2 * Mesa 3-D graphics library
5 * Copyright (C) 2009 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 * VMWARE 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 #include "main/glheader.h"
28 #include "main/context.h"
29 #include "main/macros.h"
31 #include "prog_instruction.h"
32 #include "prog_optimize.h"
33 #include "prog_print.h"
36 #define MAX_LOOP_NESTING 50
37 /* MAX_PROGRAM_TEMPS is a low number (256), and we want to be able to
38 * register allocate many temporary values into that small number of
39 * temps. So allow large temporary indices coming into the register
42 #define REG_ALLOCATE_MAX_PROGRAM_TEMPS ((1 << INST_INDEX_BITS) - 1)
44 static GLboolean dbg
= GL_FALSE
;
49 * Returns the mask of channels (bitmask of WRITEMASK_X,Y,Z,W) which
50 * are read from the given src in this instruction, We also provide
51 * one optional masks which may mask other components in the dst
55 get_src_arg_mask(const struct prog_instruction
*inst
,
56 GLuint arg
, GLuint dst_mask
)
58 GLuint read_mask
, channel_mask
;
61 ASSERT(arg
< _mesa_num_inst_src_regs(inst
->Opcode
));
63 /* Form the dst register, find the written channels */
64 if (inst
->CondUpdate
) {
65 channel_mask
= WRITEMASK_XYZW
;
68 switch (inst
->Opcode
) {
77 channel_mask
= inst
->DstReg
.WriteMask
& dst_mask
;
86 channel_mask
= WRITEMASK_X
;
89 channel_mask
= WRITEMASK_XY
;
93 channel_mask
= WRITEMASK_XYZ
;
96 channel_mask
= WRITEMASK_XYZW
;
101 /* Now, given the src swizzle and the written channels, find which
102 * components are actually read
105 for (comp
= 0; comp
< 4; ++comp
) {
106 const GLuint coord
= GET_SWZ(inst
->SrcReg
[arg
].Swizzle
, comp
);
108 if (channel_mask
& (1 << comp
) && coord
<= SWIZZLE_W
)
109 read_mask
|= 1 << coord
;
117 * For a MOV instruction, compute a write mask when src register also has
121 get_dst_mask_for_mov(const struct prog_instruction
*mov
, GLuint src_mask
)
123 const GLuint mask
= mov
->DstReg
.WriteMask
;
125 GLuint updated_mask
= 0x0;
127 ASSERT(mov
->Opcode
== OPCODE_MOV
);
129 for (comp
= 0; comp
< 4; ++comp
) {
131 if ((mask
& (1 << comp
)) == 0)
133 src_comp
= GET_SWZ(mov
->SrcReg
[0].Swizzle
, comp
);
134 if ((src_mask
& (1 << src_comp
)) == 0)
136 updated_mask
|= 1 << comp
;
144 * Ensure that the swizzle is regular. That is, all of the swizzle
145 * terms are SWIZZLE_X,Y,Z,W and not SWIZZLE_ZERO or SWIZZLE_ONE.
148 is_swizzle_regular(GLuint swz
)
150 return GET_SWZ(swz
,0) <= SWIZZLE_W
&&
151 GET_SWZ(swz
,1) <= SWIZZLE_W
&&
152 GET_SWZ(swz
,2) <= SWIZZLE_W
&&
153 GET_SWZ(swz
,3) <= SWIZZLE_W
;
158 * In 'prog' remove instruction[i] if removeFlags[i] == TRUE.
159 * \return number of instructions removed
162 remove_instructions(struct gl_program
*prog
, const GLboolean
*removeFlags
)
164 GLint i
, removeEnd
= 0, removeCount
= 0;
165 GLuint totalRemoved
= 0;
168 for (i
= prog
->NumInstructions
- 1; i
>= 0; i
--) {
169 if (removeFlags
[i
]) {
171 if (removeCount
== 0) {
172 /* begin a run of instructions to remove */
177 /* extend the run of instructions to remove */
182 /* don't remove this instruction, but check if the preceeding
183 * instructions are to be removed.
185 if (removeCount
> 0) {
186 GLint removeStart
= removeEnd
- removeCount
+ 1;
187 _mesa_delete_instructions(prog
, removeStart
, removeCount
);
188 removeStart
= removeCount
= 0; /* reset removal info */
192 /* Finish removing if the first instruction was to be removed. */
193 if (removeCount
> 0) {
194 GLint removeStart
= removeEnd
- removeCount
+ 1;
195 _mesa_delete_instructions(prog
, removeStart
, removeCount
);
202 * Remap register indexes according to map.
