5 * Copyright (c) 2013 Connor Abbott (connor@abbott.cx)
6 * Copyright (c) 2018 Alyssa Rosenzweig (alyssa@rosenzweig.io)
8 * Permission is hereby granted, free of charge, to any person obtaining a copy
9 * of this software and associated documentation files (the "Software"), to deal
10 * in the Software without restriction, including without limitation the rights
11 * to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
12 * copies of the Software, and to permit persons to whom the Software is
13 * furnished to do so, subject to the following conditions:
15 * The above copyright notice and this permission notice shall be included in
16 * all copies or substantial portions of the Software.
18 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
19 * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
20 * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
21 * AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
22 * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
23 * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN
35 #include "midgard-parse.h"
36 #include "midgard_ops.h"
37 #include "disassemble.h"
39 #include "util/half_float.h"
40 #include "util/u_math.h"
42 #define DEFINE_CASE(define, str) case define: { printf(str); break; }
44 static unsigned *midg_tags
;
45 static bool is_instruction_int
= false;
49 static struct midgard_disasm_stats midg_stats
;
51 /* Prints a short form of the tag for branching, the minimum needed to be
52 * legible and unambiguous */
55 print_tag_short(unsigned tag
)
57 switch (midgard_word_types
[tag
]) {
58 case midgard_word_type_texture
:
59 printf("tex/%X", tag
);
62 case midgard_word_type_load_store
:
66 case midgard_word_type_alu
:
67 printf("alu%u/%X", midgard_word_size
[tag
], tag
);
71 printf("%s%X", (tag
> 0) ? "" : "unk", tag
);
77 print_alu_opcode(midgard_alu_op op
)
81 if (alu_opcode_props
[op
].name
) {
82 printf("%s", alu_opcode_props
[op
].name
);
84 int_op
= midgard_is_integer_op(op
);
86 printf("alu_op_%02X", op
);
88 /* For constant analysis */
89 is_instruction_int
= int_op
;
93 print_ld_st_opcode(midgard_load_store_op op
)
95 if (load_store_opcode_props
[op
].name
)
96 printf("%s", load_store_opcode_props
[op
].name
);
98 printf("ldst_op_%02X", op
);
101 static bool is_embedded_constant_half
= false;
102 static bool is_embedded_constant_int
= false;
105 prefix_for_bits(unsigned bits
)
119 /* For static analysis to ensure all registers are written at least once before
120 * use along the source code path (TODO: does this break done for complex CF?)
123 uint16_t midg_ever_written
= 0;
126 print_reg(unsigned reg
, unsigned bits
)
128 /* Perform basic static analysis for expanding constants correctly */
131 is_embedded_constant_int
= is_instruction_int
;
132 is_embedded_constant_half
= (bits
< 32);
135 unsigned uniform_reg
= 23 - reg
;
136 bool is_uniform
= false;
138 /* For r8-r15, it could be a work or uniform. We distinguish based on
139 * the fact work registers are ALWAYS written before use, but uniform
140 * registers are NEVER written before use. */
142 if ((reg
>= 8 && reg
< 16) && !(midg_ever_written
& (1 << reg
)))
145 /* r16-r23 are always uniform */
147 if (reg
>= 16 && reg
<= 23)
150 /* Update the uniform count appropriately */
153 midg_stats
.uniform_count
=
154 MAX2(uniform_reg
+ 1, midg_stats
.uniform_count
);
156 char prefix
= prefix_for_bits(bits
);
164 static char *outmod_names_float
[4] = {
171 static char *outmod_names_int
[4] = {
178 static char *srcmod_names_int
[4] = {
186 print_outmod(unsigned outmod
, bool is_int
)
188 printf("%s", is_int
? outmod_names_int
[outmod
] :
189 outmod_names_float
[outmod
]);
193 print_quad_word(uint32_t *words
, unsigned tabs
)
197 for (i
= 0; i
< 4; i
++)
198 printf("0x%08X%s ", words
[i
], i
== 3 ? "" : ",");
203 static const char components
[16] = "xyzwefghijklmnop";
205 /* Helper to print 4 chars of a swizzle */
207 print_swizzle_helper(unsigned swizzle
, bool upper
)
209 for (unsigned i
= 0; i
< 4; ++i
) {
210 unsigned c
= (swizzle
>> (i
* 2)) & 3;
212 printf("%c", components
[c
]);
216 /* Helper to print 8 chars of a swizzle, duplicating over */
218 print_swizzle_helper_8(unsigned swizzle
, bool upper
)
220 for (unsigned i
= 0; i
< 4; ++i
) {
221 unsigned c
= (swizzle
>> (i
* 2)) & 3;
224 printf("%c%c", components
[c
], components
[c
+1]);
229 print_swizzle_vec16(unsigned swizzle
, bool rep_high
, bool rep_low
,
230 midgard_dest_override override
)
234 if (override
== midgard_dest_override_upper
) {
236 printf(" /* rep_high */ ");
238 printf(" /* rep_low */ ");
