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 /* When the mode of the instruction is itself 16-bit,
359 * rep_low/high work more or less as expected. But if the mode
360 * is 32-bit and we're stepping down, you only have vec4 and
361 * the meaning shifts to rep_low as higher-half and rep_high is
362 * never seen. TODO: are other modes similar? */
364 if (mode
== midgard_reg_mode_32
) {
366 print_swizzle_helper(src
->swizzle
, src
->rep_low
);
367 assert(!src
->rep_high
);
369 print_swizzle_vec8(src
->swizzle
, src
->rep_high
, src
->rep_low
);
371 } else if (bits
== 8)
372 print_swizzle_vec16(src
->swizzle
, src
->rep_high
, src
->rep_low
, override
);
374 print_swizzle_vec4(src
->swizzle
, src
->rep_high
, src
->rep_low
);
376 print_swizzle_vec2(src
->swizzle
, src
->rep_high
, src
->rep_low
);
378 /* Since we wrapped with a function-looking thing */
380 if (is_int
&& int_mod
== midgard_int_shift
)
381 printf(") << %u", bits
);
382 else if ((is_int
&& (int_mod
!= midgard_int_normal
))
383 || (!is_int
&& src
->mod
& MIDGARD_FLOAT_MOD_ABS
))
388 decode_vector_imm(unsigned src2_reg
, unsigned imm
)
391 ret
= src2_reg
<< 11;
392 ret
|= (imm
& 0x7) << 8;
393 ret
|= (imm
>> 3) & 0xFF;
398 print_immediate(uint16_t imm
)
400 if (is_instruction_int
)
403 printf("#%g", _mesa_half_to_float(imm
));
407 update_dest(unsigned reg
)
409 /* We should record writes as marking this as a work register. Store
410 * the max register in work_count; we'll add one at the end */
413 midg_stats
.work_count
= MAX2(reg
, midg_stats
.work_count
);
414 midg_ever_written
|= (1 << reg
);
419 print_dest(unsigned reg
, midgard_reg_mode mode
, midgard_dest_override override
)
421 /* Depending on the mode and override, we determine the type of
422 * destination addressed. Absent an override, we address just the
423 * type of the operation itself */
425 unsigned bits
= bits_for_mode(mode
);
427 if (override
!= midgard_dest_override_none
)
431 print_reg(reg
, bits
);
435 print_mask_vec16(uint8_t mask
, midgard_dest_override override
)
439 for (unsigned i
= 0; i
< 8; i
++) {
443 components
[i
*2 + 1]);
447 /* For 16-bit+ masks, we read off from the 8-bit mask field. For 16-bit (vec8),
448 * it's just one bit per channel, easy peasy. For 32-bit (vec4), it's one bit
449 * per channel with one duplicate bit in the middle. For 64-bit (vec2), it's
450 * one-bit per channel with _3_ duplicate bits in the middle. Basically, just
451 * subdividing the 128-bit word in 16-bit increments. For 64-bit, we uppercase
452 * the mask to make it obvious what happened */
455 print_mask(uint8_t mask
, unsigned bits
, midgard_dest_override override
)
458 print_mask_vec16(mask
, override
);
462 /* Skip 'complete' masks */
464 if (override
== midgard_dest_override_none
) {
465 if (bits
>= 32 && mask
== 0xFF) return;
470 else if (mask
== 0xF0) {
479 unsigned skip
= (bits
/ 16);
480 bool uppercase
= bits
> 32;
481 bool tripped
= false;
483 /* To apply an upper destination override, we "shift" the alphabet.
484 * E.g. with an upper override on 32-bit, instead of xyzw, print efgh.