203 * \param prog the program to search/replace
204 * \param file the type of register file to search/replace
205 * \param map maps old register indexes to new indexes
208 replace_regs(struct gl_program
*prog
, gl_register_file file
, const GLint map
[])
212 for (i
= 0; i
< prog
->NumInstructions
; i
++) {
213 struct prog_instruction
*inst
= prog
->Instructions
+ i
;
214 const GLuint numSrc
= _mesa_num_inst_src_regs(inst
->Opcode
);
216 for (j
= 0; j
< numSrc
; j
++) {
217 if (inst
->SrcReg
[j
].File
== file
) {
218 GLuint index
= inst
->SrcReg
[j
].Index
;
219 ASSERT(map
[index
] >= 0);
220 inst
->SrcReg
[j
].Index
= map
[index
];
223 if (inst
->DstReg
.File
== file
) {
224 const GLuint index
= inst
->DstReg
.Index
;
225 ASSERT(map
[index
] >= 0);
226 inst
->DstReg
.Index
= map
[index
];
233 * Remove dead instructions from the given program.
234 * This is very primitive for now. Basically look for temp registers
235 * that are written to but never read. Remove any instructions that
236 * write to such registers. Be careful with condition code setters.
239 _mesa_remove_dead_code_global(struct gl_program
*prog
)
241 GLboolean tempRead
[REG_ALLOCATE_MAX_PROGRAM_TEMPS
][4];
242 GLboolean
*removeInst
; /* per-instruction removal flag */
243 GLuint i
, rem
= 0, comp
;
245 memset(tempRead
, 0, sizeof(tempRead
));
248 printf("Optimize: Begin dead code removal\n");
249 /*_mesa_print_program(prog);*/
252 removeInst
= (GLboolean
*)
253 calloc(1, prog
->NumInstructions
* sizeof(GLboolean
));
255 /* Determine which temps are read and written */
256 for (i
= 0; i
< prog
->NumInstructions
; i
++) {
257 const struct prog_instruction
*inst
= prog
->Instructions
+ i
;
258 const GLuint numSrc
= _mesa_num_inst_src_regs(inst
->Opcode
);
262 for (j
= 0; j
< numSrc
; j
++) {
263 if (inst
->SrcReg
[j
].File
== PROGRAM_TEMPORARY
) {
264 const GLuint index
= inst
->SrcReg
[j
].Index
;
266 ASSERT(index
< REG_ALLOCATE_MAX_PROGRAM_TEMPS
);
267 read_mask
= get_src_arg_mask(inst
, j
, NO_MASK
);
269 if (inst
->SrcReg
[j
].RelAddr
) {
271 printf("abort remove dead code (indirect temp)\n");
275 for (comp
= 0; comp
< 4; comp
++) {
276 const GLuint swz
= GET_SWZ(inst
->SrcReg
[j
].Swizzle
, comp
);
278 if ((read_mask
& (1 << swz
)) == 0)
280 if (swz
<= SWIZZLE_W
)
281 tempRead
[index
][swz
] = GL_TRUE
;
287 if (inst
->DstReg
.File
== PROGRAM_TEMPORARY
) {
288 const GLuint index
= inst
->DstReg
.Index
;
289 ASSERT(index
< REG_ALLOCATE_MAX_PROGRAM_TEMPS
);
291 if (inst
->DstReg
.RelAddr
) {
293 printf("abort remove dead code (indirect temp)\n");
297 if (inst
->CondUpdate
) {
298 /* If we're writing to this register and setting condition
299 * codes we cannot remove the instruction. Prevent removal
300 * by setting the 'read' flag.
302 tempRead
[index
][0] = GL_TRUE
;
303 tempRead
[index
][1] = GL_TRUE
;
304 tempRead
[index
][2] = GL_TRUE
;
305 tempRead
[index
][3] = GL_TRUE
;
310 /* find instructions that write to dead registers, flag for removal */
311 for (i
= 0; i
< prog
->NumInstructions
; i
++) {
312 struct prog_instruction
*inst
= prog
->Instructions
+ i
;
313 const GLuint numDst
= _mesa_num_inst_dst_regs(inst
->Opcode
);
315 if (numDst
!= 0 && inst
->DstReg
.File
== PROGRAM_TEMPORARY
) {
316 GLint chan
, index
= inst
->DstReg
.Index
;
318 for (chan
= 0; chan
< 4; chan
++) {
319 if (!tempRead
[index
][chan
] &&
320 inst
->DstReg
.WriteMask
& (1 << chan
)) {
322 printf("Remove writemask on %u.%c\n", i
,
323 chan
== 3 ? 'w' : 'x' + chan
);
325 inst
->DstReg
.WriteMask
&= ~(1 << chan
);
330 if (inst
->DstReg
.WriteMask
== 0) {
331 /* If we cleared all writes, the instruction can be removed. */
333 printf("Remove instruction %u: \n", i
);
334 removeInst
[i
] = GL_TRUE
;
339 /* now remove the instructions which aren't needed */
340 rem
= remove_instructions(prog
, removeInst
);
343 printf("Optimize: End dead code removal.\n");
344 printf(" %u channel writes removed\n", rem
);
345 printf(" %u instructions removed\n", rem
);
346 /*_mesa_print_program(prog);*/
365 * Scan forward in program from 'start' for the next occurances of TEMP[index].
