat
*val
= prog
->Parameters
->ParameterValues
[st
->Index
];
642 if (st
->Swizzle
== SWIZZLE_NOOP
)
643 sprintf(s
, "{%g, %g, %g, %g}", val
[0], val
[1], val
[2], val
[3]);
645 sprintf(s
, "%g", val
[GET_SWZ(st
->Swizzle
, 0)]);
649 case PROGRAM_TEMPORARY
:
651 sprintf(s
, "%s", (char *) n
->Var
->a_name
);
653 sprintf(s
, "t[%d]", st
->Index
);
655 case PROGRAM_STATE_VAR
:
656 case PROGRAM_UNIFORM
:
657 sprintf(s
, "%s", prog
->Parameters
->Parameters
[st
->Index
].Name
);
659 case PROGRAM_VARYING
:
660 sprintf(s
, "%s", prog
->Varying
->Parameters
[st
->Index
].Name
);
663 sprintf(s
, "input[%d]", st
->Index
);
666 sprintf(s
, "output[%d]", st
->Index
);
671 return _mesa_strdup(s
);
679 * Return an annotation string for an instruction.
682 instruction_annotation(gl_inst_opcode opcode
, char *dstAnnot
,
683 char *srcAnnot0
, char *srcAnnot1
, char *srcAnnot2
)
686 const char *operator;
691 len
+= strlen(dstAnnot
);
693 dstAnnot
= _mesa_strdup("");
696 len
+= strlen(srcAnnot0
);
698 srcAnnot0
= _mesa_strdup("");
701 len
+= strlen(srcAnnot1
);
703 srcAnnot1
= _mesa_strdup("");
706 len
+= strlen(srcAnnot2
);
708 srcAnnot2
= _mesa_strdup("");
739 s
= (char *) malloc(len
);
740 sprintf(s
, "%s = %s %s %s %s", dstAnnot
,
741 srcAnnot0
, operator, srcAnnot1
, srcAnnot2
);
742 assert(_mesa_strlen(s
) < len
);
757 * Emit an instruction that's just a comment.
759 static struct prog_instruction
*
760 emit_comment(slang_emit_info
*emitInfo
, const char *comment
)
762 struct prog_instruction
*inst
= new_instruction(emitInfo
, OPCODE_NOP
);
763 inst_comment(inst
, comment
);
769 * Generate code for a simple arithmetic instruction.
770 * Either 1, 2 or 3 operands.
772 static struct prog_instruction
*
773 emit_arith(slang_emit_info
*emitInfo
, slang_ir_node
*n
)
775 const slang_ir_info
*info
= _slang_ir_info(n
->Opcode
);
776 struct prog_instruction
*inst
;
780 assert(info
->InstOpcode
!= OPCODE_NOP
);
782 #if PEEPHOLE_OPTIMIZATIONS
783 /* Look for MAD opportunity */
784 if (info
->NumParams
== 2 &&
785 n
->Opcode
== IR_ADD
&& n
->Children
[0]->Opcode
== IR_MUL
) {
786 /* found pattern IR_ADD(IR_MUL(A, B), C) */
787 emit(emitInfo
, n
->Children
[0]->Children
[0]); /* A */
788 emit(emitInfo
, n
->Children
[0]->Children
[1]); /* B */
789 emit(emitInfo
, n
->Children
[1]); /* C */
790 alloc_node_storage(emitInfo
, n
, -1); /* dest */
792 inst
= emit_instruction(emitInfo
,
795 n
->Children
[0]->Children
[0]->Store
,
796 n
->Children
[0]->Children
[1]->Store
,
797 n
->Children
[1]->Store
);
799 free_node_storage(emitInfo
->vt
, n
->Children
[0]->Children
[0]);
800 free_node_storage(emitInfo
->vt
, n
->Children
[0]->Children
[1]);
801 free_node_storage(emitInfo
->vt
, n
->Children
[1]);
805 if (info
->NumParams
== 2 &&
806 n
->Opcode
== IR_ADD
&& n
->Children
[1]->Opcode
== IR_MUL
) {
807 /* found pattern IR_ADD(A, IR_MUL(B, C)) */
808 emit(emitInfo
, n
->Children
[0]); /* A */
809 emit(emitInfo
, n
->Children
[1]->Children
[0]); /* B */
810 emit(emitInfo
, n
->Children
[1]->Children
[1]); /* C */
811 alloc_node_storage(emitInfo
, n
, -1); /* dest */
813 inst
= emit_instruction(emitInfo
,
816 n
->Children
[1]->Children
[0]->Store
,
817 n
->Children
[1]->Children
[1]->Store
,
818 n
->Children
[0]->Store
);
820 free_node_storage(emitInfo
->vt
, n
->Children
[1]->Children
[0]);
821 free_node_storage(emitInfo
->vt
, n
->Children
[1]->Children
[1]);
822 free_node_storage(emitInfo
->vt
, n
->Children
[0]);
827 /* gen code for children, may involve temp allocation */
828 for (i
= 0; i
< info
->NumParams
; i
++) {
829 emit(emitInfo
, n
->Children
[i
]);
830 if (!n
->Children
[i
] || !n
->Children
[i
]->Store
) {
837 alloc_node_storage(emitInfo
, n
, -1);
839 inst
= emit_instruction(emitInfo
,
842 (info
->NumParams
> 0 ? n
->Children
[0]->Store
: NULL
),
843 (info
->NumParams
> 1 ? n
->Children
[1]->Store
: NULL
),
844 (info
->NumParams
> 2 ? n
->Children
[2]->Store
: NULL
)
848 for (i
= 0; i
< info
->NumParams
; i
++)
849 free_node_storage(emitInfo
->vt
, n
->Children
[i
]);
856 * Emit code for == and != operators. These could normally be handled
857 * by emit_arith() except we need to be able to handle structure comparisons.
859 static struct prog_instruction
*
860 emit_compare(slang_emit_info
*emitInfo
, slang_ir_node
*n
)
862 struct prog_instruction
*inst
= NULL
;
865 assert(n
->Opcode
== IR_EQUAL
|| n
->Opcode
== IR_NOTEQUAL
);
867 /* gen code for children */
868 emit(emitInfo
, n
->Children
[0]);
869 emit(emitInfo
, n
->Children
[1]);
871 if (n
->Children
[0]->Store
->Size
!= n
->Children
[1]->Store
->Size
) {
872 slang_info_log_error(emitInfo
->log
, "invalid operands to == or !=");
876 /* final result is 1 bool */
877 if (!alloc_node_storage(emitInfo
, n
, 1))
880 size
= n
->Children
[0]->Store
->Size
;
883 gl_inst_opcode opcode
= n
->Opcode
== IR_EQUAL
? OPCODE_SEQ
: OPCODE_SNE
;
884 inst
= emit_instruction(emitInfo
,
887 n
->Children
[0]->Store
,
888 n
->Children
[1]->Store
,
891 else if (size
<= 4) {
892 /* compare two vectors.