485 * For upper 16-bit, instead of xyzwefgh, print ijklmnop */
487 const char *alphabet
= components
;
489 if (override
== midgard_dest_override_upper
) {
490 unsigned components
= 128 / bits
;
491 alphabet
+= components
;
494 for (unsigned i
= 0; i
< 8; i
+= skip
) {
495 bool a
= (mask
& (1 << i
)) != 0;
497 for (unsigned j
= 1; j
< skip
; ++j
) {
498 bool dupe
= (mask
& (1 << (i
+ j
))) != 0;
499 tripped
|= (dupe
!= a
);
503 char c
= alphabet
[i
/ skip
];
513 printf(" /* %X */", mask
);
516 /* Prints the 4-bit masks found in texture and load/store ops, as opposed to
517 * the 8-bit masks found in (vector) ALU ops. Supports texture-style 16-bit
518 * mode as well, but not load/store-style 16-bit mode. */
521 print_mask_4(unsigned mask
, bool upper
)
532 for (unsigned i
= 0; i
< 4; ++i
) {
533 bool a
= (mask
& (1 << i
)) != 0;
535 printf("%c", components
[i
+ (upper
? 4 : 0)]);
540 print_vector_field(const char *name
, uint16_t *words
, uint16_t reg_word
,
543 midgard_reg_info
*reg_info
= (midgard_reg_info
*)®_word
;
544 midgard_vector_alu
*alu_field
= (midgard_vector_alu
*) words
;
545 midgard_reg_mode mode
= alu_field
->reg_mode
;
546 unsigned override
= alu_field
->dest_override
;
548 /* For now, prefix instruction names with their unit, until we
549 * understand how this works on a deeper level */
552 print_alu_opcode(alu_field
->op
);
554 /* Postfix with the size to disambiguate if necessary */
555 char postfix
= prefix_for_bits(bits_for_mode(mode
));
556 bool size_ambiguous
= override
!= midgard_dest_override_none
;
559 printf("%c", postfix
? postfix
: 'r');
561 /* Print the outmod, if there is one */
562 print_outmod(alu_field
->outmod
,
563 midgard_is_integer_out_op(alu_field
->op
));
567 /* Mask denoting status of 8-lanes */
568 uint8_t mask
= alu_field
->mask
;
570 /* First, print the destination */
571 print_dest(reg_info
->out_reg
, mode
, alu_field
->dest_override
);
573 if (override
!= midgard_dest_override_none
) {
574 bool modeable
= (mode
!= midgard_reg_mode_8
);
575 bool known
= override
!= 0x3; /* Unused value */
577 if (!(modeable
&& known
))
578 printf("/* do%u */ ", override
);
581 print_mask(mask
, bits_for_mode(mode
), override
);
585 bool is_int
= midgard_is_integer_op(alu_field
->op
);
586 print_vector_src(alu_field
->src1
, mode
, reg_info
->src1_reg
, override
, is_int
);
590 if (reg_info
->src2_imm
) {
591 uint16_t imm
= decode_vector_imm(reg_info
->src2_reg
, alu_field
->src2
>> 2);
592 print_immediate(imm
);
594 print_vector_src(alu_field
->src2
, mode
,
595 reg_info
->src2_reg
, override
, is_int
);
598 midg_stats
.instruction_count
++;
603 print_scalar_src(unsigned src_binary
, unsigned reg
)
605 midgard_scalar_alu_src
*src
= (midgard_scalar_alu_src
*)&src_binary
;
613 print_reg(reg
, src
->full
? 32 : 16);
615 unsigned c
= src
->component
;
618 assert((c
& 1) == 0);
622 printf(".%c", components
[c
]);
630 decode_scalar_imm(unsigned src2_reg
, unsigned imm
)
633 ret
= src2_reg
<< 11;
634 ret
|= (imm
& 3) << 9;
635 ret
|= (imm
& 4) << 6;
636 ret
|= (imm
& 0x38) << 2;
642 print_scalar_field(const char *name
, uint16_t *words
, uint16_t reg_word
