2 * Copyright © 2012 Intel Corporation
4 * Permission is hereby granted, free of charge, to any person obtaining a
5 * copy of this software and associated documentation files (the "Software"),
6 * to deal in the Software without restriction, including without limitation
7 * the rights to use, copy, modify, merge, publish, distribute, sublicense,
8 * and/or sell copies of the Software, and to permit persons to whom the
9 * Software is furnished to do so, subject to the following conditions:
11 * The above copyright notice and this permission notice (including the next
12 * paragraph) shall be included in all copies or substantial portions of the
15 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
16 * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
17 * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL
18 * THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
19 * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING
20 * FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS
24 /** @file brw_eu_compact.c
26 * Instruction compaction is a feature of G45 and newer hardware that allows
27 * for a smaller instruction encoding.
29 * The instruction cache is on the order of 32KB, and many programs generate
30 * far more instructions than that. The instruction cache is built to barely
31 * keep up with instruction dispatch ability in cache hit cases -- L1
32 * instruction cache misses that still hit in the next level could limit
33 * throughput by around 50%.
35 * The idea of instruction compaction is that most instructions use a tiny
36 * subset of the GPU functionality, so we can encode what would be a 16 byte
37 * instruction in 8 bytes using some lookup tables for various fields.
40 * Instruction compaction capabilities vary subtly by generation.
42 * G45's support for instruction compaction is very limited. Jump counts on
43 * this generation are in units of 16-byte uncompacted instructions. As such,
44 * all jump targets must be 16-byte aligned. Also, all instructions must be
45 * naturally aligned, i.e. uncompacted instructions must be 16-byte aligned.
46 * A G45-only instruction, NENOP, must be used to provide padding to align
47 * uncompacted instructions.
49 * Gen5 removes these restrictions and changes jump counts to be in units of
50 * 8-byte compacted instructions, allowing jump targets to be only 8-byte
51 * aligned. Uncompacted instructions can also be placed on 8-byte boundaries.
53 * Gen6 adds the ability to compact instructions with a limited range of
54 * immediate values. Compactable immediates have 12 unrestricted bits, and a
55 * 13th bit that's replicated through the high 20 bits, to create the 32-bit
56 * value of DW3 in the uncompacted instruction word.
58 * On Gen7 we can compact some control flow instructions with a small positive
59 * immediate in the low bits of DW3, like ENDIF with the JIP field. Other
60 * control flow instructions with UIP cannot be compacted, because of the
61 * replicated 13th bit. No control flow instructions can be compacted on Gen6
62 * since the jump count field is not in DW3.
68 * else JIP (plus UIP on BDW+)
70 * while JIP (must be negative)
72 * Gen 8 adds support for compacting 3-src instructions.
76 #include "brw_shader.h"
77 #include "intel_asm_annotation.h"
78 #include "common/gen_debug.h"
80 static const uint32_t g45_control_index_table
[32] = {
115 static const uint32_t g45_datatype_table
[32] = {
116 0b001000000000100001,
117 0b001011010110101101,
118 0b001000001000110001,
119 0b001111011110111101,
120 0b001011010110101100,
121 0b001000000110101101,
122 0b001000000000100000,
123 0b010100010110110001,
124 0b001100011000101101,
125 0b001000000000100010,
126 0b001000001000110110,
127 0b010000001000110001,
128 0b001000001000110010,
129 0b011000001000110010,
130 0b001111011110111100,
131 0b001000000100101000,
132 0b010100011000110001,
133 0b001010010100101001,
134 0b001000001000101001,
135 0b010000001000110110,
136 0b101000001000110001,
137 0b001011011000101101,
138 0b001000000100001001,
139 0b001011011000101100,
140 0b110100011000110001,
141 0b001000001110111101,
142 0b110000001000110001,
143 0b011000000100101010,
144 0b101000001000101001,
145 0b001011010110001100,
146 0b001000000110100001,
150 static const uint16_t g45_subreg_table
[32] = {
185 static const uint16_t g45_src_index_table
[32] = {
220 static const uint32_t gen6_control_index_table
[32] = {
255 static const uint32_t gen6_datatype_table
[32] = {
256 0b001001110000000000,
257 0b001000110000100000,
258 0b001001110000000001,
259 0b001000000001100000,
260 0b001010110100101001,
261 0b001000000110101101,
262 0b001100011000101100,
263 0b001011110110101101,
264 0b001000000111101100,
265 0b001000000001100001,
266 0b001000110010100101,
267 0b001000000001000001,
268 0b001000001000110001,
269 0b001000001000101001,
270 0b001000000000100000,
271 0b001000001000110010,
272 0b001010010100101001,
273 0b001011010010100101,
274 0b001000000110100101,
275 0b001100011000101001,
276 0b001011011000101100,
277 0b001011010110100101,
278 0b001011110110100101,
279 0b001111011110111101,
280 0b001111011110111100,
281 0b001111011110111101,
282 0b001111011110011101,
283 0b001111011110111110,
284 0b001000000000100001,
285 0b001000000000100010,
286 0b001001111111011101,
287 0b001000001110111110,
290 static const uint16_t gen6_subreg_table
[32] = {
325 static const uint16_t gen6_src_index_table
[32] = {
360 static const uint32_t gen7_control_index_table
[32] = {
361 0b0000000000000000010,
362 0b0000100000000000000,
363 0b0000100000000000001,
364 0b0000100000000000010,
365 0b0000100000000000011,
366 0b0000100000000000100,
367 0b0000100000000000101,
368 0b0000100000000000111,
369 0b0000100000000001000,
370 0b0000100000000001001,
371 0b0000100000000001101,
372 0b0000110000000000000,
373 0b0000110000000000001,
374 0b0000110000000000010,