240 if (!rep_high
&& rep_low
)
241 print_swizzle_helper_8(swizzle
, true);
243 print_swizzle_helper_8(swizzle
, false);
245 print_swizzle_helper_8(swizzle
, rep_high
& 1);
246 print_swizzle_helper_8(swizzle
, !(rep_low
& 1));
251 print_swizzle_vec8(unsigned swizzle
, bool rep_high
, bool rep_low
)
255 print_swizzle_helper(swizzle
, rep_high
& 1);
256 print_swizzle_helper(swizzle
, !(rep_low
& 1));
260 print_swizzle_vec4(unsigned swizzle
, bool rep_high
, bool rep_low
)
263 printf(" /* rep_high */ ");
265 printf(" /* rep_low */ ");
267 if (swizzle
== 0xE4) return; /* xyzw */
270 print_swizzle_helper(swizzle
, 0);
273 print_swizzle_vec2(unsigned swizzle
, bool rep_high
, bool rep_low
)
276 printf(" /* rep_high */ ");
278 printf(" /* rep_low */ ");
280 if (swizzle
== 0xE4) return; /* XY */
284 for (unsigned i
= 0; i
< 4; i
+= 2) {
285 unsigned a
= (swizzle
>> (i
* 2)) & 3;
286 unsigned b
= (swizzle
>> ((i
+1) * 2)) & 3;
288 /* Normally we're adjacent, but if there's an issue, don't make
292 printf("[%c%c]", components
[a
], components
[b
]);
294 printf("%c", components
[a
>> 1]);
295 else if (b
== (a
+ 1))
296 printf("%c", "XY"[a
>> 1]);
298 printf("[%c%c]", components
[a
], components
[b
]);
303 bits_for_mode(midgard_reg_mode mode
)
306 case midgard_reg_mode_8
:
308 case midgard_reg_mode_16
:
310 case midgard_reg_mode_32
:
312 case midgard_reg_mode_64
:
315 unreachable("Invalid reg mode");
321 bits_for_mode_halved(midgard_reg_mode mode
, bool half
)
323 unsigned bits
= bits_for_mode(mode
);
332 print_vector_src(unsigned src_binary
,
333 midgard_reg_mode mode
, unsigned reg
,
334 midgard_dest_override override
, bool is_int
)
336 midgard_vector_alu_src
*src
= (midgard_vector_alu_src
*)&src_binary
;
338 /* Modifiers change meaning depending on the op's context */
340 midgard_int_mod int_mod
= src
->mod
;
343 printf("%s", srcmod_names_int
[int_mod
]);
345 if (src
->mod
& MIDGARD_FLOAT_MOD_NEG
)
348 if (src
->mod
& MIDGARD_FLOAT_MOD_ABS
)
353 unsigned bits
= bits_for_mode_halved(mode
, src
->half
);
354 print_reg(reg
, bits
);
358 print_swizzle_vec8(src
->swizzle
, src
->rep_high
, src
->rep_low
);
360 print_swizzle_vec16(src
->swizzle
, src
->rep_high
, src
->rep_low
, override
);
362 print_swizzle_vec4(src
->swizzle
, src
->rep_high
, src
->rep_low
);
364 print_swizzle_vec2(src
->swizzle
, src
->rep_high
, src
->rep_low
);
366 /* Since we wrapped with a function-looking thing */
368 if (is_int
&& int_mod
== midgard_int_shift
)
369 printf(") << %u", bits
);
370 else if ((is_int
&& (int_mod
!= midgard_int_normal
))
371 || (!is_int
&& src
->mod
& MIDGARD_FLOAT_MOD_ABS
))
376 decode_vector_imm(unsigned src2_reg
, unsigned imm
)
379 ret
= src2_reg
<< 11;
380 ret
|= (imm
& 0x7) << 8;
381 ret
|= (imm
>> 3) & 0xFF;
386 print_immediate(uint16_t imm
)
388 if (is_instruction_int
)
391 printf("#%g", _mesa_half_to_float(imm
));
395 update_dest(unsigned reg
)
397 /* We should record writes as marking this as a work register. Store
398 * the max register in work_count; we'll add one at the end */
401 midg_stats
.work_count
= MAX2(reg
, midg_stats
.work_count
);
402 midg_ever_written
|= (1 << reg
);
407 print_dest(unsigned reg
, midgard_reg_mode mode
, midgard_dest_override override
)
409 /* Depending on the mode and override, we determine the type of
410 * destination addressed. Absent an override, we address just the
411 * type of the operation itself */
413 unsigned bits
= bits_for_mode(mode
);
415 if (override
!= midgard_dest_override_none
)
419 print_reg(reg
, bits
);
423 print_mask_vec16(uint8_t mask
, midgard_dest_override override
)
427 for (unsigned i
= 0; i
< 8; i
++) {
431 components
[i
*2 + 1]);
435 /* For 16-bit+ masks, we read off from the 8-bit mask field. For 16-bit (vec8),
436 * it's just one bit per channel, easy peasy. For 32-bit (vec4), it's one bit
437 * per channel with one duplicate bit in the middle. For 64-bit (vec2), it's
438 * one-bit per channel with _3_ duplicate bits in the middle. Basically, just
439 * subdividing the 128-bit word in 16-bit increments. For 64-bit, we uppercase
440 * the mask to make it obvious what happened */
443 print_mask(uint8_t mask
, unsigned bits
, midgard_dest_override override
)
446 print_mask_vec16(mask
, override
);
450 /* Skip 'complete' masks */
452 if (override
== midgard_dest_override_none
) {
453 if (bits
>= 32 && mask
== 0xFF) return;
458 else if (mask
== 0xF0) {
467 unsigned skip
= (bits
/ 16);
468 bool uppercase
= bits
> 32;
469 bool tripped
= false;
471 /* To apply an upper destination override, we "shift" the alphabet.