366 * We look if an instruction reads the component given by the masks and if they
368 * Return READ, WRITE, FLOW or END to indicate the next usage or an indicator
369 * that we can't look further.
372 find_next_use(const struct gl_program
*prog
,
379 for (i
= start
; i
< prog
->NumInstructions
; i
++) {
380 const struct prog_instruction
*inst
= prog
->Instructions
+ i
;
381 switch (inst
->Opcode
) {
398 const GLuint numSrc
= _mesa_num_inst_src_regs(inst
->Opcode
);
400 for (j
= 0; j
< numSrc
; j
++) {
401 if (inst
->SrcReg
[j
].RelAddr
||
402 (inst
->SrcReg
[j
].File
== PROGRAM_TEMPORARY
&&
403 inst
->SrcReg
[j
].Index
== index
&&
404 (get_src_arg_mask(inst
,j
,NO_MASK
) & mask
)))
407 if (_mesa_num_inst_dst_regs(inst
->Opcode
) == 1 &&
408 inst
->DstReg
.File
== PROGRAM_TEMPORARY
&&
409 inst
->DstReg
.Index
== index
) {
410 mask
&= ~inst
->DstReg
.WriteMask
;
422 * Is the given instruction opcode a flow-control opcode?
423 * XXX maybe move this into prog_instruction.[ch]
426 _mesa_is_flow_control_opcode(enum prog_opcode opcode
)
449 * Test if the given instruction is a simple MOV (no conditional updating,
450 * not relative addressing, no negation/abs, etc).
453 can_downward_mov_be_modifed(const struct prog_instruction
*mov
)
456 mov
->Opcode
== OPCODE_MOV
&&
457 mov
->CondUpdate
== GL_FALSE
&&
458 mov
->SrcReg
[0].RelAddr
== 0 &&
459 mov
->SrcReg
[0].Negate
== 0 &&
460 mov
->SrcReg
[0].Abs
== 0 &&
461 mov
->SrcReg
[0].HasIndex2
== 0 &&
462 mov
->SrcReg
[0].RelAddr2
== 0 &&
463 mov
->DstReg
.RelAddr
== 0 &&
464 mov
->DstReg
.CondMask
== COND_TR
&&
465 mov
->SaturateMode
== SATURATE_OFF
;
470 can_upward_mov_be_modifed(const struct prog_instruction
*mov
)
473 can_downward_mov_be_modifed(mov
) &&
474 mov
->DstReg
.File
== PROGRAM_TEMPORARY
;
479 * Try to remove use of extraneous MOV instructions, to free them up for dead
483 _mesa_remove_extra_move_use(struct gl_program
*prog
)
488 printf("Optimize: Begin remove extra move use\n");
489 _mesa_print_program(prog
);
493 * Look for sequences such as this:
496 * FOO tmpY, tmpX, arg1;
500 * FOO tmpY, arg0, arg1;
503 for (i
= 0; i
+ 1 < prog
->NumInstructions
; i
++) {
504 const struct prog_instruction
*mov
= prog
->Instructions
+ i
;
505 GLuint dst_mask
, src_mask
;
506 if (can_upward_mov_be_modifed(mov
) == GL_FALSE
)
509 /* Scanning the code, we maintain the components which are still active in
512 dst_mask
= mov
->DstReg
.WriteMask
;
513 src_mask
= get_src_arg_mask(mov
, 0, NO_MASK
);
515 /* Walk through remaining instructions until the or src reg gets
516 * rewritten or we get into some flow-control, eliminating the use of
519 for (j
= i
+ 1; j
< prog
->NumInstructions
; j
++) {
520 struct prog_instruction
*inst2
= prog
->Instructions
+ j
;
523 if (_mesa_is_flow_control_opcode(inst2
->Opcode
))
526 /* First rewrite this instruction's args if appropriate. */
527 for (arg
= 0; arg
< _mesa_num_inst_src_regs(inst2
->Opcode
); arg
++) {
528 GLuint comp
, read_mask
;
530 if (inst2
->SrcReg
[arg
].File
!= mov
->DstReg
.File
||
531 inst2
->SrcReg
[arg
].Index
!= mov
->DstReg
.Index
||
532 inst2
->SrcReg
[arg
].RelAddr
||
533 inst2
->SrcReg
[arg
].Abs
)
535 read_mask
= get_src_arg_mask(inst2
, arg
, NO_MASK
);
537 /* Adjust the swizzles of inst2 to point at MOV's source if ALL the
538 * components read still come from the mov instructions