375 0b0000110000000000011,
376 0b0000110000000000100,
377 0b0000110000000000101,
378 0b0000110000000000111,
379 0b0000110000000001001,
380 0b0000110000000001101,
381 0b0000110000000010000,
382 0b0000110000100000000,
383 0b0001000000000000000,
384 0b0001000000000000010,
385 0b0001000000000000100,
386 0b0001000000100000000,
387 0b0010110000000000000,
388 0b0010110000000010000,
389 0b0011000000000000000,
390 0b0011000000100000000,
391 0b0101000000000000000,
392 0b0101000000100000000
395 static const uint32_t gen7_datatype_table
[32] = {
396 0b001000000000000001,
397 0b001000000000100000,
398 0b001000000000100001,
399 0b001000000001100001,
400 0b001000000010111101,
401 0b001000001011111101,
402 0b001000001110100001,
403 0b001000001110100101,
404 0b001000001110111101,
405 0b001000010000100001,
406 0b001000110000100000,
407 0b001000110000100001,
408 0b001001010010100101,
409 0b001001110010100100,
410 0b001001110010100101,
411 0b001111001110111101,
412 0b001111011110011101,
413 0b001111011110111100,
414 0b001111011110111101,
415 0b001111111110111100,
416 0b000000001000001100,
417 0b001000000000111101,
418 0b001000000010100101,
419 0b001000010000100000,
420 0b001001010010100100,
421 0b001001110010000100,
422 0b001010010100001001,
423 0b001101111110111101,
424 0b001111111110111101,
425 0b001011110110101100,
426 0b001010010100101000,
430 static const uint16_t gen7_subreg_table
[32] = {
465 static const uint16_t gen7_src_index_table
[32] = {
500 static const uint32_t gen8_control_index_table
[32] = {
501 0b0000000000000000010,
502 0b0000100000000000000,
503 0b0000100000000000001,
504 0b0000100000000000010,
505 0b0000100000000000011,
506 0b0000100000000000100,
507 0b0000100000000000101,
508 0b0000100000000000111,
509 0b0000100000000001000,
510 0b0000100000000001001,
511 0b0000100000000001101,
512 0b0000110000000000000,
513 0b0000110000000000001,
514 0b0000110000000000010,
515 0b0000110000000000011,
516 0b0000110000000000100,
517 0b0000110000000000101,
518 0b0000110000000000111,
519 0b0000110000000001001,
520 0b0000110000000001101,
521 0b0000110000000010000,
522 0b0000110000100000000,
523 0b0001000000000000000,
524 0b0001000000000000010,
525 0b0001000000000000100,
526 0b0001000000100000000,
527 0b0010110000000000000,
528 0b0010110000000010000,
529 0b0011000000000000000,
530 0b0011000000100000000,
531 0b0101000000000000000,
532 0b0101000000100000000
535 static const uint32_t gen8_datatype_table
[32] = {
536 0b001000000000000000001,
537 0b001000000000001000000,
538 0b001000000000001000001,
539 0b001000000000011000001,
540 0b001000000000101011101,
541 0b001000000010111011101,
542 0b001000000011101000001,
543 0b001000000011101000101,
544 0b001000000011101011101,
545 0b001000001000001000001,
546 0b001000011000001000000,
547 0b001000011000001000001,
548 0b001000101000101000101,
549 0b001000111000101000100,
550 0b001000111000101000101,
551 0b001011100011101011101,
552 0b001011101011100011101,
553 0b001011101011101011100,
554 0b001011101011101011101,
555 0b001011111011101011100,
556 0b000000000010000001100,
557 0b001000000000001011101,
558 0b001000000000101000101,
559 0b001000001000001000000,
560 0b001000101000101000100,
561 0b001000111000100000100,
562 0b001001001001000001001,
563 0b001010111011101011101,
564 0b001011111011101011101,
565 0b001001111001101001100,
566 0b001001001001001001000,
567 0b001001011001001001000
570 static const uint16_t gen8_subreg_table
[32] = {
605 static const uint16_t gen8_src_index_table
[32] = {
640 /* This is actually the control index table for Cherryview (26 bits), but the
641 * only difference from Broadwell (24 bits) is that it has two extra 0-bits at
644 * The low 24 bits have the same mappings on both hardware.
646 static const uint32_t gen8_3src_control_index_table
[4] = {
647 0b00100000000110000000000001,
648 0b00000000000110000000000001,
649 0b00000000001000000000000001,
650 0b00000000001000000000100001
653 /* This is actually the control index table for Cherryview (49 bits), but the
654 * only difference from Broadwell (46 bits) is that it has three extra 0-bits
657 * The low 44 bits have the same mappings on both hardware, and since the high
658 * three bits on Broadwell are zero, we can reuse Cherryview's table.
660 static const uint64_t gen8_3src_source_index_table
[4] = {
661 0b0000001110010011100100111001000001111000000000000,
662 0b0000001110010011100100111001000001111000000000010,
663 0b0000001110010011100100111001000001111000000001000,
664 0b0000001110010011100100111001000001111000000100000
667 static const uint32_t *control_index_table
;
668 static const uint32_t *datatype_table
;
669 static const uint16_t *subreg_table
;
670 static const uint16_t *src_index_table
;
673 set_control_index(const struct gen_device_info
*devinfo
,
674 brw_compact_inst
*dst
, const brw_inst
*src
)
676 uint32_t uncompacted
= devinfo
->gen
>= 8 /* 17b/G45; 19b/IVB+ */
677 ? (brw_inst_bits(src
, 33, 31) << 16) | /* 3b */
678 (brw_inst_bits(src
, 23, 12) << 4) | /* 12b */
679 (brw_inst_bits(src
, 10, 9) << 2) | /* 2b */
680 (brw_inst_bits(src
, 34, 34) << 1) | /* 1b */
681 (brw_inst_bits(src
, 8, 8)) /* 1b */
682 : (brw_inst_bits(src
, 31, 31) << 16) | /* 1b */
683 (brw_inst_bits(src
, 23, 8)); /* 16b */
685 /* On gen7, the flag register and subregister numbers are integrated into
688 if (devinfo
->gen
== 7)
689 uncompacted
|= brw_inst_bits(src
, 90, 89) << 17; /* 2b */
691 for (int i
= 0; i
< 32; i
++) {
692 if (control_index_table
[i
] == uncompacted
) {
693 brw_compact_inst_set_control_index(devinfo
, dst
, i
);
702 set_datatype_index(const struct gen_device_info
*devinfo
, brw_compact_inst
*dst
,
705 uint32_t uncompacted
= devinfo
->gen
>= 8 /* 18b/G45+; 21b/BDW+ */
706 ? (brw_inst_bits(src
, 63, 61) << 18) | /* 3b */
707 (brw_inst_bits(src
, 94, 89) << 12) | /* 6b */
708 (brw_inst_bits(src
, 46, 35)) /* 12b */