472 * E.g. with an upper override on 32-bit, instead of xyzw, print efgh.
473 * For upper 16-bit, instead of xyzwefgh, print ijklmnop */
475 const char *alphabet
= components
;
477 if (override
== midgard_dest_override_upper
) {
478 unsigned components
= 128 / bits
;
479 alphabet
+= components
;
482 for (unsigned i
= 0; i
< 8; i
+= skip
) {
483 bool a
= (mask
& (1 << i
)) != 0;
485 for (unsigned j
= 1; j
< skip
; ++j
) {
486 bool dupe
= (mask
& (1 << (i
+ j
))) != 0;
487 tripped
|= (dupe
!= a
);
491 char c
= alphabet
[i
/ skip
];
501 printf(" /* %X */", mask
);
504 /* Prints the 4-bit masks found in texture and load/store ops, as opposed to
505 * the 8-bit masks found in (vector) ALU ops. Supports texture-style 16-bit
506 * mode as well, but not load/store-style 16-bit mode. */
509 print_mask_4(unsigned mask
, bool upper
)
520 for (unsigned i
= 0; i
< 4; ++i
) {
521 bool a
= (mask
& (1 << i
)) != 0;
523 printf("%c", components
[i
+ (upper
? 4 : 0)]);
528 print_vector_field(const char *name
, uint16_t *words
, uint16_t reg_word
,
531 midgard_reg_info
*reg_info
= (midgard_reg_info
*)®_word
;
532 midgard_vector_alu
*alu_field
= (midgard_vector_alu
*) words
;
533 midgard_reg_mode mode
= alu_field
->reg_mode
;
534 unsigned override
= alu_field
->dest_override
;
536 /* For now, prefix instruction names with their unit, until we
537 * understand how this works on a deeper level */
540 print_alu_opcode(alu_field
->op
);
542 /* Postfix with the size to disambiguate if necessary */
543 char postfix
= prefix_for_bits(bits_for_mode(mode
));
544 bool size_ambiguous
= override
!= midgard_dest_override_none
;
547 printf("%c", postfix
? postfix
: 'r');
549 /* Print the outmod, if there is one */
550 print_outmod(alu_field
->outmod
,
551 midgard_is_integer_out_op(alu_field
->op
));
555 /* Mask denoting status of 8-lanes */
556 uint8_t mask
= alu_field
->mask
;
558 /* First, print the destination */
559 print_dest(reg_info
->out_reg
, mode
, alu_field
->dest_override
);
561 if (override
!= midgard_dest_override_none
) {
562 bool modeable
= (mode
!= midgard_reg_mode_8
);
563 bool known
= override
!= 0x3; /* Unused value */
565 if (!(modeable
&& known
))
566 printf("/* do%u */ ", override
);
569 print_mask(mask
, bits_for_mode(mode
), override
);
573 bool is_int
= midgard_is_integer_op(alu_field
->op
);
574 print_vector_src(alu_field
->src1
, mode
, reg_info
->src1_reg
, override
, is_int
);
578 if (reg_info
->src2_imm
) {
579 uint16_t imm
= decode_vector_imm(reg_info
->src2_reg
, alu_field
->src2
>> 2);
580 print_immediate(imm
);
582 print_vector_src(alu_field
->src2
, mode
,
583 reg_info
->src2_reg
, override
, is_int
);
586 midg_stats
.instruction_count
++;
591 print_scalar_src(unsigned src_binary
, unsigned reg
)
593 midgard_scalar_alu_src
*src
= (midgard_scalar_alu_src
*)&src_binary
;
601 print_reg(reg
, src
->full
? 32 : 16);
603 unsigned c
= src
->component
;
606 assert((c
& 1) == 0);
610 printf(".%c", components
[c
]);
618 decode_scalar_imm(unsigned src2_reg
, unsigned imm
)
621 ret
= src2_reg
<< 11;
622 ret
|= (imm
& 3) << 9;
623 ret
|= (imm
& 4) << 6;
624 ret
|= (imm
& 0x38) << 2;
630 print_scalar_field(const char *name
, uint16_t *words
, uint16_t reg_word
,
633 midgard_reg_info
*reg_info
= (midgard_reg_info
*)®_word
;
634 midgard_scalar_alu
*alu_field
= (midgard_scalar_alu
*) words
;
636 if (alu_field
->unknown
)
637 printf("scalar ALU unknown bit set\n");
640 print_alu_opcode(alu_field
->op
);
641 print_outmod(alu_field
->outmod
,
642 midgard_is_integer_out_op(alu_field
->op
));
645 bool full
= alu_field
->output_full
;
646 update_dest(reg_info
->out_reg
);
647 print_reg(reg_info
->out_reg
, full
? 32 : 16);
648 unsigned c
= alu_field
->output_component
;
651 assert((c