540 if (is_swizzle_regular(inst2
->SrcReg
[arg
].Swizzle
) &&
541 (read_mask
& dst_mask
) == read_mask
) {
542 for (comp
= 0; comp
< 4; comp
++) {
543 const GLuint inst2_swz
=
544 GET_SWZ(inst2
->SrcReg
[arg
].Swizzle
, comp
);
545 const GLuint s
= GET_SWZ(mov
->SrcReg
[0].Swizzle
, inst2_swz
);
546 inst2
->SrcReg
[arg
].Swizzle
&= ~(7 << (3 * comp
));
547 inst2
->SrcReg
[arg
].Swizzle
|= s
<< (3 * comp
);
548 inst2
->SrcReg
[arg
].Negate
^= (((mov
->SrcReg
[0].Negate
>>
549 inst2_swz
) & 0x1) << comp
);
551 inst2
->SrcReg
[arg
].File
= mov
->SrcReg
[0].File
;
552 inst2
->SrcReg
[arg
].Index
= mov
->SrcReg
[0].Index
;
556 /* The source of MOV is written. This potentially deactivates some
557 * components from the src and dst of the MOV instruction
559 if (inst2
->DstReg
.File
== mov
->DstReg
.File
&&
560 (inst2
->DstReg
.RelAddr
||
561 inst2
->DstReg
.Index
== mov
->DstReg
.Index
)) {
562 dst_mask
&= ~inst2
->DstReg
.WriteMask
;
563 src_mask
= get_src_arg_mask(mov
, 0, dst_mask
);
566 /* Idem when the destination of mov is written */
567 if (inst2
->DstReg
.File
== mov
->SrcReg
[0].File
&&
568 (inst2
->DstReg
.RelAddr
||
569 inst2
->DstReg
.Index
== mov
->SrcReg
[0].Index
)) {
570 src_mask
&= ~inst2
->DstReg
.WriteMask
;
571 dst_mask
&= get_dst_mask_for_mov(mov
, src_mask
);
579 printf("Optimize: End remove extra move use.\n");
580 /*_mesa_print_program(prog);*/
586 * Complements dead_code_global. Try to remove code in block of code by
587 * carefully monitoring the swizzles. Both functions should be merged into one
588 * with a proper control flow graph
591 _mesa_remove_dead_code_local(struct gl_program
*prog
)
593 GLboolean
*removeInst
;
594 GLuint i
, arg
, rem
= 0;
596 removeInst
= (GLboolean
*)
597 calloc(1, prog
->NumInstructions
* sizeof(GLboolean
));
599 for (i
= 0; i
< prog
->NumInstructions
; i
++) {
600 const struct prog_instruction
*inst
= prog
->Instructions
+ i
;
601 const GLuint index
= inst
->DstReg
.Index
;
602 const GLuint mask
= inst
->DstReg
.WriteMask
;
605 /* We must deactivate the pass as soon as some indirection is used */
606 if (inst
->DstReg
.RelAddr
)
608 for (arg
= 0; arg
< _mesa_num_inst_src_regs(inst
->Opcode
); arg
++)
609 if (inst
->SrcReg
[arg
].RelAddr
)
612 if (_mesa_is_flow_control_opcode(inst
->Opcode
) ||
613 _mesa_num_inst_dst_regs(inst
->Opcode
) == 0 ||
614 inst
->DstReg
.File
!= PROGRAM_TEMPORARY
||
615 inst
->DstReg
.RelAddr
)
618 use
= find_next_use(prog
, i
+1, index
, mask
);
619 if (use
== WRITE
|| use
== END
)
620 removeInst
[i
] = GL_TRUE
;
623 rem
= remove_instructions(prog
, removeInst
);
632 * Try to inject the destination of mov as the destination of inst and recompute
633 * the swizzles operators for the sources of inst if required. Return GL_TRUE
634 * of the substitution was possible, GL_FALSE otherwise
637 _mesa_merge_mov_into_inst(struct prog_instruction
*inst
,
638 const struct prog_instruction
*mov
)
640 /* Indirection table which associates destination and source components for
641 * the mov instruction
643 const GLuint mask
= get_src_arg_mask(mov
, 0, NO_MASK
);
645 /* Some components are not written by inst. We cannot remove the mov */
646 if (mask
!= (inst
->DstReg
.WriteMask
& mask
))
649 /* Depending on the instruction, we may need to recompute the swizzles.