709 : (brw_inst_bits(src
, 63, 61) << 15) | /* 3b */
710 (brw_inst_bits(src
, 46, 32)); /* 15b */
712 for (int i
= 0; i
< 32; i
++) {
713 if (datatype_table
[i
] == uncompacted
) {
714 brw_compact_inst_set_datatype_index(devinfo
, dst
, i
);
723 set_subreg_index(const struct gen_device_info
*devinfo
, brw_compact_inst
*dst
,
724 const brw_inst
*src
, bool is_immediate
)
726 uint16_t uncompacted
= /* 15b */
727 (brw_inst_bits(src
, 52, 48) << 0) | /* 5b */
728 (brw_inst_bits(src
, 68, 64) << 5); /* 5b */
731 uncompacted
|= brw_inst_bits(src
, 100, 96) << 10; /* 5b */
733 for (int i
= 0; i
< 32; i
++) {
734 if (subreg_table
[i
] == uncompacted
) {
735 brw_compact_inst_set_subreg_index(devinfo
, dst
, i
);
744 get_src_index(uint16_t uncompacted
,
747 for (int i
= 0; i
< 32; i
++) {
748 if (src_index_table
[i
] == uncompacted
) {
758 set_src0_index(const struct gen_device_info
*devinfo
,
759 brw_compact_inst
*dst
, const brw_inst
*src
)
762 uint16_t uncompacted
= brw_inst_bits(src
, 88, 77); /* 12b */
764 if (!get_src_index(uncompacted
, &compacted
))
767 brw_compact_inst_set_src0_index(devinfo
, dst
, compacted
);
773 set_src1_index(const struct gen_device_info
*devinfo
, brw_compact_inst
*dst
,
774 const brw_inst
*src
, bool is_immediate
)
779 compacted
= (brw_inst_imm_ud(devinfo
, src
) >> 8) & 0x1f;
781 uint16_t uncompacted
= brw_inst_bits(src
, 120, 109); /* 12b */
783 if (!get_src_index(uncompacted
, &compacted
))
787 brw_compact_inst_set_src1_index(devinfo
, dst
, compacted
);
793 set_3src_control_index(const struct gen_device_info
*devinfo
,
794 brw_compact_inst
*dst
, const brw_inst
*src
)
796 assert(devinfo
->gen
>= 8);
798 uint32_t uncompacted
= /* 24b/BDW; 26b/CHV */
799 (brw_inst_bits(src
, 34, 32) << 21) | /* 3b */
800 (brw_inst_bits(src
, 28, 8)); /* 21b */
802 if (devinfo
->gen
>= 9 || devinfo
->is_cherryview
)
803 uncompacted
|= brw_inst_bits(src
, 36, 35) << 24; /* 2b */
805 for (unsigned i
= 0; i
< ARRAY_SIZE(gen8_3src_control_index_table
); i
++) {
806 if (gen8_3src_control_index_table
[i
] == uncompacted
) {
807 brw_compact_inst_set_3src_control_index(devinfo
, dst
, i
);
816 set_3src_source_index(const struct gen_device_info
*devinfo
,
817 brw_compact_inst
*dst
, const brw_inst
*src
)
819 assert(devinfo
->gen
>= 8);
821 uint64_t uncompacted
= /* 46b/BDW; 49b/CHV */
822 (brw_inst_bits(src
, 83, 83) << 43) | /* 1b */
823 (brw_inst_bits(src
, 114, 107) << 35) | /* 8b */
824 (brw_inst_bits(src
, 93, 86) << 27) | /* 8b */
825 (brw_inst_bits(src
, 72, 65) << 19) | /* 8b */
826 (brw_inst_bits(src
, 55, 37)); /* 19b */
828 if (devinfo
->gen
>= 9 || devinfo
->is_cherryview
) {
830 (brw_inst_bits(src
, 126, 125) << 47) | /* 2b */
831 (brw_inst_bits(src
, 105, 104) << 45) | /* 2b */
832 (brw_inst_bits(src
, 84, 84) << 44); /* 1b */
835 (brw_inst_bits(src
, 125, 125) << 45) | /* 1b */
836 (brw_inst_bits(src
, 104, 104) << 44); /* 1b */
839 for (unsigned i
= 0; i
< ARRAY_SIZE(gen8_3src_source_index_table
); i
++) {
840 if (gen8_3src_source_index_table
[i
] == uncompacted
) {
841 brw_compact_inst_set_3src_source_index(devinfo
, dst
, i
);
850 has_unmapped_bits(const struct gen_device_info
*devinfo
, const brw_inst
*src
)
852 /* EOT can only be mapped on a send if the src1 is an immediate */
853 if ((brw_inst_opcode(devinfo
, src
) == BRW_OPCODE_SENDC
||
854 brw_inst_opcode(devinfo
, src
) == BRW_OPCODE_SEND
) &&
855 brw_inst_eot(devinfo
, src
))
858 /* Check for instruction bits that don't map to any of the fields of the
859 * compacted instruction. The instruction cannot be compacted if any of
860 * them are set. They overlap with:
861 * - NibCtrl (bit 47 on Gen7, bit 11 on Gen8)
862 * - Dst.AddrImm[9] (bit 47 on Gen8)
863 * - Src0.AddrImm[9] (bit 95 on Gen8)
864 * - Imm64[27:31] (bits 91-95 on Gen7, bit 95 on Gen8)
865 * - UIP[31] (bit 95 on Gen8)
867 if (devinfo
->gen
>= 8) {
868 assert(!brw_inst_bits(src
, 7, 7));
869 return brw_inst_bits(src
, 95, 95) ||
870 brw_inst_bits(src
, 47, 47) ||
871 brw_inst_bits(src
, 11, 11);
873 assert(!brw_inst_bits(src
, 7, 7) &&
874 !(devinfo
->gen
< 7 && brw_inst_bits(src
, 90, 90)));
875 return brw_inst_bits(src
, 95, 91) ||
876 brw_inst_bits(src
, 47, 47);
881 has_3src_unmapped_bits(const struct gen_device_info
*devinfo
,
884 /* Check for three-source instruction bits that don't map to any of the
885 * fields of the compacted instruction. All of them seem to be reserved
888 if (devinfo
->gen
>= 9 || devinfo
->is_cherryview
) {
889 assert(!brw_inst_bits(src
, 127, 127) &&
890 !brw_inst_bits(src
, 7, 7));
892 assert(devinfo
->gen
>= 8);
893 assert(!brw_inst_bits(src
, 127, 126) &&
894 !brw_inst_bits(src
, 105, 105) &&
895 !brw_inst_bits(src
, 84, 84) &&
896 !brw_inst_bits(src
, 36, 35) &&
897 !brw_inst_bits(src
, 7, 7));
904 brw_try_compact_3src_instruction(const struct gen_device_info
*devinfo
,
905 brw_compact_inst
*dst
, const brw_inst
*src
)
907 assert(devinfo
->gen
>= 8);
909 if (has_3src_unmapped_bits(devinfo
, src
))
912 #define compact(field) \
913 brw_compact_inst_set_3src_##field(devinfo, dst, brw_inst_3src_##field(devinfo, src))
914 #define compact_a16(field) \
915 brw_compact_inst_set_3src_##field(devinfo, dst, brw_inst_3src_a16_##field(devinfo, src))
919 if (!set_3src_control_index(devinfo
, dst
, src
))
922 if (!set_3src_source_index(devinfo
, dst
, src
))
926 compact_a16(src0_rep_ctrl
);
927 brw_compact_inst_set_3src_cmpt_control(devinfo
, dst
, true);
928 compact(debug_control
);
930 compact_a16(src1_rep_ctrl
);
931 compact_a16(src2_rep_ctrl
);
932 compact(src0_reg_nr
);
933 compact(src1_reg_nr
);
934 compact(src2_reg_nr
);
935 compact_a16(src0_subreg_nr
);
936 compact_a16(src1_subreg_nr
);
937 compact_a16(src2_subreg_nr
);
945 /* Compacted instructions have 12-bits for immediate sources, and a 13th bit
946 * that's replicated through the high 20 bits.