,
645 midgard_reg_info
*reg_info
= (midgard_reg_info
*)®_word
;
646 midgard_scalar_alu
*alu_field
= (midgard_scalar_alu
*) words
;
648 if (alu_field
->unknown
)
649 printf("scalar ALU unknown bit set\n");
652 print_alu_opcode(alu_field
->op
);
653 print_outmod(alu_field
->outmod
,
654 midgard_is_integer_out_op(alu_field
->op
));
657 bool full
= alu_field
->output_full
;
658 update_dest(reg_info
->out_reg
);
659 print_reg(reg_info
->out_reg
, full
? 32 : 16);
660 unsigned c
= alu_field
->output_component
;
663 assert((c
& 1) == 0);
667 printf(".%c, ", components
[c
]);
669 print_scalar_src(alu_field
->src1
, reg_info
->src1_reg
);
673 if (reg_info
->src2_imm
) {
674 uint16_t imm
= decode_scalar_imm(reg_info
->src2_reg
,
676 print_immediate(imm
);
678 print_scalar_src(alu_field
->src2
, reg_info
->src2_reg
);
680 midg_stats
.instruction_count
++;
685 print_branch_op(unsigned op
)
688 case midgard_jmp_writeout_op_branch_uncond
:
692 case midgard_jmp_writeout_op_branch_cond
:
696 case midgard_jmp_writeout_op_writeout
:
700 case midgard_jmp_writeout_op_tilebuffer_pending
:
701 printf("tilebuffer.");
704 case midgard_jmp_writeout_op_discard
:
709 printf("unk%u.", op
);
715 print_branch_cond(int cond
)
718 case midgard_condition_write0
:
722 case midgard_condition_false
:
726 case midgard_condition_true
:
730 case midgard_condition_always
:
735 printf("unk%X", cond
);
741 print_compact_branch_writeout_field(uint16_t word
)
743 midgard_jmp_writeout_op op
= word
& 0x7;
746 case midgard_jmp_writeout_op_branch_uncond
: {
747 midgard_branch_uncond br_uncond
;
748 memcpy((char *) &br_uncond
, (char *) &word
, sizeof(br_uncond
));
749 printf("br.uncond ");
751 if (br_uncond
.unknown
!= 1)
752 printf("unknown:%u, ", br_uncond
.unknown
);
754 if (br_uncond
.offset
>= 0)
757 printf("%d -> ", br_uncond
.offset
);
758 print_tag_short(br_uncond
.dest_tag
);
764 case midgard_jmp_writeout_op_branch_cond
:
765 case midgard_jmp_writeout_op_writeout
:
766 case midgard_jmp_writeout_op_discard
:
768 midgard_branch_cond br_cond
;
769 memcpy((char *) &br_cond
, (char *) &word
, sizeof(br_cond
));
773 print_branch_op(br_cond
.op
);
774 print_branch_cond(br_cond
.cond
);
778 if (br_cond
.offset
>= 0)
781 printf("%d -> ", br_cond
.offset
);
782 print_tag_short(br_cond
.dest_tag
);
789 midg_stats
.instruction_count
++;
793 print_extended_branch_writeout_field(uint8_t *words
, unsigned next
)
795 midgard_branch_extended br
;
796 memcpy((char *) &br
, (char *) words
, sizeof(br
));
800 print_branch_op(br
.op
);
802 /* Condition codes are a LUT in the general case, but simply repeated 8 times for single-channel conditions.. Check this. */
804 bool single_channel
= true;
806 for (unsigned i
= 0; i
< 16; i
+= 2) {
807 single_channel
&= (((br
.cond
>> i
) & 0x3) == (br
.cond
& 0x3));
811 print_branch_cond(br
.cond
& 0x3);
813 printf("lut%X", br
.cond
);
816 printf(".unknown%u", br
.unknown
);
823 printf("%d -> ", br
.offset
);
824 print_tag_short(br
.dest_tag
);
827 unsigned I
= next
+ br
.offset
* 4;
829 if (midg_tags
[I
] && midg_tags
[I
] != br
.dest_tag
) {
830 printf("\t/* XXX TAG ERROR: jumping to ");