650 * Also, some other instructions (like TEX) are not linear. We will only
651 * consider completely active sources and destinations
653 switch (inst
->Opcode
) {
655 /* Carstesian instructions: we compute the swizzle */
665 GLuint dst_to_src_comp
[4] = {0,0,0,0};
666 GLuint dst_comp
, arg
;
667 for (dst_comp
= 0; dst_comp
< 4; ++dst_comp
) {
668 if (mov
->DstReg
.WriteMask
& (1 << dst_comp
)) {
669 const GLuint src_comp
= GET_SWZ(mov
->SrcReg
[0].Swizzle
, dst_comp
);
670 ASSERT(src_comp
< 4);
671 dst_to_src_comp
[dst_comp
] = src_comp
;
675 /* Patch each source of the instruction */
676 for (arg
= 0; arg
< _mesa_num_inst_src_regs(inst
->Opcode
); arg
++) {
677 const GLuint arg_swz
= inst
->SrcReg
[arg
].Swizzle
;
678 inst
->SrcReg
[arg
].Swizzle
= 0;
680 /* Reset each active component of the swizzle */
681 for (dst_comp
= 0; dst_comp
< 4; ++dst_comp
) {
682 GLuint src_comp
, arg_comp
;
683 if ((mov
->DstReg
.WriteMask
& (1 << dst_comp
)) == 0)
685 src_comp
= dst_to_src_comp
[dst_comp
];
686 ASSERT(src_comp
< 4);
687 arg_comp
= GET_SWZ(arg_swz
, src_comp
);
688 ASSERT(arg_comp
< 4);
689 inst
->SrcReg
[arg
].Swizzle
|= arg_comp
<< (3*dst_comp
);
692 inst
->DstReg
= mov
->DstReg
;
696 /* Dot products and scalar instructions: we only change the destination */
707 inst
->DstReg
= mov
->DstReg
;
710 /* All other instructions require fully active components with no swizzle */
712 if (mov
->SrcReg
[0].Swizzle
!= SWIZZLE_XYZW
||
713 inst
->DstReg
.WriteMask
!= WRITEMASK_XYZW
)
715 inst
->DstReg
= mov
->DstReg
;
722 * Try to remove extraneous MOV instructions from the given program.
725 _mesa_remove_extra_moves(struct gl_program
*prog
)
727 GLboolean
*removeInst
; /* per-instruction removal flag */
728 GLuint i
, rem
= 0, nesting
= 0;
731 printf("Optimize: Begin remove extra moves\n");
732 _mesa_print_program(prog
);
735 removeInst
= (GLboolean
*)
736 calloc(1, prog
->NumInstructions
* sizeof(GLboolean
));
739 * Look for sequences such as this:
740 * FOO tmpX, arg0, arg1;
743 * FOO tmpY, arg0, arg1;
746 for (i
= 0; i
< prog
->NumInstructions
; i
++) {
747 const struct prog_instruction
*mov
= prog
->Instructions
+ i
;
749 switch (mov
->Opcode
) {
761 if (i
> 0 && can_downward_mov_be_modifed(mov
) && nesting
== 0) {
763 /* see if this MOV can be removed */
764 const GLuint id
= mov
->SrcReg
[0].Index
;
765 struct prog_instruction
*prevInst
;
768 /* get pointer to previous instruction */
770 while (prevI
> 0 && removeInst
[prevI
])
772 prevInst
= prog
->Instructions
+ prevI
;
774 if (prevInst
->DstReg
.File
== PROGRAM_TEMPORARY
&&
775 prevInst
->DstReg
.Index
== id
&&
776 prevInst
->DstReg
.RelAddr
== 0 &&
777 prevInst
->DstReg
.CondSrc
== 0 &&
778 prevInst
->DstReg
.CondMask
== COND_TR
) {
780 const GLuint dst_mask
= prevInst
->DstReg
.WriteMask
;
781 enum inst_use next_use
= find_next_use(prog
, i
+1, id
, dst_mask
);
783 if (next_use
== WRITE
|| next_use
== END
) {
784 /* OK, we can safely remove this MOV instruction.