893 * Unfortunately, there's no instruction to compare vectors and
894 * return a scalar result. Do it with some compare and dot product
898 gl_inst_opcode dotOp
;
899 slang_ir_storage tempStore
;
901 if (!alloc_local_temp(emitInfo
, &tempStore
, 4)) {
908 swizzle
= SWIZZLE_XYZW
;
910 else if (size
== 3) {
912 swizzle
= SWIZZLE_XYZW
;
917 swizzle
= MAKE_SWIZZLE4(SWIZZLE_X
, SWIZZLE_Y
, SWIZZLE_Y
, SWIZZLE_Y
);
920 /* Compute inequality (temp = (A != B)) */
921 inst
= emit_instruction(emitInfo
,
924 n
->Children
[0]->Store
,
925 n
->Children
[1]->Store
,
927 inst_comment(inst
, "Compare values");
929 /* Compute val = DOT(temp, temp) (reduction) */
930 inst
= emit_instruction(emitInfo
,
936 inst
->SrcReg
[0].Swizzle
= inst
->SrcReg
[1].Swizzle
= swizzle
; /*override*/
937 inst_comment(inst
, "Reduce vec to bool");
939 _slang_free_temp(emitInfo
->vt
, &tempStore
); /* free temp */
941 if (n
->Opcode
== IR_EQUAL
) {
942 /* compute val = !val.x with SEQ val, val, 0; */
943 slang_ir_storage zero
;
944 constant_to_storage(emitInfo
, 0.0, &zero
);
945 inst
= emit_instruction(emitInfo
,
951 inst_comment(inst
, "Invert true/false");
955 /* size > 4, struct or array compare.
956 * XXX this won't work reliably for structs with padding!!
958 GLint i
, num
= (n
->Children
[0]->Store
->Size
+ 3) / 4;
959 slang_ir_storage accTemp
, sneTemp
;
961 if (!alloc_local_temp(emitInfo
, &accTemp
, 4))
964 if (!alloc_local_temp(emitInfo
, &sneTemp
, 4))
967 for (i
= 0; i
< num
; i
++) {
968 slang_ir_storage srcStore0
= *n
->Children
[0]->Store
;
969 slang_ir_storage srcStore1
= *n
->Children
[1]->Store
;
970 srcStore0
.Index
+= i
;
971 srcStore1
.Index
+= i
;
974 /* SNE accTemp, left[i], right[i] */
975 inst
= emit_instruction(emitInfo
, OPCODE_SNE
,
980 inst_comment(inst
, "Begin struct/array comparison");
983 /* SNE sneTemp, left[i], right[i] */
984 inst
= emit_instruction(emitInfo
, OPCODE_SNE
,
989 /* ADD accTemp, accTemp, sneTemp; # like logical-OR */
990 inst
= emit_instruction(emitInfo
, OPCODE_ADD
,
998 /* compute accTemp.x || accTemp.y || accTemp.z || accTemp.w with DOT4 */
999 inst
= emit_instruction(emitInfo
, OPCODE_DP4
,
1004 inst_comment(inst
, "End struct/array comparison");
1006 if (n
->Opcode
== IR_EQUAL
) {
1007 /* compute tmp.x = !tmp.x via tmp.x = (tmp.x == 0) */
1008 slang_ir_storage zero
;
1009 constant_to_storage(emitInfo
, 0.0, &zero
);
1010 inst
= emit_instruction(emitInfo
, OPCODE_SEQ
,
1011 n
->Store
, /* dest */
1015 inst_comment(inst
, "Invert true/false");
1018 _slang_free_temp(emitInfo
->vt
, &accTemp
);
1019 _slang_free_temp(emitInfo
->vt
, &sneTemp
);
1023 free_node_storage(emitInfo
->vt
, n
->Children
[0]);
1024 free_node_storage(emitInfo
->vt
, n
->Children
[1]);
1032 * Generate code for an IR_CLAMP instruction.
1034 static struct prog_instruction
*
1035 emit_clamp(slang_emit_info
*emitInfo
, slang_ir_node
*n
)
1037 struct prog_instruction
*inst
;
1038 slang_ir_node tmpNode
;
1040 assert(n
->Opcode
== IR_CLAMP
);
1046 inst
= emit(emitInfo
, n
->Children
[0]);
1048 /* If lower limit == 0.0 and upper limit == 1.0,
1049 * set prev instruction's SaturateMode field to SATURATE_ZERO_ONE.
1051 * emit OPCODE_MIN, OPCODE_MAX sequence.
1054 /* XXX this isn't quite finished yet */
1055 if (n
->Children
[1]->Opcode
== IR_FLOAT
&&
1056 n
->Children
[1]->Value
[0] == 0.0 &&
1057 n
->Children
[1]->Value
[1] == 0.0 &&
1058 n
->Children
[1]->Value
[2] == 0.0 &&
1059 n
->Children
[1]->Value
[3] == 0.0 &&
1060 n
->Children
[2]->Opcode
== IR_FLOAT
&&
1061 n
->Children
[2]->Value
[0] == 1.0 &&
1062 n
->Children
[2]->Value
[1] == 1.0 &&
1063 n
->Children
[2]->Value
[2] == 1.0 &&
1064 n
->Children
[2]->Value
[3] == 1.0) {
1066 inst
= prev_instruction(prog
);
1068 if (inst
&& inst
->Opcode
!= OPCODE_NOP
) {
1069 /* and prev instruction's DstReg matches n->Children[0]->Store */
1070 inst
->SaturateMode
= SATURATE_ZERO_ONE
;
1071 n
->Store
= n
->Children
[0]->Store
;
1077 if (!alloc_node_storage(emitInfo
, n
, n
->Children
[0]->Store
->Size
))
1080 emit(emitInfo
, n
->Children
[1]);
1081 emit(emitInfo
, n
->Children
[2]);
1083 /* Some GPUs don't allow reading from output registers. So if the
1084 * dest for this clamp() is an output reg, we can't use that reg for
1085 * the intermediate result. Use a temp register instead.
1087 _mesa_bzero(&tmpNode
, sizeof(tmpNode
));
1088 alloc_node_storage(emitInfo
, &tmpNode
, n
->Store
->Size
);
1090 /* tmp = max(ch[0], ch[1]) */
1091 inst
= emit_instruction(emitInfo
, OPCODE_MAX
,
1092 tmpNode
.Store
, /* dest */
1093 n
->Children
[0]->Store
,
1094 n
->Children
[1]->Store
,
1097 /* n->dest = min(tmp, ch[2]) */
1098 inst
= emit_instruction(emitInfo
, OPCODE_MIN
,
1099 n
->Store
, /* dest */
1101 n
->Children
[2]->Store
,
1104 free_node_storage(emitInfo
->vt
, &tmpNode
);
1110 static struct prog_instruction
*
1111 emit_negation(slang_emit_info
*emitInfo
, slang_ir_node
*n
)
1113 /* Implement as MOV dst, -src; */
1114 /* XXX we could look at the previous instruction and in some circumstances
1115 * modify it to accomplish the negation.