948 * Effectively this means we get 12-bit integers, 0.0f, and some limited uses
949 * of packed vectors as compactable immediates.
952 is_compactable_immediate(unsigned imm
)
954 /* We get the low 12 bits as-is. */
957 /* We get one bit replicated through the top 20 bits. */
958 return imm
== 0 || imm
== 0xfffff000;
962 * Applies some small changes to instruction types to increase chances of
966 precompact(const struct gen_device_info
*devinfo
, brw_inst inst
)
968 if (brw_inst_src0_reg_file(devinfo
, &inst
) != BRW_IMMEDIATE_VALUE
)
971 /* The Bspec's section titled "Non-present Operands" claims that if src0
972 * is an immediate that src1's type must be the same as that of src0.
974 * The SNB+ DataTypeIndex instruction compaction tables contain mappings
975 * that do not follow this rule. E.g., from the IVB/HSW table:
977 * DataTypeIndex 18-Bit Mapping Mapped Meaning
978 * 3 001000001011111101 r:f | i:vf | a:ud | <1> | dir |
980 * And from the SNB table:
982 * DataTypeIndex 18-Bit Mapping Mapped Meaning
983 * 8 001000000111101100 a:w | i:w | a:ud | <1> | dir |
985 * Neither of these cause warnings from the simulator when used,
986 * compacted or otherwise. In fact, all compaction mappings that have an
987 * immediate in src0 use a:ud for src1.
989 * The GM45 instruction compaction tables do not contain mapped meanings
990 * so it's not clear whether it has the restriction. We'll assume it was
991 * lifted on SNB. (FINISHME: decode the GM45 tables and check.)
993 * Don't do any of this for 64-bit immediates, since the src1 fields
994 * overlap with the immediate and setting them would overwrite the
997 if (devinfo
->gen
>= 6 &&
998 !(devinfo
->is_haswell
&&
999 brw_inst_opcode(devinfo
, &inst
) == BRW_OPCODE_DIM
) &&
1000 !(devinfo
->gen
>= 8 &&
1001 (brw_inst_src0_type(devinfo
, &inst
) == BRW_REGISTER_TYPE_DF
||
1002 brw_inst_src0_type(devinfo
, &inst
) == BRW_REGISTER_TYPE_UQ
||
1003 brw_inst_src0_type(devinfo
, &inst
) == BRW_REGISTER_TYPE_Q
))) {
1004 enum brw_reg_file file
= brw_inst_src1_reg_file(devinfo
, &inst
);
1005 brw_inst_set_src1_file_type(devinfo
, &inst
, file
, BRW_REGISTER_TYPE_UD
);
1008 /* Compacted instructions only have 12-bits (plus 1 for the other 20)
1009 * for immediate values. Presumably the hardware engineers realized
1010 * that the only useful floating-point value that could be represented
1011 * in this format is 0.0, which can also be represented as a VF-typed
1012 * immediate, so they gave us the previously mentioned mapping on IVB+.
1014 * Strangely, we do have a mapping for imm:f in src1, so we don't need
1017 * If we see a 0.0:F, change the type to VF so that it can be compacted.
1019 if (brw_inst_imm_ud(devinfo
, &inst
) == 0x0 &&
1020 brw_inst_src0_type(devinfo
, &inst
) == BRW_REGISTER_TYPE_F
&&
1021 brw_inst_dst_type(devinfo
, &inst
) == BRW_REGISTER_TYPE_F
&&
1022 brw_inst_dst_hstride(devinfo
, &inst
) == BRW_HORIZONTAL_STRIDE_1
) {
1023 enum brw_reg_file file
= brw_inst_src0_reg_file(devinfo
, &inst
);
1024 brw_inst_set_src0_file_type(devinfo
, &inst
, file
, BRW_REGISTER_TYPE_VF
);
1027 /* There are no mappings for dst:d | i:d, so if the immediate is suitable
1028 * set the types to :UD so the instruction can be compacted.
1030 if (is_compactable_immediate(brw_inst_imm_ud(devinfo
, &inst
)) &&
1031 brw_inst_cond_modifier(devinfo
, &inst
) == BRW_CONDITIONAL_NONE
&&
1032 brw_inst_src0_type(devinfo
, &inst
) == BRW_REGISTER_TYPE_D
&&
1033 brw_inst_dst_type(devinfo
, &inst
) == BRW_REGISTER_TYPE_D
) {
1034 enum brw_reg_file src_file
= brw_inst_src0_reg_file(devinfo
, &inst
);
1035 enum brw_reg_file dst_file
= brw_inst_dst_reg_file(devinfo
, &inst
);
1037 brw_inst_set_src0_file_type(devinfo
, &inst
, src_file
, BRW_REGISTER_TYPE_UD
);
1038 brw_inst_set_dst_file_type(devinfo
, &inst
, dst_file
, BRW_REGISTER_TYPE_UD
);
1045 * Tries to compact instruction src into dst.
1047 * It doesn't modify dst unless src is compactable, which is relied on by
1048 * brw_compact_instructions().