786 * prevI: FOO tempIndex, x, y;
787 * i: MOV z, tempIndex;
789 * prevI: FOO z, x, y;
791 if (_mesa_merge_mov_into_inst(prevInst
, mov
)) {
792 removeInst
[i
] = GL_TRUE
;
794 printf("Remove MOV at %u\n", i
);
795 printf("new prev inst %u: ", prevI
);
796 _mesa_print_instruction(prevInst
);
808 /* now remove the instructions which aren't needed */
809 rem
= remove_instructions(prog
, removeInst
);
814 printf("Optimize: End remove extra moves. %u instructions removed\n", rem
);
815 /*_mesa_print_program(prog);*/
822 /** A live register interval */
825 GLuint Reg
; /** The temporary register index */
826 GLuint Start
, End
; /** Start/end instruction numbers */
830 /** A list of register intervals */
834 struct interval Intervals
[REG_ALLOCATE_MAX_PROGRAM_TEMPS
];
839 append_interval(struct interval_list
*list
, const struct interval
*inv
)
841 list
->Intervals
[list
->Num
++] = *inv
;
845 /** Insert interval inv into list, sorted by interval end */
847 insert_interval_by_end(struct interval_list
*list
, const struct interval
*inv
)
849 /* XXX we could do a binary search insertion here since list is sorted */
850 GLint i
= list
->Num
- 1;
851 while (i
>= 0 && list
->Intervals
[i
].End
> inv
->End
) {
852 list
->Intervals
[i
+ 1] = list
->Intervals
[i
];
855 list
->Intervals
[i
+ 1] = *inv
;
861 for (i
= 0; i
+ 1 < list
->Num
; i
++) {
862 ASSERT(list
->Intervals
[i
].End
<= list
->Intervals
[i
+ 1].End
);
869 /** Remove the given interval from the interval list */
871 remove_interval(struct interval_list
*list
, const struct interval
*inv
)
873 /* XXX we could binary search since list is sorted */
875 for (k
= 0; k
< list
->Num
; k
++) {
876 if (list
->Intervals
[k
].Reg
== inv
->Reg
) {
877 /* found, remove it */
878 ASSERT(list
->Intervals
[k
].Start
== inv
->Start
);
879 ASSERT(list
->Intervals
[k
].End
== inv
->End
);
880 while (k
< list
->Num
- 1) {
881 list
->Intervals
[k
] = list
->Intervals
[k
+ 1];
891 /** called by qsort() */
893 compare_start(const void *a
, const void *b
)
895 const struct interval
*ia
= (const struct interval
*) a
;
896 const struct interval
*ib
= (const struct interval
*) b
;
897 if (ia
->Start
< ib
->Start
)
899 else if (ia
->Start
> ib
->Start
)
906 /** sort the interval list according to interval starts */
908 sort_interval_list_by_start(struct interval_list
*list
)
910 qsort(list
->Intervals
, list
->Num
, sizeof(struct interval
), compare_start
);
914 for (i
= 0; i
+ 1 < list
->Num
; i
++) {
915 ASSERT(list
->Intervals
[i
].Start
<= list
->Intervals
[i
+ 1].Start
);
923 GLuint Start
, End
; /**< Start, end instructions of loop */
927 * Update the intermediate interval info for register 'index' and
931 update_interval(GLint intBegin
[], GLint intEnd
[],
932 struct loop_info
*loopStack
, GLuint loopStackDepth
,
933 GLuint index
, GLuint ic
)
937 /* If the register is used in a loop, extend its lifetime through the end
938 * of the outermost loop that doesn't contain its definition.
940 for (i
= 0; i
< loopStackDepth
; i
++) {
941 if (intBegin
[index
] < loopStack
[i
].Start
) {
942 ic
= loopStack
[i
].End
;
947 ASSERT(index
< REG_ALLOCATE_MAX_PROGRAM_TEMPS
);
948 if (intBegin
[index
] == -1) {
949 ASSERT(intEnd
[index
] == -1);
950 intBegin
[index
] = intEnd
[index
] = ic
;
959 * Find first/last instruction that references each temporary register.