1117 struct prog_instruction
*inst
;
1119 emit(emitInfo
, n
->Children
[0]);
1121 if (!alloc_node_storage(emitInfo
, n
, n
->Children
[0]->Store
->Size
))
1124 inst
= emit_instruction(emitInfo
,
1126 n
->Store
, /* dest */
1127 n
->Children
[0]->Store
,
1130 inst
->SrcReg
[0].NegateBase
= NEGATE_XYZW
;
1135 static struct prog_instruction
*
1136 emit_label(slang_emit_info
*emitInfo
, const slang_ir_node
*n
)
1140 /* XXX this fails in loop tail code - investigate someday */
1141 assert(_slang_label_get_location(n
->Label
) < 0);
1142 _slang_label_set_location(n
->Label
, emitInfo
->prog
->NumInstructions
,
1145 if (_slang_label_get_location(n
->Label
) < 0)
1146 _slang_label_set_location(n
->Label
, emitInfo
->prog
->NumInstructions
,
1154 * Emit code for a function call.
1155 * Note that for each time a function is called, we emit the function's
1156 * body code again because the set of available registers may be different.
1158 static struct prog_instruction
*
1159 emit_fcall(slang_emit_info
*emitInfo
, slang_ir_node
*n
)
1161 struct gl_program
*progSave
;
1162 struct prog_instruction
*inst
;
1163 GLuint subroutineId
;
1166 assert(n
->Opcode
== IR_CALL
);
1169 /* save/push cur program */
1170 maxInstSave
= emitInfo
->MaxInstructions
;
1171 progSave
= emitInfo
->prog
;
1173 emitInfo
->prog
= new_subroutine(emitInfo
, &subroutineId
);
1174 emitInfo
->MaxInstructions
= emitInfo
->prog
->NumInstructions
;
1176 _slang_label_set_location(n
->Label
, emitInfo
->prog
->NumInstructions
,
1179 if (emitInfo
->EmitBeginEndSub
) {
1180 /* BGNSUB isn't a real instruction.
1181 * We require a label (i.e. "foobar:") though, if we're going to
1182 * print the program in the NV format. The BNGSUB instruction is
1183 * really just a NOP to attach the label to.
1185 inst
= new_instruction(emitInfo
, OPCODE_BGNSUB
);
1186 inst_comment(inst
, n
->Label
->Name
);
1189 /* body of function: */
1190 emit(emitInfo
, n
->Children
[0]);
1191 n
->Store
= n
->Children
[0]->Store
;
1193 /* add RET instruction now, if needed */
1194 inst
= prev_instruction(emitInfo
);
1195 if (inst
&& inst
->Opcode
!= OPCODE_RET
) {
1196 inst
= new_instruction(emitInfo
, OPCODE_RET
);
1199 if (emitInfo
->EmitBeginEndSub
) {
1200 inst
= new_instruction(emitInfo
, OPCODE_ENDSUB
);
1201 inst_comment(inst
, n
->Label
->Name
);
1204 /* pop/restore cur program */
1205 emitInfo
->prog
= progSave
;
1206 emitInfo
->MaxInstructions
= maxInstSave
;
1208 /* emit the function call */
1209 inst
= new_instruction(emitInfo
, OPCODE_CAL
);
1210 /* The branch target is just the subroutine number (changed later) */
1211 inst
->BranchTarget
= subroutineId
;
1212 inst_comment(inst
, n
->Label
->Name
);
1213 assert(inst
->BranchTarget
>= 0);
1220 * Emit code for a 'return' statement.
1222 static struct prog_instruction
*
1223 emit_return(slang_emit_info
*emitInfo
, slang_ir_node
*n
)
1225 struct prog_instruction
*inst
;
1227 assert(n
->Opcode
== IR_RETURN
);
1229 inst
= new_instruction(emitInfo
, OPCODE_RET
);
1230 inst
->DstReg
.CondMask
= COND_TR
; /* always return */
1235 static struct prog_instruction
*
1236 emit_kill(slang_emit_info
*emitInfo
)
1238 struct gl_fragment_program
*fp
;
1239 struct prog_instruction
*inst
;
1240 /* NV-KILL - discard fragment depending on condition code.
1241 * Note that ARB-KILL depends on sign of vector operand.
1243 inst
= new_instruction(emitInfo
, OPCODE_KIL_NV
);
1244 inst
->DstReg
.CondMask
= COND_TR
; /* always kill */
1246 assert(emitInfo
->prog
->Target
== GL_FRAGMENT_PROGRAM_ARB
);
1247 fp
= (struct gl_fragment_program
*) emitInfo
->prog
;
1248 fp
->UsesKill
= GL_TRUE
;
1254 static struct prog_instruction
*
1255 emit_tex(slang_emit_info
*emitInfo
, slang_ir_node
*n
)
1257 struct prog_instruction
*inst
;
1258 gl_inst_opcode opcode
;
1260 if (n
->Opcode
== IR_TEX
) {
1261 opcode
= OPCODE_TEX
;
1263 else if (n
->Opcode
== IR_TEXB
) {
1264 opcode
= OPCODE_TXB
;
1267 assert(n
->Opcode
== IR_TEXP
);
1268 opcode
= OPCODE_TXP
;
1271 /* emit code for the texcoord operand */
1272 (void) emit(emitInfo
, n
->Children
[1]);
1274 /* alloc storage for result of texture fetch */
1275 if (!alloc_node_storage(emitInfo
, n
, 4))
1278 /* emit TEX instruction; Child[1] is the texcoord */
1279 inst
= emit_instruction(emitInfo
,
1282 n
->Children
[1]->Store
,
1286 /* Child[0] is the sampler (a uniform which'll indicate the texture unit) */
1287 assert(n
->Children
[0]->Store
);
1288 assert(n
->Children
[0]->Store
->File
== PROGRAM_SAMPLER
);
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 */
1298 /* mark the sampler as being used */
1299 _mesa_use_uniform(emitInfo
->prog
->Parameters
,
1300 (char *) n
->Children
[0]->Var
->a_name
);
1309 static struct prog_instruction
*
1310 emit_copy(slang_emit_info
*emitInfo
, slang_ir_node
*n
)
1312 struct prog_instruction
*inst
;
1314 assert(n
->Opcode
== IR_COPY
);
1317 emit(emitInfo
, n
->Children
[0]);
1318 if (!n
->Children
[0]->Store
|| n
->Children
[0]->Store
->Index
< 0) {
1319 /* an error should have been already recorded */
1324 assert(n
->Children
[1]);
1325 inst
= emit(emitInfo
, n
->Children
[1]);
1327 if (!n
->Children
[1]->Store
|| n
->Children
[1]->Store
->Index
< 0) {
1328 if (!emitInfo
->log
->text
) {
1329 slang_info_log_error(emitInfo
->log
, "invalid assignment");
1334 assert(n
->Children
[1]->Store
->Index
>= 0);
1336 /*assert(n->Children[0]->Store->Size == n->Children[1]->Store->Size);*/
1338 n
->Store
= n
->Children
[0]->Store
;
1340 if (n
->Store
->File
== PROGRAM_SAMPLER
) {
1341 /* no code generated for sampler assignments,
1342 * just copy the sampler index at compile time.