1051 brw_try_compact_instruction(const struct gen_device_info
*devinfo
,
1052 brw_compact_inst
*dst
, const brw_inst
*src
)
1054 brw_compact_inst temp
;
1056 assert(brw_inst_cmpt_control(devinfo
, src
) == 0);
1058 if (is_3src(devinfo
, brw_inst_opcode(devinfo
, src
))) {
1059 if (devinfo
->gen
>= 8) {
1060 memset(&temp
, 0, sizeof(temp
));
1061 if (brw_try_compact_3src_instruction(devinfo
, &temp
, src
)) {
1073 brw_inst_src0_reg_file(devinfo
, src
) == BRW_IMMEDIATE_VALUE
||
1074 brw_inst_src1_reg_file(devinfo
, src
) == BRW_IMMEDIATE_VALUE
;
1076 (devinfo
->gen
< 6 ||
1077 !is_compactable_immediate(brw_inst_imm_ud(devinfo
, src
)))) {
1081 if (has_unmapped_bits(devinfo
, src
))
1084 memset(&temp
, 0, sizeof(temp
));
1086 #define compact(field) \
1087 brw_compact_inst_set_##field(devinfo, &temp, brw_inst_##field(devinfo, src))
1090 compact(debug_control
);
1092 if (!set_control_index(devinfo
, &temp
, src
))
1094 if (!set_datatype_index(devinfo
, &temp
, src
))
1096 if (!set_subreg_index(devinfo
, &temp
, src
, is_immediate
))
1099 if (devinfo
->gen
>= 6) {
1100 compact(acc_wr_control
);
1102 compact(mask_control_ex
);
1105 compact(cond_modifier
);
1107 if (devinfo
->gen
<= 6)
1108 compact(flag_subreg_nr
);
1110 brw_compact_inst_set_cmpt_control(devinfo
, &temp
, true);
1112 if (!set_src0_index(devinfo
, &temp
, src
))
1114 if (!set_src1_index(devinfo
, &temp
, src
, is_immediate
))
1117 brw_compact_inst_set_dst_reg_nr(devinfo
, &temp
,
1118 brw_inst_dst_da_reg_nr(devinfo
, src
));
1119 brw_compact_inst_set_src0_reg_nr(devinfo
, &temp
,
1120 brw_inst_src0_da_reg_nr(devinfo
, src
));
1123 brw_compact_inst_set_src1_reg_nr(devinfo
, &temp
,
1124 brw_inst_imm_ud(devinfo
, src
) & 0xff);
1126 brw_compact_inst_set_src1_reg_nr(devinfo
, &temp
,
1127 brw_inst_src1_da_reg_nr(devinfo
, src
));
1138 set_uncompacted_control(const struct gen_device_info
*devinfo
, brw_inst
*dst
,
1139 brw_compact_inst
*src
)
1141 uint32_t uncompacted
=
1142 control_index_table
[brw_compact_inst_control_index(devinfo
, src
)];
1144 if (devinfo
->gen
>= 8) {
1145 brw_inst_set_bits(dst
, 33, 31, (uncompacted
>> 16));
1146 brw_inst_set_bits(dst
, 23, 12, (uncompacted
>> 4) & 0xfff);
1147 brw_inst_set_bits(dst
, 10, 9, (uncompacted
>> 2) & 0x3);
1148 brw_inst_set_bits(dst
, 34, 34, (uncompacted
>> 1) & 0x1);
1149 brw_inst_set_bits(dst
, 8, 8, (uncompacted
>> 0) & 0x1);
1151 brw_inst_set_bits(dst
, 31, 31, (uncompacted
>> 16) & 0x1);
1152 brw_inst_set_bits(dst
, 23, 8, (uncompacted
& 0xffff));
1154 if (devinfo
->gen
== 7)
1155 brw_inst_set_bits(dst
, 90, 89, uncompacted
>> 17);
1160 set_uncompacted_datatype(const struct gen_device_info
*devinfo
, brw_inst
*dst
,
1161 brw_compact_inst
*src
)
1163 uint32_t uncompacted
=
1164 datatype_table
[brw_compact_inst_datatype_index(devinfo
, src
)];
1166 if (devinfo
->gen
>= 8) {
1167 brw_inst_set_bits(dst
, 63, 61, (uncompacted
>> 18));
1168 brw_inst_set_bits(dst
, 94, 89, (uncompacted
>> 12) & 0x3f);
1169 brw_inst_set_bits(dst
, 46, 35, (uncompacted
>> 0) & 0xfff);
1171 brw_inst_set_bits(dst
, 63, 61, (uncompacted
>> 15));
1172 brw_inst_set_bits(dst
, 46, 32, (uncompacted
& 0x7fff));
1177 set_uncompacted_subreg(const struct gen_device_info
*devinfo
, brw_inst
*dst
,
1178 brw_compact_inst
*src
)
1180 uint16_t uncompacted
=
1181 subreg_table
[brw_compact_inst_subreg_index(devinfo
, src
)];
1183 brw_inst_set_bits(dst
, 100, 96, (uncompacted
>> 10));
1184 brw_inst_set_bits(dst
, 68, 64, (uncompacted
>> 5) & 0x1f);
1185 brw_inst_set_bits(dst
, 52, 48, (uncompacted
>> 0) & 0x1f);
1189 set_uncompacted_src0(const struct gen_device_info
*devinfo
, brw_inst
*dst
,
1190 brw_compact_inst
*src
)
1192 uint32_t compacted
= brw_compact_inst_src0_index(devinfo
, src
);
1193 uint16_t uncompacted
= src_index_table
[compacted
];
1195 brw_inst_set_bits(dst
, 88, 77, uncompacted
);
1199 set_uncompacted_src1(const struct gen_device_info
*devinfo
, brw_inst
*dst
,
1200 brw_compact_inst
*src
, bool is_immediate
)
1203 signed high5
= brw_compact_inst_src1_index(devinfo
, src
);
1204 /* Replicate top bit of src1_index into high 20 bits of the immediate. */
1205 brw_inst_set_imm_ud(devinfo
, dst
, (high5
<< 27) >> 19);
1207 uint16_t uncompacted
=
1208 src_index_table
[brw_compact_inst_src1_index(devinfo
, src
)];
1210 brw_inst_set_bits(dst
, 120, 109, uncompacted
);
1215 set_uncompacted_3src_control_index(const struct gen_device_info
*devinfo
,
1216 brw_inst
*dst
, brw_compact_inst
*src
)
1218 assert(devinfo
->gen
>= 8);
1220 uint32_t compacted
= brw_compact_inst_3src_control_index(devinfo
, src
);
1221 uint32_t uncompacted
= gen8_3src_control_index_table
[compacted
];
1223 brw_inst_set_bits(dst
, 34, 32, (uncompacted
>> 21) & 0x7);
1224 brw_inst_set_bits(dst
, 28, 8, (uncompacted
>> 0) & 0x1fffff);
1226 if (devinfo
->gen
>= 9 || devinfo
->is_cherryview
)
1227 brw_inst_set_bits(dst
, 36, 35, (uncompacted
>> 24) & 0x3);