962 _mesa_find_temp_intervals(const struct prog_instruction
*instructions
,
963 GLuint numInstructions
,
964 GLint intBegin
[REG_ALLOCATE_MAX_PROGRAM_TEMPS
],
965 GLint intEnd
[REG_ALLOCATE_MAX_PROGRAM_TEMPS
])
967 struct loop_info loopStack
[MAX_LOOP_NESTING
];
968 GLuint loopStackDepth
= 0;
971 for (i
= 0; i
< REG_ALLOCATE_MAX_PROGRAM_TEMPS
; i
++){
972 intBegin
[i
] = intEnd
[i
] = -1;
975 /* Scan instructions looking for temporary registers */
976 for (i
= 0; i
< numInstructions
; i
++) {
977 const struct prog_instruction
*inst
= instructions
+ i
;
978 if (inst
->Opcode
== OPCODE_BGNLOOP
) {
979 loopStack
[loopStackDepth
].Start
= i
;
980 loopStack
[loopStackDepth
].End
= inst
->BranchTarget
;
983 else if (inst
->Opcode
== OPCODE_ENDLOOP
) {
986 else if (inst
->Opcode
== OPCODE_CAL
) {
990 const GLuint numSrc
= 3;/*_mesa_num_inst_src_regs(inst->Opcode);*/
992 for (j
= 0; j
< numSrc
; j
++) {
993 if (inst
->SrcReg
[j
].File
== PROGRAM_TEMPORARY
) {
994 const GLuint index
= inst
->SrcReg
[j
].Index
;
995 if (inst
->SrcReg
[j
].RelAddr
)
997 update_interval(intBegin
, intEnd
, loopStack
, loopStackDepth
,
1001 if (inst
->DstReg
.File
== PROGRAM_TEMPORARY
) {
1002 const GLuint index
= inst
->DstReg
.Index
;
1003 if (inst
->DstReg
.RelAddr
)
1005 update_interval(intBegin
, intEnd
, loopStack
, loopStackDepth
,
1016 * Find the live intervals for each temporary register in the program.
1017 * For register R, the interval [A,B] indicates that R is referenced
1018 * from instruction A through instruction B.
1019 * Special consideration is needed for loops and subroutines.
1020 * \return GL_TRUE if success, GL_FALSE if we cannot proceed for some reason
1023 find_live_intervals(struct gl_program
*prog
,
1024 struct interval_list
*liveIntervals
)
1026 GLint intBegin
[REG_ALLOCATE_MAX_PROGRAM_TEMPS
];
1027 GLint intEnd
[REG_ALLOCATE_MAX_PROGRAM_TEMPS
];
1031 * Note: we'll return GL_FALSE below if we find relative indexing
1032 * into the TEMP register file. We can't handle that yet.
1033 * We also give up on subroutines for now.
1037 printf("Optimize: Begin find intervals\n");
1040 /* build intermediate arrays */
1041 if (!_mesa_find_temp_intervals(prog
->Instructions
, prog
->NumInstructions
,
1045 /* Build live intervals list from intermediate arrays */
1046 liveIntervals
->Num
= 0;
1047 for (i
= 0; i
< REG_ALLOCATE_MAX_PROGRAM_TEMPS
; i
++) {
1048 if (intBegin
[i
] >= 0) {
1049 struct interval inv
;
1051 inv
.Start
= intBegin
[i
];
1052 inv
.End
= intEnd
[i
];
1053 append_interval(liveIntervals
, &inv
);
1057 /* Sort the list according to interval starts */
1058 sort_interval_list_by_start(liveIntervals
);
1061 /* print interval info */
1062 for (i
= 0; i
< liveIntervals
->Num
; i
++) {
1063 const struct interval
*inv
= liveIntervals
->Intervals
+ i
;
1064 printf("Reg[%d] live [%d, %d]:",
1065 inv
->Reg
, inv
->Start
, inv
->End
);
1068 for (j
= 0; j
< inv
->Start
; j
++)
1070 for (j
= inv
->Start
; j
<= inv
->End
; j
++)
1081 /** Scan the array of used register flags to find free entry */
1083 alloc_register(GLboolean usedRegs
[REG_ALLOCATE_MAX_PROGRAM_TEMPS
])
1086 for (k
= 0; k
< REG_ALLOCATE_MAX_PROGRAM_TEMPS
; k
++) {
1088 usedRegs
[k
] = GL_TRUE
;
1097 * This function implements "Linear Scan Register Allocation" to reduce
1098 * the number of temporary registers used by the program.
1100 * We compute the "live interval" for all temporary registers then
1101 * examine the overlap of the intervals to allocate new registers.
1102 * Basically, if two intervals do not overlap, they can use the same register.