& 1) == 0);
655 printf(".%c, ", components
[c
]);
657 print_scalar_src(alu_field
->src1
, reg_info
->src1_reg
);
661 if (reg_info
->src2_imm
) {
662 uint16_t imm
= decode_scalar_imm(reg_info
->src2_reg
,
664 print_immediate(imm
);
666 print_scalar_src(alu_field
->src2
, reg_info
->src2_reg
);
668 midg_stats
.instruction_count
++;
673 print_branch_op(unsigned op
)
676 case midgard_jmp_writeout_op_branch_uncond
:
680 case midgard_jmp_writeout_op_branch_cond
:
684 case midgard_jmp_writeout_op_writeout
:
688 case midgard_jmp_writeout_op_tilebuffer_pending
:
689 printf("tilebuffer.");
692 case midgard_jmp_writeout_op_discard
:
697 printf("unk%u.", op
);
703 print_branch_cond(int cond
)
706 case midgard_condition_write0
:
710 case midgard_condition_false
:
714 case midgard_condition_true
:
718 case midgard_condition_always
:
723 printf("unk%X", cond
);
729 print_compact_branch_writeout_field(uint16_t word
)
731 midgard_jmp_writeout_op op
= word
& 0x7;
734 case midgard_jmp_writeout_op_branch_uncond
: {
735 midgard_branch_uncond br_uncond
;
736 memcpy((char *) &br_uncond
, (char *) &word
, sizeof(br_uncond
));
737 printf("br.uncond ");
739 if (br_uncond
.unknown
!= 1)
740 printf("unknown:%u, ", br_uncond
.unknown
);
742 if (br_uncond
.offset
>= 0)
745 printf("%d -> ", br_uncond
.offset
);
746 print_tag_short(br_uncond
.dest_tag
);
752 case midgard_jmp_writeout_op_branch_cond
:
753 case midgard_jmp_writeout_op_writeout
:
754 case midgard_jmp_writeout_op_discard
:
756 midgard_branch_cond br_cond
;
757 memcpy((char *) &br_cond
, (char *) &word
, sizeof(br_cond
));
761 print_branch_op(br_cond
.op
);
762 print_branch_cond(br_cond
.cond
);
766 if (br_cond
.offset
>= 0)
769 printf("%d -> ", br_cond
.offset
);
770 print_tag_short(br_cond
.dest_tag
);
777 midg_stats
.instruction_count
++;
781 print_extended_branch_writeout_field(uint8_t *words
, unsigned next
)
783 midgard_branch_extended br
;
784 memcpy((char *) &br
, (char *) words
, sizeof(br
));
788 print_branch_op(br
.op
);
790 /* Condition codes are a LUT in the general case, but simply repeated 8 times for single-channel conditions.. Check this. */
792 bool single_channel
= true;
794 for (unsigned i
= 0; i
< 16; i
+= 2) {
795 single_channel
&= (((br
.cond
>> i
) & 0x3) == (br
.cond
& 0x3));
799 print_branch_cond(br
.cond
& 0x3);
801 printf("lut%X", br
.cond
);
804 printf(".unknown%u", br
.unknown
);
811 printf("%d -> ", br
.offset
);
812 print_tag_short(br
.dest_tag
);
815 unsigned I
= next
+ br
.offset
* 4;
817 if (midg_tags
[I
] && midg_tags
[I
] != br
.dest_tag
) {
818 printf("\t/* XXX TAG ERROR: jumping to ");
819 print_tag_short(br
.dest_tag
);
820 printf(" but tagged ");
821 print_tag_short(midg_tags
[I
]);
825 midg_tags
[I
] = br
.dest_tag
;
827 midg_stats
.instruction_count
++;
831 num_alu_fields_enabled(uint32_t control_word
)
835 if ((control_word
>> 17) & 1)
838 if ((control_word
>> 19) & 1)
841 if ((control_word
>> 21) & 1)
844 if ((control_word
>> 23) & 1)
847 if ((control_word
>> 25) & 1)
854 float_bitcast(uint32_t integer
)
866 print_alu_word(uint32_t *words
, unsigned num_quad_words
,
867 unsigned tabs
, unsigned next
)
869 uint32_t control_word
= words
[0];
870 uint16_t *beginning_ptr
= (uint16_t *)(words
+ 1);
871 unsigned num_fields
= num_alu_fields_enabled(control_word
);
872 uint16_t *word_ptr
= beginning_ptr
+ num_fields
;
873 unsigned num_words
= 2 + num_fields
;
875 if ((control_word
>> 16) & 1)
876 printf("unknown bit 16 enabled\n");
878 if ((control_word
>> 17) & 1) {
879 print_vector_field("vmul", word_ptr
, *beginning_ptr
, tabs
);
885 if ((control_word
>> 18) & 1)
886 printf("unknown bit 18 enabled\n");
888 if ((control_word
>> 19) & 1) {
889 print_scalar_field("sadd", word_ptr
, *beginning_ptr