1231 set_uncompacted_3src_source_index(const struct gen_device_info
*devinfo
,
1232 brw_inst
*dst
, brw_compact_inst
*src
)
1234 assert(devinfo
->gen
>= 8);
1236 uint32_t compacted
= brw_compact_inst_3src_source_index(devinfo
, src
);
1237 uint64_t uncompacted
= gen8_3src_source_index_table
[compacted
];
1239 brw_inst_set_bits(dst
, 83, 83, (uncompacted
>> 43) & 0x1);
1240 brw_inst_set_bits(dst
, 114, 107, (uncompacted
>> 35) & 0xff);
1241 brw_inst_set_bits(dst
, 93, 86, (uncompacted
>> 27) & 0xff);
1242 brw_inst_set_bits(dst
, 72, 65, (uncompacted
>> 19) & 0xff);
1243 brw_inst_set_bits(dst
, 55, 37, (uncompacted
>> 0) & 0x7ffff);
1245 if (devinfo
->gen
>= 9 || devinfo
->is_cherryview
) {
1246 brw_inst_set_bits(dst
, 126, 125, (uncompacted
>> 47) & 0x3);
1247 brw_inst_set_bits(dst
, 105, 104, (uncompacted
>> 45) & 0x3);
1248 brw_inst_set_bits(dst
, 84, 84, (uncompacted
>> 44) & 0x1);
1250 brw_inst_set_bits(dst
, 125, 125, (uncompacted
>> 45) & 0x1);
1251 brw_inst_set_bits(dst
, 104, 104, (uncompacted
>> 44) & 0x1);
1256 brw_uncompact_3src_instruction(const struct gen_device_info
*devinfo
,
1257 brw_inst
*dst
, brw_compact_inst
*src
)
1259 assert(devinfo
->gen
>= 8);
1261 #define uncompact(field) \
1262 brw_inst_set_3src_##field(devinfo, dst, brw_compact_inst_3src_##field(devinfo, src))
1263 #define uncompact_a16(field) \
1264 brw_inst_set_3src_a16_##field(devinfo, dst, brw_compact_inst_3src_##field(devinfo, src))
1268 set_uncompacted_3src_control_index(devinfo
, dst
, src
);
1269 set_uncompacted_3src_source_index(devinfo
, dst
, src
);
1271 uncompact(dst_reg_nr
);
1272 uncompact_a16(src0_rep_ctrl
);
1273 brw_inst_set_3src_cmpt_control(devinfo
, dst
, false);
1274 uncompact(debug_control
);
1275 uncompact(saturate
);
1276 uncompact_a16(src1_rep_ctrl
);
1277 uncompact_a16(src2_rep_ctrl
);
1278 uncompact(src0_reg_nr
);
1279 uncompact(src1_reg_nr
);
1280 uncompact(src2_reg_nr
);
1281 uncompact_a16(src0_subreg_nr
);
1282 uncompact_a16(src1_subreg_nr
);
1283 uncompact_a16(src2_subreg_nr
);
1286 #undef uncompact_a16
1290 brw_uncompact_instruction(const struct gen_device_info
*devinfo
, brw_inst
*dst
,
1291 brw_compact_inst
*src
)
1293 memset(dst
, 0, sizeof(*dst
));
1295 if (devinfo
->gen
>= 8 &&
1296 is_3src(devinfo
, brw_compact_inst_3src_opcode(devinfo
, src
))) {
1297 brw_uncompact_3src_instruction(devinfo
, dst
, src
);
1301 #define uncompact(field) \
1302 brw_inst_set_##field(devinfo, dst, brw_compact_inst_##field(devinfo, src))
1305 uncompact(debug_control
);
1307 set_uncompacted_control(devinfo
, dst
, src
);
1308 set_uncompacted_datatype(devinfo
, dst
, src
);
1310 /* src0/1 register file fields are in the datatype table. */
1311 bool is_immediate
= brw_inst_src0_reg_file(devinfo
, dst
) == BRW_IMMEDIATE_VALUE
||
1312 brw_inst_src1_reg_file(devinfo
, dst
) == BRW_IMMEDIATE_VALUE
;
1314 set_uncompacted_subreg(devinfo
, dst
, src
);
1316 if (devinfo
->gen
>= 6) {
1317 uncompact(acc_wr_control
);
1319 uncompact(mask_control_ex
);
1322 uncompact(cond_modifier
);
1324 if (devinfo
->gen
<= 6)
1325 uncompact(flag_subreg_nr
);
1327 set_uncompacted_src0(devinfo
, dst
, src
);
1328 set_uncompacted_src1(devinfo
, dst
, src
, is_immediate
);
1330 brw_inst_set_dst_da_reg_nr(devinfo
, dst
,
1331 brw_compact_inst_dst_reg_nr(devinfo
, src
));
1332 brw_inst_set_src0_da_reg_nr(devinfo
, dst
,
1333 brw_compact_inst_src0_reg_nr(devinfo
, src
));
1336 brw_inst_set_imm_ud(devinfo
, dst
,
1337 brw_inst_imm_ud(devinfo
, dst
) |
1338 brw_compact_inst_src1_reg_nr(devinfo
, src
));
1340 brw_inst_set_src1_da_reg_nr(devinfo
, dst
,
1341 brw_compact_inst_src1_reg_nr(devinfo
, src
));
1347 void brw_debug_compact_uncompact(const struct gen_device_info
*devinfo
,
1349 brw_inst
*uncompacted
)
1351 fprintf(stderr
, "Instruction compact/uncompact changed (gen%d):\n",
1354 fprintf(stderr
, " before: ");
1355 brw_disassemble_inst(stderr
, devinfo
, orig
, true);
1357 fprintf(stderr
, " after: ");
1358 brw_disassemble_inst(stderr
, devinfo
, uncompacted
, false);
1360 uint32_t *before_bits
= (uint32_t *)orig
;
1361 uint32_t *after_bits
= (uint32_t *)uncompacted
;
1362 fprintf(stderr
, " changed bits:\n");
1363 for (int i
= 0; i
< 128; i
++) {
1364 uint32_t before
= before_bits
[i
/ 32] & (1 << (i
& 31));
1365 uint32_t after
= after_bits
[i
/ 32] & (1 << (i
& 31));
1367 if (before
!= after
) {
1368 fprintf(stderr
, " bit %d, %s to %s\n", i
,
1369 before
? "set" : "unset",
1370 after
? "set" : "unset");
1376 compacted_between(int old_ip
, int old_target_ip
, int *compacted_counts
)
1378 int this_compacted_count
= compacted_counts
[old_ip
];
1379 int target_compacted_count
= compacted_counts
[old_target_ip
];
1380 return target_compacted_count
- this_compacted_count
;
1384 update_uip_jip(const struct gen_device_info
*devinfo
, brw_inst
*insn
,
1385 int this_old_ip
, int *compacted_counts
)
1387 /* JIP and UIP are in units of:
1388 * - bytes on Gen8+; and
1389 * - compacted instructions on Gen6+.