1105 _mesa_reallocate_registers(struct gl_program
*prog
)
1107 struct interval_list liveIntervals
;
1108 GLint registerMap
[REG_ALLOCATE_MAX_PROGRAM_TEMPS
];
1109 GLboolean usedRegs
[REG_ALLOCATE_MAX_PROGRAM_TEMPS
];
1114 printf("Optimize: Begin live-interval register reallocation\n");
1115 _mesa_print_program(prog
);
1118 for (i
= 0; i
< REG_ALLOCATE_MAX_PROGRAM_TEMPS
; i
++){
1119 registerMap
[i
] = -1;
1120 usedRegs
[i
] = GL_FALSE
;
1123 if (!find_live_intervals(prog
, &liveIntervals
)) {
1125 printf("Aborting register reallocation\n");
1130 struct interval_list activeIntervals
;
1131 activeIntervals
.Num
= 0;
1133 /* loop over live intervals, allocating a new register for each */
1134 for (i
= 0; i
< liveIntervals
.Num
; i
++) {
1135 const struct interval
*live
= liveIntervals
.Intervals
+ i
;
1138 printf("Consider register %u\n", live
->Reg
);
1140 /* Expire old intervals. Intervals which have ended with respect
1141 * to the live interval can have their remapped registers freed.
1145 for (j
= 0; j
< (GLint
) activeIntervals
.Num
; j
++) {
1146 const struct interval
*inv
= activeIntervals
.Intervals
+ j
;
1147 if (inv
->End
>= live
->Start
) {
1148 /* Stop now. Since the activeInterval list is sorted
1149 * we know we don't have to go further.
1154 /* Interval 'inv' has expired */
1155 const GLint regNew
= registerMap
[inv
->Reg
];
1156 ASSERT(regNew
>= 0);
1159 printf(" expire interval for reg %u\n", inv
->Reg
);
1161 /* remove interval j from active list */
1162 remove_interval(&activeIntervals
, inv
);
1163 j
--; /* counter-act j++ in for-loop above */
1165 /* return register regNew to the free pool */
1167 printf(" free reg %d\n", regNew
);
1168 ASSERT(usedRegs
[regNew
] == GL_TRUE
);
1169 usedRegs
[regNew
] = GL_FALSE
;
1174 /* find a free register for this live interval */
1176 const GLint k
= alloc_register(usedRegs
);
1178 /* out of registers, give up */
1181 registerMap
[live
->Reg
] = k
;
1182 maxTemp
= MAX2(maxTemp
, k
);
1184 printf(" remap register %u -> %d\n", live
->Reg
, k
);
1187 /* Insert this live interval into the active list which is sorted
1188 * by increasing end points.
1190 insert_interval_by_end(&activeIntervals
, live
);
1194 if (maxTemp
+ 1 < (GLint
) liveIntervals
.Num
) {
1195 /* OK, we've reduced the number of registers needed.
1196 * Scan the program and replace all the old temporary register
1197 * indexes with the new indexes.
1199 replace_regs(prog
, PROGRAM_TEMPORARY
, registerMap
);
1201 prog
->NumTemporaries
= maxTemp
+ 1;
1205 printf("Optimize: End live-interval register reallocation\n");
1206 printf("Num temp regs before: %u after: %u\n",
1207 liveIntervals
.Num
, maxTemp
+ 1);
1208 _mesa_print_program(prog
);
1215 print_it(GLcontext
*ctx
, struct gl_program
*program
, const char *txt
) {
1216 fprintf(stderr
, "%s (%u inst):\n", txt
, program
->NumInstructions
);
1217 _mesa_print_program(program
);
1218 _mesa_print_program_parameters(ctx
, program
);
1219 fprintf(stderr
, "\n\n");
1225 * Apply optimizations to the given program to eliminate unnecessary
1226 * instructions, temp regs, etc.
1229 _mesa_optimize_program(GLcontext
*ctx
, struct gl_program
*program
)
1231 GLboolean any_change
;
1233 /* Stop when no modifications were output */
1235 any_change
= GL_FALSE
;
1236 _mesa_remove_extra_move_use(program
);
1237 if (_mesa_remove_dead_code_global(program
))
1238 any_change
= GL_TRUE
;
1239 if (_mesa_remove_extra_moves(program
))
1240 any_change
= GL_TRUE
;
1241 if (_mesa_remove_dead_code_local(program
))
1242 any_change
= GL_TRUE
;
1243 _mesa_reallocate_registers(program
);
1244 } while (any_change
);