, tabs
);
895 if ((control_word
>> 20) & 1)
896 printf("unknown bit 20 enabled\n");
898 if ((control_word
>> 21) & 1) {
899 print_vector_field("vadd", word_ptr
, *beginning_ptr
, tabs
);
905 if ((control_word
>> 22) & 1)
906 printf("unknown bit 22 enabled\n");
908 if ((control_word
>> 23) & 1) {
909 print_scalar_field("smul", word_ptr
, *beginning_ptr
, tabs
);
915 if ((control_word
>> 24) & 1)
916 printf("unknown bit 24 enabled\n");
918 if ((control_word
>> 25) & 1) {
919 print_vector_field("lut", word_ptr
, *beginning_ptr
, tabs
);
924 if ((control_word
>> 26) & 1) {
925 print_compact_branch_writeout_field(*word_ptr
);
930 if ((control_word
>> 27) & 1) {
931 print_extended_branch_writeout_field((uint8_t *) word_ptr
, next
);
936 if (num_quad_words
> (num_words
+ 7) / 8) {
937 assert(num_quad_words
== (num_words
+ 15) / 8);
938 //Assume that the extra quadword is constants
939 void *consts
= words
+ (4 * num_quad_words
- 4);
941 if (is_embedded_constant_int
) {
942 if (is_embedded_constant_half
) {
943 int16_t *sconsts
= (int16_t *) consts
;
944 printf("sconstants %d, %d, %d, %d\n",
950 uint32_t *iconsts
= (uint32_t *) consts
;
951 printf("iconstants 0x%X, 0x%X, 0x%X, 0x%X\n",
958 if (is_embedded_constant_half
) {
959 uint16_t *hconsts
= (uint16_t *) consts
;
960 printf("hconstants %g, %g, %g, %g\n",
961 _mesa_half_to_float(hconsts
[0]),
962 _mesa_half_to_float(hconsts
[1]),
963 _mesa_half_to_float(hconsts
[2]),
964 _mesa_half_to_float(hconsts
[3]));
966 uint32_t *fconsts
= (uint32_t *) consts
;
967 printf("fconstants %g, %g, %g, %g\n",
968 float_bitcast(fconsts
[0]),
969 float_bitcast(fconsts
[1]),
970 float_bitcast(fconsts
[2]),
971 float_bitcast(fconsts
[3]));
979 print_varying_parameters(midgard_load_store_word
*word
)
981 midgard_varying_parameter param
;
982 unsigned v
= word
->varying_parameters
;
983 memcpy(¶m
, &v
, sizeof(param
));
985 if (param
.is_varying
) {
986 /* If a varying, there are qualifiers */
990 if (param
.interpolation
!= midgard_interp_default
) {
991 if (param
.interpolation
== midgard_interp_centroid
)
994 printf(".interp%d", param
.interpolation
);
997 if (param
.modifier
!= midgard_varying_mod_none
) {
998 if (param
.modifier
== midgard_varying_mod_perspective_w
)
999 printf(".perspectivew");
1000 else if (param
.modifier
== midgard_varying_mod_perspective_z
)
1001 printf(".perspectivez");
1003 printf(".mod%d", param
.modifier
);
1005 } else if (param
.flat
|| param
.interpolation
|| param
.modifier
) {
1006 printf(" /* is_varying not set but varying metadata attached */");
1009 if (param
.zero0
|| param
.zero1
|| param
.zero2
)
1010 printf(" /* zero tripped, %u %u %u */ ", param
.zero0
, param
.zero1
, param
.zero2
);
1014 is_op_varying(unsigned op
)
1017 case midgard_op_st_vary_16
:
1018 case midgard_op_st_vary_32
:
1019 case midgard_op_st_vary_32i
:
1020 case midgard_op_st_vary_32u
:
1021 case midgard_op_ld_vary_16
:
1022 case midgard_op_ld_vary_32
:
1023 case midgard_op_ld_vary_32i
:
1024 case midgard_op_ld_vary_32u
:
1032 is_op_attribute(unsigned op
)
1035 case midgard_op_ld_attr_16
:
1036 case midgard_op_ld_attr_32
:
1037 case midgard_op_ld_attr_32i
:
1038 case midgard_op_ld_attr_32u
:
1046 print_load_store_arg(uint8_t arg
, unsigned index
)
1048 /* Try to interpret as a register */
1049 midgard_ldst_register_select sel
;
1050 memcpy(&sel
, &arg
, sizeof(arg
));
1052 /* If unknown is set, we're not sure what this is or how to
1053 * interpret it. But if it's zero, we get it. */
1056 printf("0x%02X", arg
);
1060 unsigned reg
= REGISTER_LDST_BASE
+ sel
.select
;
1061 char comp
= components
[sel
.component
];
1063 printf("r%u.%c", reg
, comp
);
1065 /* Only print a shift if it's non-zero. Shifts only make sense for the