1344 n
->Store
->Index
= n
->Children
[1]->Store
->Index
;
1348 #if PEEPHOLE_OPTIMIZATIONS
1350 _slang_is_temp(emitInfo
->vt
, n
->Children
[1]->Store
) &&
1351 (inst
->DstReg
.File
== n
->Children
[1]->Store
->File
) &&
1352 (inst
->DstReg
.Index
== n
->Children
[1]->Store
->Index
) &&
1353 !n
->Children
[0]->Store
->IsIndirect
&&
1354 n
->Children
[0]->Store
->Size
<= 4) {
1355 /* Peephole optimization:
1356 * The Right-Hand-Side has its results in a temporary place.
1357 * Modify the RHS (and the prev instruction) to store its results
1358 * in the destination specified by n->Children[0].
1359 * Then, this MOVE is a no-op.
1366 if (n
->Children
[1]->Opcode
!= IR_SWIZZLE
)
1367 _slang_free_temp(emitInfo
->vt
, n
->Children
[1]->Store
);
1368 *n
->Children
[1]->Store
= *n
->Children
[0]->Store
;
1370 /* fixup the previous instruction (which stored the RHS result) */
1371 assert(n
->Children
[0]->Store
->Index
>= 0);
1373 storage_to_dst_reg(&inst
->DstReg
, n
->Children
[0]->Store
);
1379 if (n
->Children
[0]->Store
->Size
> 4) {
1380 /* move matrix/struct etc (block of registers) */
1381 slang_ir_storage dstStore
= *n
->Children
[0]->Store
;
1382 slang_ir_storage srcStore
= *n
->Children
[1]->Store
;
1383 GLint size
= srcStore
.Size
;
1384 ASSERT(n
->Children
[1]->Store
->Swizzle
== SWIZZLE_NOOP
);
1388 inst
= emit_instruction(emitInfo
, OPCODE_MOV
,
1393 inst_comment(inst
, "IR_COPY block");
1400 /* single register move */
1401 char *srcAnnot
, *dstAnnot
;
1402 assert(n
->Children
[0]->Store
->Index
>= 0);
1403 inst
= emit_instruction(emitInfo
, OPCODE_MOV
,
1404 n
->Children
[0]->Store
, /* dest */
1405 n
->Children
[1]->Store
,
1408 dstAnnot
= storage_annotation(n
->Children
[0], emitInfo
->prog
);
1409 srcAnnot
= storage_annotation(n
->Children
[1], emitInfo
->prog
);
1410 inst
->Comment
= instruction_annotation(inst
->Opcode
, dstAnnot
,
1411 srcAnnot
, NULL
, NULL
);
1413 free_node_storage(emitInfo
->vt
, n
->Children
[1]);
1420 * An IR_COND node wraps a boolean expression which is used by an
1421 * IF or WHILE test. This is where we'll set condition codes, if needed.
1423 static struct prog_instruction
*
1424 emit_cond(slang_emit_info
*emitInfo
, slang_ir_node
*n
)
1426 struct prog_instruction
*inst
;
1428 assert(n
->Opcode
== IR_COND
);
1430 if (!n
->Children
[0])
1433 /* emit code for the expression */
1434 inst
= emit(emitInfo
, n
->Children
[0]);
1436 if (!n
->Children
[0]->Store
) {
1437 /* error recovery */
1441 assert(n
->Children
[0]->Store
);
1442 /*assert(n->Children[0]->Store->Size == 1);*/
1444 if (emitInfo
->EmitCondCodes
) {
1446 n
->Children
[0]->Store
&&
1447 inst
->DstReg
.File
== n
->Children
[0]->Store
->File
&&
1448 inst
->DstReg
.Index
== n
->Children
[0]->Store
->Index
) {
1449 /* The previous instruction wrote to the register who's value
1450 * we're testing. Just fix that instruction so that the
1451 * condition codes are computed.
1453 inst
->CondUpdate
= GL_TRUE
;
1454 n
->Store
= n
->Children
[0]->Store
;
1458 /* This'll happen for things like "if (i) ..." where no code
1459 * is normally generated for the expression "i".
1460 * Generate a move instruction just to set condition codes.
1462 if (!alloc_node_storage(emitInfo
, n
, 1))
1464 inst
= emit_instruction(emitInfo
, OPCODE_MOV
,
1465 n
->Store
, /* dest */
1466 n
->Children
[0]->Store
,
1469 inst
->CondUpdate
= GL_TRUE
;
1470 inst_comment(inst
, "COND expr");
1471 _slang_free_temp(emitInfo
->vt
, n
->Store
);
1476 /* No-op: the boolean result of the expression is in a regular reg */
1477 n
->Store
= n
->Children
[0]->Store
;
1486 static struct prog_instruction
*
1487 emit_not(slang_emit_info
*emitInfo
, slang_ir_node
*n
)
1489 static const struct {
1490 gl_inst_opcode op
, opNot
;
1492 { OPCODE_SLT
, OPCODE_SGE
},
1493 { OPCODE_SLE
, OPCODE_SGT
},
1494 { OPCODE_SGT
, OPCODE_SLE
},
1495 { OPCODE_SGE
, OPCODE_SLT
},
1496 { OPCODE_SEQ
, OPCODE_SNE
},
1497 { OPCODE_SNE
, OPCODE_SEQ
},
1500 struct prog_instruction
*inst
;
1501 slang_ir_storage zero
;
1505 inst
= emit(emitInfo
, n
->Children
[0]);
1507 #if PEEPHOLE_OPTIMIZATIONS
1509 /* if the prev instruction was a comparison instruction, invert it */
1510 for (i
= 0; operators
[i
].op
; i
++) {
1511 if (inst
->Opcode
== operators
[i
].op
) {
1512 inst
->Opcode
= operators
[i
].opNot
;
1513 n
->Store
= n
->Children
[0]->Store
;
1520 /* else, invert using SEQ (v = v == 0) */
1521 if (!alloc_node_storage(emitInfo
, n
, n
->Children
[0]->Store
->Size
))
1524 constant_to_storage(emitInfo
, 0.0, &zero
);
1525 inst
= emit_instruction(emitInfo
,
1528 n
->Children
[0]->Store
,
1531 inst_comment(inst
, "NOT");
1533 free_node_storage(emitInfo
->vt
, n
->Children
[0]);
1539 static struct prog_instruction
*
1540 emit_if(slang_emit_info
*emitInfo
, slang_ir_node
*n
)
1542 struct gl_program
*prog
= emitInfo
->prog
;
1543 GLuint ifInstLoc
, elseInstLoc
= 0;
1544 GLuint condWritemask
= 0;
1546 /* emit condition expression code */
1548 struct prog_instruction
*inst
;
1549 inst
= emit(emitInfo
, n
->Children
[0]);
1550 if (emitInfo
->EmitCondCodes
) {
1552 /* error recovery */
1555 condWritemask
= inst
->DstReg
.WriteMask
;
1559 if (!n
->Children
[0]->Store
)
1563 assert(n
->Children
[0]->Store
->Size
== 1); /* a bool! */
1566 ifInstLoc
= prog
->NumInstructions
;
1567 if (emitInfo
->EmitHighLevelInstructions
) {
1568 if (emitInfo
->EmitCondCodes
) {
1569 /* IF condcode THEN ... */
1570 struct prog_instruction
*ifInst
;
1571 ifInst
= new_instruction(emitInfo
, OPCODE_IF
);
1572 ifInst
->DstReg
.CondMask
= COND_NE
; /* if cond is non-zero */
1573 /* only test the cond code (1 of 4) that was updated by the
1574 * previous instruction.