1391 int shift
= devinfo
->gen
>= 8 ? 3 : 0;
1393 int32_t jip_compacted
= brw_inst_jip(devinfo
, insn
) >> shift
;
1394 jip_compacted
-= compacted_between(this_old_ip
,
1395 this_old_ip
+ (jip_compacted
/ 2),
1397 brw_inst_set_jip(devinfo
, insn
, jip_compacted
<< shift
);
1399 if (brw_inst_opcode(devinfo
, insn
) == BRW_OPCODE_ENDIF
||
1400 brw_inst_opcode(devinfo
, insn
) == BRW_OPCODE_WHILE
||
1401 (brw_inst_opcode(devinfo
, insn
) == BRW_OPCODE_ELSE
&& devinfo
->gen
<= 7))
1404 int32_t uip_compacted
= brw_inst_uip(devinfo
, insn
) >> shift
;
1405 uip_compacted
-= compacted_between(this_old_ip
,
1406 this_old_ip
+ (uip_compacted
/ 2),
1408 brw_inst_set_uip(devinfo
, insn
, uip_compacted
<< shift
);
1412 update_gen4_jump_count(const struct gen_device_info
*devinfo
, brw_inst
*insn
,
1413 int this_old_ip
, int *compacted_counts
)
1415 assert(devinfo
->gen
== 5 || devinfo
->is_g4x
);
1417 /* Jump Count is in units of:
1418 * - uncompacted instructions on G45; and
1419 * - compacted instructions on Gen5.
1421 int shift
= devinfo
->is_g4x
? 1 : 0;
1423 int jump_count_compacted
= brw_inst_gen4_jump_count(devinfo
, insn
) << shift
;
1425 int target_old_ip
= this_old_ip
+ (jump_count_compacted
/ 2);
1427 int this_compacted_count
= compacted_counts
[this_old_ip
];
1428 int target_compacted_count
= compacted_counts
[target_old_ip
];
1430 jump_count_compacted
-= (target_compacted_count
- this_compacted_count
);
1431 brw_inst_set_gen4_jump_count(devinfo
, insn
, jump_count_compacted
>> shift
);
1435 brw_init_compaction_tables(const struct gen_device_info
*devinfo
)
1437 assert(g45_control_index_table
[ARRAY_SIZE(g45_control_index_table
) - 1] != 0);
1438 assert(g45_datatype_table
[ARRAY_SIZE(g45_datatype_table
) - 1] != 0);
1439 assert(g45_subreg_table
[ARRAY_SIZE(g45_subreg_table
) - 1] != 0);
1440 assert(g45_src_index_table
[ARRAY_SIZE(g45_src_index_table
) - 1] != 0);
1441 assert(gen6_control_index_table
[ARRAY_SIZE(gen6_control_index_table
) - 1] != 0);
1442 assert(gen6_datatype_table
[ARRAY_SIZE(gen6_datatype_table
) - 1] != 0);
1443 assert(gen6_subreg_table
[ARRAY_SIZE(gen6_subreg_table
) - 1] != 0);
1444 assert(gen6_src_index_table
[ARRAY_SIZE(gen6_src_index_table
) - 1] != 0);
1445 assert(gen7_control_index_table
[ARRAY_SIZE(gen7_control_index_table
) - 1] != 0);
1446 assert(gen7_datatype_table
[ARRAY_SIZE(gen7_datatype_table
) - 1] != 0);
1447 assert(gen7_subreg_table
[ARRAY_SIZE(gen7_subreg_table
) - 1] != 0);
1448 assert(gen7_src_index_table
[ARRAY_SIZE(gen7_src_index_table
) - 1] != 0);
1449 assert(gen8_control_index_table
[ARRAY_SIZE(gen8_control_index_table
) - 1] != 0);
1450 assert(gen8_datatype_table
[ARRAY_SIZE(gen8_datatype_table
) - 1] != 0);
1451 assert(gen8_subreg_table
[ARRAY_SIZE(gen8_subreg_table
) - 1] != 0);
1452 assert(gen8_src_index_table
[ARRAY_SIZE(gen8_src_index_table
) - 1] != 0);
1454 switch (devinfo
->gen
) {
1458 control_index_table
= gen8_control_index_table
;
1459 datatype_table
= gen8_datatype_table
;
1460 subreg_table
= gen8_subreg_table
;
1461 src_index_table
= gen8_src_index_table
;
1464 control_index_table
= gen7_control_index_table
;
1465 datatype_table
= gen7_datatype_table
;
1466 subreg_table
= gen7_subreg_table
;
1467 src_index_table
= gen7_src_index_table
;
1470 control_index_table
= gen6_control_index_table
;
1471 datatype_table
= gen6_datatype_table
;
1472 subreg_table
= gen6_subreg_table
;
1473 src_index_table
= gen6_src_index_table
;
1477 control_index_table
= g45_control_index_table
;
1478 datatype_table
= g45_datatype_table
;
1479 subreg_table
= g45_subreg_table
;
1480 src_index_table
= g45_src_index_table
;
1483 unreachable("unknown generation");
1488 brw_compact_instructions(struct brw_codegen
*p
, int start_offset
,
1489 struct disasm_info
*disasm
)
1491 if (unlikely(INTEL_DEBUG
& DEBUG_NO_COMPACTION
))
1494 const struct gen_device_info
*devinfo
= p
->devinfo
;
1495 void *store
= p
->store
+ start_offset
/ 16;
1496 /* For an instruction at byte offset 16*i before compaction, this is the
1497 * number of compacted instructions minus the number of padding NOP/NENOPs
1500 int compacted_counts
[(p
->next_insn_offset
- start_offset
) / sizeof(brw_inst
)];
1501 /* For an instruction at byte offset 8*i after compaction, this was its IP
1502 * (in 16-byte units) before compaction.