1066 * second index. For the first, we're not sure what it means yet */
1070 printf(" << %u", sel
.shift
);
1072 printf(" /* %X */", sel
.shift
);
1077 update_stats(signed *stat
, unsigned address
)
1080 *stat
= MAX2(*stat
, address
+ 1);
1084 print_load_store_instr(uint64_t data
,
1087 midgard_load_store_word
*word
= (midgard_load_store_word
*) &data
;
1089 print_ld_st_opcode(word
->op
);
1091 unsigned address
= word
->address
;
1093 if (is_op_varying(word
->op
)) {
1094 print_varying_parameters(word
);
1096 /* Do some analysis: check if direct cacess */
1098 if ((word
->arg_2
== 0x1E) && midg_stats
.varying_count
>= 0)
1099 update_stats(&midg_stats
.varying_count
, address
);
1101 midg_stats
.varying_count
= -16;
1102 } else if (is_op_attribute(word
->op
)) {
1103 if ((word
->arg_2
== 0x1E) && midg_stats
.attribute_count
>= 0)
1104 update_stats(&midg_stats
.attribute_count
, address
);
1106 midg_stats
.attribute_count
= -16;
1109 printf(" r%u", word
->reg
);
1110 print_mask_4(word
->mask
, false);
1112 if (!OP_IS_STORE(word
->op
))
1113 update_dest(word
->reg
);
1115 bool is_ubo
= OP_IS_UBO_READ(word
->op
);
1118 /* UBOs use their own addressing scheme */
1120 int lo
= word
->varying_parameters
>> 7;
1121 int hi
= word
->address
;
1123 /* TODO: Combine fields logically */
1124 address
= (hi
<< 3) | lo
;
1127 printf(", %u", address
);
1129 print_swizzle_vec4(word
->swizzle
, false, false);
1134 printf("ubo%u", word
->arg_1
);
1135 update_stats(&midg_stats
.uniform_buffer_count
, word
->arg_1
);
1137 print_load_store_arg(word
->arg_1
, 0);
1140 print_load_store_arg(word
->arg_2
, 1);
1141 printf(" /* %X */\n", word
->varying_parameters
);
1143 midg_stats
.instruction_count
++;
1147 print_load_store_word(uint32_t *word
, unsigned tabs
)
1149 midgard_load_store
*load_store
= (midgard_load_store
*) word
;
1151 if (load_store
->word1
!= 3) {
1152 print_load_store_instr(load_store
->word1
, tabs
);
1155 if (load_store
->word2
!= 3) {
1156 print_load_store_instr(load_store
->word2
, tabs
);
1161 print_texture_reg_select(uint8_t u
)
1163 midgard_tex_register_select sel
;
1164 memcpy(&sel
, &u
, sizeof(u
));
1169 printf("r%u", REG_TEX_BASE
+ sel
.select
);
1171 unsigned component
= sel
.component
;
1173 /* Use the upper half in half-reg mode */
1179 printf(".%c", components
[component
]);
1181 assert(sel
.zero
== 0);
1185 print_texture_format(int format
)
1187 /* Act like a modifier */
1191 DEFINE_CASE(MALI_TEX_1D
, "1d");
1192 DEFINE_CASE(MALI_TEX_2D
, "2d");
1193 DEFINE_CASE(MALI_TEX_3D
, "3d");
1194 DEFINE_CASE(MALI_TEX_CUBE
, "cube");
1197 unreachable("Bad format");
1202 midgard_op_has_helpers(unsigned op
, bool gather
)
1208 case TEXTURE_OP_NORMAL
:
1209 case TEXTURE_OP_DFDX
:
1210 case TEXTURE_OP_DFDY
:
1218 print_texture_op(unsigned op
, bool gather
)
1220 /* Act like a bare name, like ESSL functions */
1223 printf("textureGather");
1225 unsigned component
= op
>> 4;
1226 unsigned bottom
= op
& 0xF;
1229 printf("_unk%u", bottom
);
1231 printf(".%c", components
[component
]);
1236 DEFINE_CASE(TEXTURE_OP_NORMAL
, "texture");
1237 DEFINE_CASE(TEXTURE_OP_LOD
, "textureLod");
1238 DEFINE_CASE(TEXTURE_OP_TEXEL_FETCH
, "texelFetch");
1239 DEFINE_CASE(TEXTURE_OP_DFDX
, "dFdx");
1240 DEFINE_CASE(TEXTURE_OP_DFDY
, "dFdy");
1243 printf("tex_%X", op
);
1249 texture_op_takes_bias(unsigned op
)
1251 return op
== TEXTURE_OP_NORMAL
;
1255 sampler_type_name(enum mali_sampler_type t
)
1258 case MALI_SAMPLER_FLOAT
:
1260 case MALI_SAMPLER_UNSIGNED
:
1262 case MALI_SAMPLER_SIGNED
:
1273 print_texture_word(uint32_t *word
, unsigned tabs
)
1275 midgard_texture_word
*texture
= (midgard_texture_word
*) word
;
1277 midg_stats
.helper_invocations
|=
1278 midgard_op_has_helpers(texture
->op
, texture