1576 ifInst
->DstReg
.CondSwizzle
= writemask_to_swizzle(condWritemask
);
1579 /* IF src[0] THEN ... */
1580 emit_instruction(emitInfo
, OPCODE_IF
,
1582 n
->Children
[0]->Store
, /* op0 */
1588 /* conditional jump to else, or endif */
1589 struct prog_instruction
*ifInst
= new_instruction(emitInfo
, OPCODE_BRA
);
1590 ifInst
->DstReg
.CondMask
= COND_EQ
; /* BRA if cond is zero */
1591 inst_comment(ifInst
, "if zero");
1592 ifInst
->DstReg
.CondSwizzle
= writemask_to_swizzle(condWritemask
);
1596 emit(emitInfo
, n
->Children
[1]);
1598 if (n
->Children
[2]) {
1599 /* have else body */
1600 elseInstLoc
= prog
->NumInstructions
;
1601 if (emitInfo
->EmitHighLevelInstructions
) {
1602 (void) new_instruction(emitInfo
, OPCODE_ELSE
);
1605 /* jump to endif instruction */
1606 struct prog_instruction
*inst
;
1607 inst
= new_instruction(emitInfo
, OPCODE_BRA
);
1608 inst_comment(inst
, "else");
1609 inst
->DstReg
.CondMask
= COND_TR
; /* always branch */
1611 prog
->Instructions
[ifInstLoc
].BranchTarget
= prog
->NumInstructions
;
1612 emit(emitInfo
, n
->Children
[2]);
1616 prog
->Instructions
[ifInstLoc
].BranchTarget
= prog
->NumInstructions
;
1619 if (emitInfo
->EmitHighLevelInstructions
) {
1620 (void) new_instruction(emitInfo
, OPCODE_ENDIF
);
1623 if (n
->Children
[2]) {
1624 prog
->Instructions
[elseInstLoc
].BranchTarget
= prog
->NumInstructions
;
1630 static struct prog_instruction
*
1631 emit_loop(slang_emit_info
*emitInfo
, slang_ir_node
*n
)
1633 struct gl_program
*prog
= emitInfo
->prog
;
1634 struct prog_instruction
*endInst
;
1635 GLuint beginInstLoc
, tailInstLoc
, endInstLoc
;
1638 /* emit OPCODE_BGNLOOP */
1639 beginInstLoc
= prog
->NumInstructions
;
1640 if (emitInfo
->EmitHighLevelInstructions
) {
1641 (void) new_instruction(emitInfo
, OPCODE_BGNLOOP
);
1645 emit(emitInfo
, n
->Children
[0]);
1648 tailInstLoc
= prog
->NumInstructions
;
1649 if (n
->Children
[1]) {
1650 if (emitInfo
->EmitComments
)
1651 emit_comment(emitInfo
, "Loop tail code:");
1652 emit(emitInfo
, n
->Children
[1]);
1655 endInstLoc
= prog
->NumInstructions
;
1656 if (emitInfo
->EmitHighLevelInstructions
) {
1657 /* emit OPCODE_ENDLOOP */
1658 endInst
= new_instruction(emitInfo
, OPCODE_ENDLOOP
);
1661 /* emit unconditional BRA-nch */
1662 endInst
= new_instruction(emitInfo
, OPCODE_BRA
);
1663 endInst
->DstReg
.CondMask
= COND_TR
; /* always true */
1665 /* ENDLOOP's BranchTarget points to the BGNLOOP inst */
1666 endInst
->BranchTarget
= beginInstLoc
;
1668 if (emitInfo
->EmitHighLevelInstructions
) {
1669 /* BGNLOOP's BranchTarget points to the ENDLOOP inst */
1670 prog
->Instructions
[beginInstLoc
].BranchTarget
= prog
->NumInstructions
-1;
1673 /* Done emitting loop code. Now walk over the loop's linked list of
1674 * BREAK and CONT nodes, filling in their BranchTarget fields (which
1675 * will point to the ENDLOOP+1 or BGNLOOP instructions, respectively).
1677 for (ir
= n
->List
; ir
; ir
= ir
->List
) {
1678 struct prog_instruction
*inst
= prog
->Instructions
+ ir
->InstLocation
;
1679 assert(inst
->BranchTarget
< 0);
1680 if (ir
->Opcode
== IR_BREAK
||
1681 ir
->Opcode
== IR_BREAK_IF_TRUE
) {
1682 assert(inst
->Opcode
== OPCODE_BRK
||
1683 inst
->Opcode
== OPCODE_BRA
);
1684 /* go to instruction after end of loop */
1685 inst
->BranchTarget
= endInstLoc
+ 1;
1688 assert(ir
->Opcode
== IR_CONT
||
1689 ir
->Opcode
== IR_CONT_IF_TRUE
);
1690 assert(inst
->Opcode
== OPCODE_CONT
||
1691 inst
->Opcode
== OPCODE_BRA
);
1692 /* go to instruction at tail of loop */
1693 inst
->BranchTarget
= endInstLoc
;
1701 * Unconditional "continue" or "break" statement.
1702 * Either OPCODE_CONT, OPCODE_BRK or OPCODE_BRA will be emitted.
1704 static struct prog_instruction
*
1705 emit_cont_break(slang_emit_info
*emitInfo
, slang_ir_node
*n
)
1707 gl_inst_opcode opcode
;
1708 struct prog_instruction
*inst
;
1710 if (n
->Opcode
== IR_CONT
) {
1711 /* we need to execute the loop's tail code before doing CONT */
1713 assert(n
->Parent
->Opcode
== IR_LOOP
);
1714 if (n
->Parent
->Children
[1]) {
1715 /* emit tail code */
1716 if (emitInfo
->EmitComments
) {
1717 emit_comment(emitInfo
, "continue - tail code:");
1719 emit(emitInfo
, n
->Parent
->Children
[1]);
1723 /* opcode selection */
1724 if (emitInfo
->EmitHighLevelInstructions
) {
1725 opcode
= (n
->Opcode
== IR_CONT
) ? OPCODE_CONT
: OPCODE_BRK
;
1728 opcode
= OPCODE_BRA
;
1730 n
->InstLocation
= emitInfo
->prog
->NumInstructions
;
1731 inst
= new_instruction(emitInfo
, opcode
);
1732 inst
->DstReg
.CondMask
= COND_TR
; /* always true */
1738 * Conditional "continue" or "break" statement.
1739 * Either OPCODE_CONT, OPCODE_BRK or OPCODE_BRA will be emitted.