1504 int old_ip
[(p
->next_insn_offset
- start_offset
) / sizeof(brw_compact_inst
) + 1];
1506 if (devinfo
->gen
== 4 && !devinfo
->is_g4x
)
1510 int compacted_count
= 0;
1511 for (int src_offset
= 0; src_offset
< p
->next_insn_offset
- start_offset
;
1512 src_offset
+= sizeof(brw_inst
)) {
1513 brw_inst
*src
= store
+ src_offset
;
1514 void *dst
= store
+ offset
;
1516 old_ip
[offset
/ sizeof(brw_compact_inst
)] = src_offset
/ sizeof(brw_inst
);
1517 compacted_counts
[src_offset
/ sizeof(brw_inst
)] = compacted_count
;
1519 brw_inst inst
= precompact(devinfo
, *src
);
1520 brw_inst saved
= inst
;
1522 if (brw_try_compact_instruction(devinfo
, dst
, &inst
)) {
1526 brw_inst uncompacted
;
1527 brw_uncompact_instruction(devinfo
, &uncompacted
, dst
);
1528 if (memcmp(&saved
, &uncompacted
, sizeof(uncompacted
))) {
1529 brw_debug_compact_uncompact(devinfo
, &saved
, &uncompacted
);
1533 offset
+= sizeof(brw_compact_inst
);
1535 /* All uncompacted instructions need to be aligned on G45. */
1536 if ((offset
& sizeof(brw_compact_inst
)) != 0 && devinfo
->is_g4x
){
1537 brw_compact_inst
*align
= store
+ offset
;
1538 memset(align
, 0, sizeof(*align
));
1539 brw_compact_inst_set_opcode(devinfo
, align
, BRW_OPCODE_NENOP
);
1540 brw_compact_inst_set_cmpt_control(devinfo
, align
, true);
1541 offset
+= sizeof(brw_compact_inst
);
1543 compacted_counts
[src_offset
/ sizeof(brw_inst
)] = compacted_count
;
1544 old_ip
[offset
/ sizeof(brw_compact_inst
)] = src_offset
/ sizeof(brw_inst
);
1546 dst
= store
+ offset
;
1549 /* If we didn't compact this intruction, we need to move it down into
1552 if (offset
!= src_offset
) {
1553 memmove(dst
, src
, sizeof(brw_inst
));
1555 offset
+= sizeof(brw_inst
);
1559 /* Add an entry for the ending offset of the program. This greatly
1560 * simplifies the linked list walk at the end of the function.
1562 old_ip
[offset
/ sizeof(brw_compact_inst
)] =
1563 (p
->next_insn_offset
- start_offset
) / sizeof(brw_inst
);
1565 /* Fix up control flow offsets. */
1566 p
->next_insn_offset
= start_offset
+ offset
;
1567 for (offset
= 0; offset
< p
->next_insn_offset
- start_offset
;
1568 offset
= next_offset(devinfo
, store
, offset
)) {
1569 brw_inst
*insn
= store
+ offset
;
1570 int this_old_ip
= old_ip
[offset
/ sizeof(brw_compact_inst
)];
1571 int this_compacted_count
= compacted_counts
[this_old_ip
];
1573 switch (brw_inst_opcode(devinfo
, insn
)) {
1574 case BRW_OPCODE_BREAK
:
1575 case BRW_OPCODE_CONTINUE
:
1576 case BRW_OPCODE_HALT
:
1577 if (devinfo
->gen
>= 6) {
1578 update_uip_jip(devinfo
, insn
, this_old_ip
, compacted_counts
);
1580 update_gen4_jump_count(devinfo
, insn
, this_old_ip
,
1586 case BRW_OPCODE_IFF
:
1587 case BRW_OPCODE_ELSE
:
1588 case BRW_OPCODE_ENDIF
:
1589 case BRW_OPCODE_WHILE
:
1590 if (devinfo
->gen
>= 7) {
1591 if (brw_inst_cmpt_control(devinfo
, insn
)) {
1592 brw_inst uncompacted
;
1593 brw_uncompact_instruction(devinfo
, &uncompacted
,
1594 (brw_compact_inst
*)insn
);
1596 update_uip_jip(devinfo
, &uncompacted
, this_old_ip
,
1599 bool ret
= brw_try_compact_instruction(devinfo
,
1600 (brw_compact_inst
*)insn
,
1602 assert(ret
); (void)ret
;
1604 update_uip_jip(devinfo
, insn
, this_old_ip
, compacted_counts
);
1606 } else if (devinfo
->gen
== 6) {
1607 assert(!brw_inst_cmpt_control(devinfo
, insn
));
1609 /* Jump Count is in units of compacted instructions on Gen6. */
1610 int jump_count_compacted
= brw_inst_gen6_jump_count(devinfo
, insn
);
1612 int target_old_ip
= this_old_ip
+ (jump_count_compacted
/ 2);
1613 int target_compacted_count
= compacted_counts
[target_old_ip
];
1614 jump_count_compacted
-= (target_compacted_count
- this_compacted_count
);
1615 brw_inst_set_gen6_jump_count(devinfo
, insn
, jump_count_compacted
);
1617 update_gen4_jump_count(devinfo
, insn
, this_old_ip
,
1622 case BRW_OPCODE_ADD
:
1623 /* Add instructions modifying the IP register use an immediate src1,
1624 * and Gens that use this cannot compact instructions with immediate
1627 if (brw_inst_cmpt_control(devinfo
, insn
))
1630 if (brw_inst_dst_reg_file(devinfo
, insn
) == BRW_ARCHITECTURE_REGISTER_FILE
&&
1631 brw_inst_dst_da_reg_nr(devinfo
, insn
) == BRW_ARF_IP
) {
1632 assert(brw_inst_src1_reg_file(devinfo
, insn
) == BRW_IMMEDIATE_VALUE
);
1635 int jump_compacted
= brw_inst_imm_d(devinfo
, insn
) >> shift
;
1637 int target_old_ip
= this_old_ip
+ (jump_compacted
/ 2);
1638 int target_compacted_count
= compacted_counts
[target_old_ip
];
1639 jump_compacted
-= (target_compacted_count
- this_compacted_count
);
1640 brw_inst_set_imm_ud(devinfo
, insn
, jump_compacted
<< shift
);
1646 /* p->nr_insn is counting the number of uncompacted instructions still, so
1647 * divide. We do want to be sure there's a valid instruction in any
1648 * alignment padding, so that the next compression pass (for the FS 8/16
1649 * compile passes) parses correctly.
1651 if (p
->next_insn_offset
& sizeof(brw_compact_inst
)) {
1652 brw_compact_inst
*align
= store
+ offset
;
1653 memset(align
, 0, sizeof(*align
));
1654 brw_compact_inst_set_opcode(devinfo
, align
, BRW_OPCODE_NOP
);
1655 brw_compact_inst_set_cmpt_control(devinfo
, align
, true);
1656 p
->next_insn_offset
+= sizeof(brw_compact_inst
);
1658 p
->nr_insn
= p
->next_insn_offset
/ sizeof(brw_inst
);
1660 /* Update the instruction offsets for each group. */
1664 foreach_list_typed(struct inst_group
, group
, link
, &disasm
->group_list
) {
1665 while (start_offset
+ old_ip
[offset
/ sizeof(brw_compact_inst
)] *
1666 sizeof(brw_inst
) != group
->offset
) {
1667 assert(start_offset
+ old_ip
[offset
/ sizeof(brw_compact_inst
)] *
1668 sizeof(brw_inst
) < group
->offset
);
1669 offset
= next_offset(devinfo
, store
, offset
);
1672 group
->offset
= start_offset
+ offset
;
1674 offset
= next_offset(devinfo
, store
, offset
);