->is_gather
);
1280 /* Broad category of texture operation in question */
1281 print_texture_op(texture
->op
, texture
->is_gather
);
1283 /* Specific format in question */
1284 print_texture_format(texture
->format
);
1286 /* Instruction "modifiers" parallel the ALU instructions. */
1288 if (texture
->shadow
)
1297 /* Output modifiers are always interpreted floatly */
1298 print_outmod(texture
->outmod
, false);
1300 printf(" %sr%d", texture
->out_full
? "" : "h",
1301 REG_TEX_BASE
+ texture
->out_reg_select
);
1302 print_mask_4(texture
->mask
, texture
->out_upper
);
1303 assert(!(texture
->out_full
&& texture
->out_upper
));
1306 /* Depending on whether we read from textures directly or indirectly,
1307 * we may be able to update our analysis */
1309 if (texture
->texture_register
) {
1311 print_texture_reg_select(texture
->texture_handle
);
1314 /* Indirect, tut tut */
1315 midg_stats
.texture_count
= -16;
1317 printf("texture%u, ", texture
->texture_handle
);
1318 update_stats(&midg_stats
.texture_count
, texture
->texture_handle
);
1321 /* Print the type, GL style */
1322 printf("%csampler", sampler_type_name(texture
->sampler_type
));
1324 if (texture
->sampler_register
) {
1326 print_texture_reg_select(texture
->sampler_handle
);
1329 midg_stats
.sampler_count
= -16;
1331 printf("%u", texture
->sampler_handle
);
1332 update_stats(&midg_stats
.sampler_count
, texture
->sampler_handle
);
1335 print_swizzle_vec4(texture
->swizzle
, false, false);
1336 printf(", %sr%d", texture
->in_reg_full
? "" : "h", REG_TEX_BASE
+ texture
->in_reg_select
);
1337 assert(!(texture
->in_reg_full
&& texture
->in_reg_upper
));
1339 /* TODO: integrate with swizzle */
1340 if (texture
->in_reg_upper
)
1343 print_swizzle_vec4(texture
->in_reg_swizzle
, false, false);
1345 /* There is *always* an offset attached. Of
1346 * course, that offset is just immediate #0 for a
1347 * GLES call that doesn't take an offset. If there
1348 * is a non-negative non-zero offset, this is
1349 * specified in immediate offset mode, with the
1350 * values in the offset_* fields as immediates. If
1351 * this is a negative offset, we instead switch to
1352 * a register offset mode, where the offset_*
1353 * fields become register triplets */
1355 if (texture
->offset_register
) {
1358 bool full
= texture
->offset_x
& 1;
1359 bool select
= texture
->offset_x
& 2;
1360 bool upper
= texture
->offset_x
& 4;
1362 printf("%sr%d", full
? "" : "h", REG_TEX_BASE
+ select
);
1363 assert(!(texture
->out_full
&& texture
->out_upper
));
1365 /* TODO: integrate with swizzle */
1369 /* The less questions you ask, the better. */
1371 unsigned swizzle_lo
, swizzle_hi
;
1372 unsigned orig_y
= texture
->offset_y
;
1373 unsigned orig_z
= texture
->offset_z
;
1375 memcpy(&swizzle_lo
, &orig_y
, sizeof(unsigned));
1376 memcpy(&swizzle_hi
, &orig_z
, sizeof(unsigned));
1378 /* Duplicate hi swizzle over */
1379 assert(swizzle_hi
< 4);
1380 swizzle_hi
= (swizzle_hi
<< 2) | swizzle_hi
;
1382 unsigned swiz
= (swizzle_lo
<< 4) | swizzle_hi
;
1383 unsigned reversed
= util_bitreverse(swiz
) >> 24;
1384 print_swizzle_vec4(reversed
, false, false);
1387 } else if (texture
->offset_x
|| texture
->offset_y
|| texture
->offset_z
) {
1388 /* Only select ops allow negative immediate offsets, verify */
1390 bool neg_x
= texture
->offset_x
< 0;
1391 bool neg_y
= texture
->offset_y
< 0;
1392 bool neg_z
= texture
->offset_z
< 0;
1393 bool any_neg
= neg_x
|| neg_y
|| neg_z
;
1395 if (any_neg
&& texture
->op
!= TEXTURE_OP_TEXEL_FETCH
)
1396 printf("/* invalid negative */ ");
1398 /* Regardless, just print the immediate offset */
1400 printf(" + <%d, %d, %d>, ",
1408 char lod_operand
= texture_op_takes_bias(texture
->op
) ? '+' : '=';