1741 static struct prog_instruction
*
1742 emit_cont_break_if_true(slang_emit_info
*emitInfo
, slang_ir_node
*n
)
1744 struct prog_instruction
*inst
;
1746 assert(n
->Opcode
== IR_CONT_IF_TRUE
||
1747 n
->Opcode
== IR_BREAK_IF_TRUE
);
1749 /* evaluate condition expr, setting cond codes */
1750 inst
= emit(emitInfo
, n
->Children
[0]);
1751 if (emitInfo
->EmitCondCodes
) {
1753 inst
->CondUpdate
= GL_TRUE
;
1756 n
->InstLocation
= emitInfo
->prog
->NumInstructions
;
1758 /* opcode selection */
1759 if (emitInfo
->EmitHighLevelInstructions
) {
1760 const gl_inst_opcode opcode
1761 = (n
->Opcode
== IR_CONT_IF_TRUE
) ? OPCODE_CONT
: OPCODE_BRK
;
1762 if (emitInfo
->EmitCondCodes
) {
1763 /* Get the writemask from the previous instruction which set
1764 * the condcodes. Use that writemask as the CondSwizzle.
1766 const GLuint condWritemask
= inst
->DstReg
.WriteMask
;
1767 inst
= new_instruction(emitInfo
, opcode
);
1768 inst
->DstReg
.CondMask
= COND_NE
;
1769 inst
->DstReg
.CondSwizzle
= writemask_to_swizzle(condWritemask
);
1778 ifInstLoc
= emitInfo
->prog
->NumInstructions
;
1779 inst
= emit_instruction(emitInfo
, OPCODE_IF
,
1781 n
->Children
[0]->Store
,
1784 n
->InstLocation
= emitInfo
->prog
->NumInstructions
;
1786 inst
= new_instruction(emitInfo
, opcode
);
1787 inst
= new_instruction(emitInfo
, OPCODE_ENDIF
);
1789 emitInfo
->prog
->Instructions
[ifInstLoc
].BranchTarget
1790 = emitInfo
->prog
->NumInstructions
;
1795 const GLuint condWritemask
= inst
->DstReg
.WriteMask
;
1796 assert(emitInfo
->EmitCondCodes
);
1797 inst
= new_instruction(emitInfo
, OPCODE_BRA
);
1798 inst
->DstReg
.CondMask
= COND_NE
;
1799 inst
->DstReg
.CondSwizzle
= writemask_to_swizzle(condWritemask
);
1805 static struct prog_instruction
*
1806 emit_swizzle(slang_emit_info
*emitInfo
, slang_ir_node
*n
)
1808 struct prog_instruction
*inst
;
1810 inst
= emit(emitInfo
, n
->Children
[0]);
1813 assert(n
->Store
->Parent
);
1814 /* Apply this node's swizzle to parent's storage */
1815 GLuint swizzle
= n
->Store
->Swizzle
;
1816 _slang_copy_ir_storage(n
->Store
, n
->Store
->Parent
);
1817 n
->Store
->Swizzle
= _slang_swizzle_swizzle(n
->Store
->Swizzle
, swizzle
);
1818 assert(!n
->Store
->Parent
);
1825 * Dereference array element: element == array[index]
1826 * This basically involves emitting code for computing the array index
1827 * and updating the node/element's storage info.
1829 static struct prog_instruction
*
1830 emit_array_element(slang_emit_info
*emitInfo
, slang_ir_node
*n
)
1832 slang_ir_storage
*arrayStore
, *indexStore
;
1833 const int elemSize
= n
->Store
->Size
; /* number of floats */
1834 const GLint elemSizeVec
= (elemSize
+ 3) / 4; /* number of vec4 */
1835 struct prog_instruction
*inst
;
1837 assert(n
->Opcode
== IR_ELEMENT
);
1838 assert(elemSize
> 0);
1840 /* special case for built-in state variables, like light state */
1842 slang_ir_storage
*root
= n
->Store
;
1843 assert(!root
->Parent
);
1844 while (root
->Parent
)
1845 root
= root
->Parent
;
1847 if (root
->File
== PROGRAM_STATE_VAR
) {
1850 _slang_alloc_statevar(n
, emitInfo
->prog
->Parameters
, &direct
);
1856 n
->Store
->Index
= index
;
1857 return NULL
; /* all done */
1862 /* do codegen for array itself */
1863 emit(emitInfo
, n
->Children
[0]);
1864 arrayStore
= n
->Children
[0]->Store
;
1866 /* The initial array element storage is the array's storage,
1867 * then modified below.
1869 _slang_copy_ir_storage(n
->Store
, arrayStore
);
1872 if (n
->Children
[1]->Opcode
== IR_FLOAT
) {
1873 /* Constant array index */
1874 const GLint element
= (GLint
) n
->Children
[1]->Value
[0];
1876 /* this element's storage is the array's storage, plus constant offset */
1877 n
->Store
->Index
+= elemSizeVec
* element
;
1880 /* Variable array index */
1882 /* do codegen for array index expression */
1883 emit(emitInfo
, n
->Children
[1]);
1884 indexStore
= n
->Children
[1]->Store
;
1886 if (indexStore
->IsIndirect
) {
1887 /* need to put the array index into a temporary since we can't
1888 * directly support a[b[i]] constructs.
1892 /*indexStore = tempstore();*/
1897 /* need to multiply array index by array element size */
1898 struct prog_instruction
*inst
;
1899 slang_ir_storage
*indexTemp
;
1900 slang_ir_storage elemSizeStore
;
1902 /* allocate 1 float indexTemp */
1903 indexTemp
= _slang_new_ir_storage(PROGRAM_TEMPORARY
, -1, 1);
1904 _slang_alloc_temp(emitInfo
->vt
, indexTemp
);
1906 /* allocate a constant containing the element size */
1907 constant_to_storage(emitInfo
, (float) elemSizeVec
, &elemSizeStore
);
1909 /* multiply array index by element size */
1910 inst
= emit_instruction(emitInfo
,
1912 indexTemp
, /* dest */
1913 indexStore
, /* the index */
1917 indexStore
= indexTemp
;
1920 if (arrayStore
->IsIndirect
) {
1921 /* ex: in a[i][j], a[i] (the arrayStore) is indirect */
1922 /* Need to add indexStore to arrayStore->Indirect store */
1923 slang_ir_storage indirectArray
;
1924 slang_ir_storage
*indexTemp
;
1926 _slang_init_ir_storage(&indirectArray
,
1927 arrayStore
->IndirectFile
,
1928 arrayStore
->IndirectIndex
,
1930 arrayStore
->IndirectSwizzle
);
1932 /* allocate 1 float indexTemp */
1933 indexTemp
= _slang_new_ir_storage(PROGRAM_TEMPORARY
, -1, 1);
1934 _slang_alloc_temp(emitInfo
->vt
, indexTemp
);
1936 inst
= emit_instruction(emitInfo
,
1938 indexTemp
, /* dest */
1939 indexStore
, /* the index */
1940 &indirectArray
, /* indirect array base */
1943 indexStore
= indexTemp
;
1946 /* update the array element storage info */
1947 n
->Store
->IsIndirect
= GL_TRUE
;
1948 n
->Store
->IndirectFile
= indexStore
->File
;
1949 n
->Store
->IndirectIndex
= indexStore
->Index
;
1950 n
->Store
->IndirectSwizzle
= indexStore
->Swizzle
;
1953 n
->Store
->Size
= elemSize
;
1954 n
->Store
->Swizzle
= _slang_var_swizzle(elemSize
, 0);
1956 return NULL
; /* no instruction */
1961 * Resolve storage for accessing a structure field.
1963 static struct prog_instruction
*
1964 emit_struct_field(slang_emit_info
*emitInfo
, slang_ir_node
*n
)
1966 slang_ir_storage
*root
= n
->Store
;
1967 GLint fieldOffset
, fieldSize
;
1969 assert(n
->Opcode
== IR_FIELD
);
1971 assert(!root
->Parent
);
1972 while (root
->Parent
)
1973 root
= root
->Parent
;
1975 /* If this is the field of a state var, allocate constant/uniform
1976 * storage for it now if we haven't already.
1977 * Note that we allocate storage (uniform/constant slots) for state
1978 * variables here rather than at declaration time so we only allocate
1979 * space for the ones that we actually use!
1981 if (root
->File
== PROGRAM_STATE_VAR
) {
1983 GLint index
= _slang_alloc_statevar(n
, emitInfo
->prog
->Parameters
, &direct
);
1985 slang_info_log_error(emitInfo
->log
, "Error parsing state variable");
1989 root
->Index
= index
;
1990 return NULL
; /* all done */
1994 /* do codegen for struct */
1995 emit(emitInfo
, n
->Children
[0]);
1996 assert(n
->Children
[0]->Store
->Index
>= 0);
1999 fieldOffset
= n
->Store
->Index
;
2000 fieldSize
= n
->Store
->Size
;
2002 _slang_copy_ir_storage(n
->Store
, n
->Children
[0]->Store
);
2004 n
->Store
->Index
= n
->Children
[0]->Store
->Index
+ fieldOffset
/ 4;
2005 n
->Store
->Size
= fieldSize
;
2007 switch (fieldSize
) {
2010 GLint swz
= fieldOffset
% 4;
2011 n
->Store
->Swizzle
= MAKE_SWIZZLE4(swz
, swz
, swz
, swz
);
2015 n
->Store
->Swizzle
= MAKE_SWIZZLE4(SWIZZLE_X
, SWIZZLE_Y
,
2016 SWIZZLE_NIL
, SWIZZLE_NIL
);
2019 n
->Store
->Swizzle
= MAKE_SWIZZLE4(SWIZZLE_X
, SWIZZLE_Y
,
2020 SWIZZLE_Z
, SWIZZLE_NIL
);
2023 n
->Store
->Swizzle
= SWIZZLE_XYZW
;
2026 assert(n
->Store
->Index
>= 0);
2028 return NULL
; /* no instruction */
2033 * Emit code for a variable declaration.
2034 * This usually doesn't result in any code generation, but just
2035 * memory allocation.
2037 static struct prog_instruction
*
2038 emit_var_decl(slang_emit_info
*emitInfo
, slang_ir_node
*n
)
2041 assert(n
->Store
->File
!= PROGRAM_UNDEFINED
);
2042 assert(n
->Store
->Size
> 0);
2043 /*assert(n->Store->Index < 0);*/
2045 if (!n
->Var
|| n
->Var
->isTemp
) {
2046 /* a nameless/temporary variable, will be freed after first use */
2048 if (n
->Store
->Index
< 0 && !_slang_alloc_temp(emitInfo
->vt
, n
->Store
)) {
2049 slang_info_log_error(emitInfo
->log
,
2050 "Ran out of registers, too many temporaries");
2055 /* a regular variable */
2056 _slang_add_variable(emitInfo
->vt
, n
->Var
);
2057 if (!_slang_alloc_var(emitInfo
->vt
, n
->Store
)) {
2058 slang_info_log_error(emitInfo
->log
,
2059 "Ran out of registers, too many variables");
2063 printf("IR_VAR_DECL %s %d store %p\n",
2064 (char*) n->Var->a_name, n->Store->Index, (void*) n->Store);
2066 assert(n
->Var
->store
== n
->Store
);
2068 if (emitInfo
->EmitComments
) {
2069 /* emit NOP with comment describing the variable's storage location */
2071 sprintf(s
, "TEMP[%d]%s = variable %s (size %d)",
2073 _mesa_swizzle_string(n
->Store
->Swizzle
, 0, GL_FALSE
),
2074 (n
->Var
? (char *) n
->Var
->a_name
: "anonymous"),
2076 emit_comment(emitInfo
, s
);
2083 * Emit code for a reference to a variable.
2084 * Actually, no code is generated but we may do some memory allocation.
2085 * In particular, state vars (uniforms) are allocated on an as-needed basis.
2087 static struct prog_instruction
*
2088 emit_var_ref(slang_emit_info
*emitInfo
, slang_ir_node
*n
)
2091 assert(n
->Store
->File
!= PROGRAM_UNDEFINED
);
2093 if (n
->Store
->File
== PROGRAM_STATE_VAR
&& n
->Store
->Index
< 0) {
2095 GLint index
= _slang_alloc_statevar(n
, emitInfo
->prog
->Parameters
, &direct
);
2099 _mesa_snprintf(s
, sizeof(s
), "Undefined variable '%s'",
2100 (char *) n
->Var
->a_name
);
2101 slang_info_log_error(emitInfo
->log
, s
);
2105 n
->Store
->Index
= index
;
2107 else if (n
->Store
->File
== PROGRAM_UNIFORM
||
2108 n
->Store
->File
== PROGRAM_SAMPLER
) {
2109 /* mark var as used */
2110 _mesa_use_uniform(emitInfo
->prog
->Parameters
, (char *) n
->Var
->a_name
);
2113 if (n
->Store
->Index
< 0) {
2114 /* probably ran out of registers */
2117 assert(n
->Store
->Size
> 0);
2123 static struct prog_instruction
*
2124 emit(slang_emit_info
*emitInfo
, slang_ir_node
*n
)
2126 struct prog_instruction
*inst
;
2130 if (emitInfo
->log
->error_flag
) {
2134 switch (n
->Opcode
) {
2136 /* sequence of two sub-trees */
2137 assert(n
->Children
[0]);
2138 assert(n
->Children
[1]);
2139 emit(emitInfo
, n
->Children
[0]);
2140 if (emitInfo
->log
->error_flag
)
2142 inst
= emit(emitInfo
, n
->Children
[1]);
2146 n
->Store
= n
->Children
[1]->Store
;
2150 /* new variable scope */
2151 _slang_push_var_table(emitInfo
->vt
);
2152 inst
= emit(emitInfo
, n
->Children
[0]);
2153 _slang_pop_var_table(emitInfo
->vt
);
2157 /* Variable declaration - allocate a register for it */
2158 inst
= emit_var_decl(emitInfo
, n
);
2162 /* Reference to a variable
2163 * Storage should have already been resolved/allocated.