1410 if (texture
->lod_register
) {
1411 printf("lod %c ", lod_operand
);
1412 print_texture_reg_select(texture
->bias
);
1415 if (texture
->bias_int
)
1416 printf(" /* bias_int = 0x%X */", texture
->bias_int
);
1417 } else if (texture
->op
== TEXTURE_OP_TEXEL_FETCH
) {
1418 /* For texel fetch, the int LOD is in the fractional place and
1419 * there is no fraction / possibility of bias. We *always* have
1420 * an explicit LOD, even if it's zero. */
1422 if (texture
->bias_int
)
1423 printf(" /* bias_int = 0x%X */ ", texture
->bias_int
);
1425 printf("lod = %u, ", texture
->bias
);
1426 } else if (texture
->bias
|| texture
->bias_int
) {
1427 signed bias_int
= texture
->bias_int
;
1428 float bias_frac
= texture
->bias
/ 256.0f
;
1429 float bias
= bias_int
+ bias_frac
;
1431 bool is_bias
= texture_op_takes_bias(texture
->op
);
1432 char sign
= (bias
>= 0.0) ? '+' : '-';
1433 char operand
= is_bias
? sign
: '=';
1435 printf("lod %c %f, ", operand
, fabsf(bias
));
1440 /* While not zero in general, for these simple instructions the
1441 * following unknowns are zero, so we don't include them */
1443 if (texture
->unknown4
||
1444 texture
->unknownA
||
1445 texture
->unknown8
) {
1446 printf("// unknown4 = 0x%x\n", texture
->unknown4
);
1447 printf("// unknownA = 0x%x\n", texture
->unknownA
);
1448 printf("// unknown8 = 0x%x\n", texture
->unknown8
);
1451 midg_stats
.instruction_count
++;
1454 struct midgard_disasm_stats
1455 disassemble_midgard(uint8_t *code
, size_t size
, unsigned gpu_id
)
1457 uint32_t *words
= (uint32_t *) code
;
1458 unsigned num_words
= size
/ 4;
1461 bool prefetch_flag
= false;
1463 int last_next_tag
= -1;
1467 midg_tags
= calloc(sizeof(midg_tags
[0]), num_words
);
1469 /* Stats for shader-db */
1470 memset(&midg_stats
, 0, sizeof(midg_stats
));
1471 midg_ever_written
= 0;
1473 while (i
< num_words
) {
1474 unsigned tag
= words
[i
] & 0xF;
1475 unsigned next_tag
= (words
[i
] >> 4) & 0xF;
1476 unsigned num_quad_words
= midgard_word_size
[tag
];
1478 if (midg_tags
[i
] && midg_tags
[i
] != tag
) {
1479 printf("\t/* XXX: TAG ERROR branch, got ");
1480 print_tag_short(tag
);
1481 printf(" expected ");
1482 print_tag_short(midg_tags
[i
]);
1489 if (last_next_tag
> 1) {
1490 if (last_next_tag
!= tag
) {
1491 printf("\t/* XXX: TAG ERROR sequence, got ");
1492 print_tag_short(tag
);
1493 printf(" expected ");
1494 print_tag_short(last_next_tag
);
1498 /* TODO: Check ALU case */
1501 last_next_tag
= next_tag
;
1503 switch (midgard_word_types
[tag
]) {
1504 case midgard_word_type_texture
:
1505 print_texture_word(&words
[i
], tabs
);
1508 case midgard_word_type_load_store
:
1509 print_load_store_word(&words
[i
], tabs
);
1512 case midgard_word_type_alu
:
1513 print_alu_word(&words
[i
], num_quad_words
, tabs
, i
+ 4*num_quad_words
);
1515 /* Reset word static analysis state */
1516 is_embedded_constant_half
= false;
1517 is_embedded_constant_int
= false;
1522 printf("Unknown word type %u:\n", words
[i
] & 0xF);
1524 print_quad_word(&words
[i
], tabs
);
1529 if (prefetch_flag
&& midgard_word_types
[tag
] == midgard_word_type_alu
)
1534 unsigned next
= (words
[i
] & 0xF0) >> 4;
1536 /* We are parsing per bundle anyway */
1537 midg_stats
.bundle_count
++;
1538 midg_stats
.quadword_count
+= num_quad_words
;
1540 /* Break based on instruction prefetch flag */
1542 if (i
< num_words
&& next
== 1) {
1543 prefetch_flag
= true;
1545 if (midgard_word_types
[words
[i
] & 0xF] != midgard_word_type_alu
)
1549 i
+= 4 * num_quad_words
;
1552 /* We computed work_count as max_work_registers, so add one to get the
1553 * count. If no work registers are written, you still have one work
1554 * reported, which is exactly what the hardware expects */
1556 midg_stats
.work_count
++;