2165 return emit_var_ref(emitInfo
, n
);
2168 return emit_array_element(emitInfo
, n
);
2170 return emit_struct_field(emitInfo
, n
);
2172 return emit_swizzle(emitInfo
, n
);
2174 /* Simple arithmetic */
2211 /* trinary operators */
2213 return emit_arith(emitInfo
, n
);
2217 return emit_compare(emitInfo
, n
);
2220 return emit_clamp(emitInfo
, n
);
2224 return emit_tex(emitInfo
, n
);
2226 return emit_negation(emitInfo
, n
);
2228 /* find storage location for this float constant */
2229 n
->Store
->Index
= _mesa_add_unnamed_constant(emitInfo
->prog
->Parameters
,
2232 &n
->Store
->Swizzle
);
2233 if (n
->Store
->Index
< 0) {
2234 slang_info_log_error(emitInfo
->log
, "Ran out of space for constants");
2240 return emit_copy(emitInfo
, n
);
2243 return emit_cond(emitInfo
, n
);
2246 return emit_not(emitInfo
, n
);
2249 return emit_label(emitInfo
, n
);
2252 return emit_kill(emitInfo
);
2255 /* new variable scope for subroutines/function calls */
2256 _slang_push_var_table(emitInfo
->vt
);
2257 inst
= emit_fcall(emitInfo
, n
);
2258 _slang_pop_var_table(emitInfo
->vt
);
2262 return emit_if(emitInfo
, n
);
2265 return emit_loop(emitInfo
, n
);
2266 case IR_BREAK_IF_TRUE
:
2267 case IR_CONT_IF_TRUE
:
2268 return emit_cont_break_if_true(emitInfo
, n
);
2272 return emit_cont_break(emitInfo
, n
);
2275 return new_instruction(emitInfo
, OPCODE_BGNSUB
);
2277 return new_instruction(emitInfo
, OPCODE_ENDSUB
);
2279 return emit_return(emitInfo
, n
);
2285 _mesa_problem(NULL
, "Unexpected IR opcode in emit()\n");
2292 * After code generation, any subroutines will be in separate program
2293 * objects. This function appends all the subroutines onto the main
2294 * program and resolves the linking of all the branch/call instructions.
2295 * XXX this logic should really be part of the linking process...
2298 _slang_resolve_subroutines(slang_emit_info
*emitInfo
)
2300 GET_CURRENT_CONTEXT(ctx
);
2301 struct gl_program
*mainP
= emitInfo
->prog
;
2302 GLuint
*subroutineLoc
, i
, total
;
2305 = (GLuint
*) _mesa_malloc(emitInfo
->NumSubroutines
* sizeof(GLuint
));
2307 /* total number of instructions */
2308 total
= mainP
->NumInstructions
;
2309 for (i
= 0; i
< emitInfo
->NumSubroutines
; i
++) {
2310 subroutineLoc
[i
] = total
;
2311 total
+= emitInfo
->Subroutines
[i
]->NumInstructions
;
2314 /* adjust BranchTargets within the functions */
2315 for (i
= 0; i
< emitInfo
->NumSubroutines
; i
++) {
2316 struct gl_program
*sub
= emitInfo
->Subroutines
[i
];
2318 for (j
= 0; j
< sub
->NumInstructions
; j
++) {
2319 struct prog_instruction
*inst
= sub
->Instructions
+ j
;
2320 if (inst
->Opcode
!= OPCODE_CAL
&& inst
->BranchTarget
>= 0) {
2321 inst
->BranchTarget
+= subroutineLoc
[i
];
2326 /* append subroutines' instructions after main's instructions */
2327 mainP
->Instructions
= _mesa_realloc_instructions(mainP
->Instructions
,
2328 mainP
->NumInstructions
,
2330 mainP
->NumInstructions
= total
;
2331 for (i
= 0; i
< emitInfo
->NumSubroutines
; i
++) {
2332 struct gl_program
*sub
= emitInfo
->Subroutines
[i
];
2333 _mesa_copy_instructions(mainP
->Instructions
+ subroutineLoc
[i
],
2335 sub
->NumInstructions
);
2336 /* delete subroutine code */
2337 sub
->Parameters
= NULL
; /* prevent double-free */
2338 _mesa_reference_program(ctx
, &emitInfo
->Subroutines
[i
], NULL
);
2341 /* free subroutine list */
2342 if (emitInfo
->Subroutines
) {
2343 _mesa_free(emitInfo
->Subroutines
);
2344 emitInfo
->Subroutines
= NULL
;
2346 emitInfo
->NumSubroutines
= 0;
2348 /* Examine CAL instructions.
2349 * At this point, the BranchTarget field of the CAL instruction is
2350 * the number/id of the subroutine to call (an index into the
2351 * emitInfo->Subroutines list).
2352 * Translate that into an actual instruction location now.
2354 for (i
= 0; i
< mainP
->NumInstructions
; i
++) {
2355 struct prog_instruction
*inst
= mainP
->Instructions
+ i
;
2356 if (inst
->Opcode
== OPCODE_CAL
) {
2357 const GLuint f
= inst
->BranchTarget
;
2358 inst
->BranchTarget
= subroutineLoc
[f
];
2362 _mesa_free(subroutineLoc
);
2369 _slang_emit_code(slang_ir_node
*n
, slang_var_table
*vt
,
2370 struct gl_program
*prog
, GLboolean withEnd
,
2371 slang_info_log
*log
)
2373 GET_CURRENT_CONTEXT(ctx
);
2375 slang_emit_info emitInfo
;
2380 emitInfo
.prog
= prog
;
2381 emitInfo
.Subroutines
= NULL
;
2382 emitInfo
.NumSubroutines
= 0;
2383 emitInfo
.MaxInstructions
= prog
->NumInstructions
;
2385 emitInfo
.EmitHighLevelInstructions
= ctx
->Shader
.EmitHighLevelInstructions
;
2386 emitInfo
.EmitCondCodes
= ctx
->Shader
.EmitCondCodes
;
2387 emitInfo
.EmitComments
= ctx
->Shader
.EmitComments
;
2388 emitInfo
.EmitBeginEndSub
= GL_TRUE
;
2390 if (!emitInfo
.EmitCondCodes
) {
2391 emitInfo
.EmitHighLevelInstructions
= GL_TRUE
;
2394 /* Check uniform/constant limits */
2395 if (prog
->Target
== GL_FRAGMENT_PROGRAM_ARB
) {
2396 maxUniforms
= ctx
->Const
.FragmentProgram
.MaxUniformComponents
/ 4;
2399 assert(prog
->Target
== GL_VERTEX_PROGRAM_ARB
);
2400 maxUniforms
= ctx
->Const
.VertexProgram
.MaxUniformComponents
/ 4;
2402 if (prog
->Parameters
->NumParameters
> maxUniforms
) {
2403 slang_info_log_error(log
, "Constant/uniform register limit exceeded");
2407 (void) emit(&emitInfo
, n
);
2409 /* finish up by adding the END opcode to program */
2411 struct prog_instruction
*inst
;
2412 inst
= new_instruction(&emitInfo
, OPCODE_END
);
2415 _slang_resolve_subroutines(&emitInfo
);
2420 printf("*********** End emit code (%u inst):\n", prog
->NumInstructions
);
2421 _mesa_print_program(prog
);
2422 _mesa_print_program_parameters(ctx
,prog
);