831 print_tag_short(br
.dest_tag
);
832 printf(" but tagged ");
833 print_tag_short(midg_tags
[I
]);
837 midg_tags
[I
] = br
.dest_tag
;
839 midg_stats
.instruction_count
++;
843 num_alu_fields_enabled(uint32_t control_word
)
847 if ((control_word
>> 17) & 1)
850 if ((control_word
>> 19) & 1)
853 if ((control_word
>> 21) & 1)
856 if ((control_word
>> 23) & 1)
859 if ((control_word
>> 25) & 1)
866 float_bitcast(uint32_t integer
)
878 print_alu_word(uint32_t *words
, unsigned num_quad_words
,
879 unsigned tabs
, unsigned next
)
881 uint32_t control_word
= words
[0];
882 uint16_t *beginning_ptr
= (uint16_t *)(words
+ 1);
883 unsigned num_fields
= num_alu_fields_enabled(control_word
);
884 uint16_t *word_ptr
= beginning_ptr
+ num_fields
;
885 unsigned num_words
= 2 + num_fields
;
887 if ((control_word
>> 16) & 1)
888 printf("unknown bit 16 enabled\n");
890 if ((control_word
>> 17) & 1) {
891 print_vector_field("vmul", word_ptr
, *beginning_ptr
, tabs
);
897 if ((control_word
>> 18) & 1)
898 printf("unknown bit 18 enabled\n");
900 if ((control_word
>> 19) & 1) {
901 print_scalar_field("sadd", word_ptr
, *beginning_ptr
, tabs
);
907 if ((control_word
>> 20) & 1)
908 printf("unknown bit 20 enabled\n");
910 if ((control_word
>> 21) & 1) {
911 print_vector_field("vadd", word_ptr
, *beginning_ptr
, tabs
);
917 if ((control_word
>> 22) & 1)
918 printf("unknown bit 22 enabled\n");
920 if ((control_word
>> 23) & 1) {
921 print_scalar_field("smul", word_ptr
, *beginning_ptr
, tabs
);
927 if ((control_word
>> 24) & 1)
928 printf("unknown bit 24 enabled\n");
930 if ((control_word
>> 25) & 1) {
931 print_vector_field("lut", word_ptr
, *beginning_ptr
, tabs
);
936 if ((control_word
>> 26) & 1) {
937 print_compact_branch_writeout_field(*word_ptr
);
942 if ((control_word
>> 27) & 1) {
943 print_extended_branch_writeout_field((uint8_t *) word_ptr
, next
);
948 if (num_quad_words
> (num_words
+ 7) / 8) {
949 assert(num_quad_words
== (num_words
+ 15) / 8);
950 //Assume that the extra quadword is constants
951 void *consts
= words
+ (4 * num_quad_words
- 4);
953 if (is_embedded_constant_int
) {
954 if (is_embedded_constant_half
) {
955 int16_t *sconsts
= (int16_t *) consts
;
956 printf("sconstants %d, %d, %d, %d\n",
962 uint32_t *iconsts
= (uint32_t *) consts
;
963 printf("iconstants 0x%X, 0x%X, 0x%X, 0x%X\n",
970 if (is_embedded_constant_half
) {
971 uint16_t *hconsts
= (uint16_t *) consts
;
972 printf("hconstants %g, %g, %g, %g\n",
973 _mesa_half_to_float(hconsts
[0]),
974 _mesa_half_to_float(hconsts
[1]),
975 _mesa_half_to_float(hconsts
[2]),
976 _mesa_half_to_float(hconsts
[3]));
978 uint32_t *fconsts
= (uint32_t *) consts
;
979 printf("fconstants %g, %g, %g, %g\n",
980 float_bitcast(fconsts
[0]),
981 float_bitcast(fconsts
[1]),
982 float_bitcast(fconsts
[2]),
983 float_bitcast(fconsts
[3]));
991 print_varying_parameters(midgard_load_store_word
*word
)
993 midgard_varying_parameter param
;
994 unsigned v
= word
->varying_parameters
;
995 memcpy(¶m
, &v
, sizeof(param
));
997 if (param
.is_varying
) {
998 /* If a varying, there are qualifiers */
1002 if (param
.interpolation
!= midgard_interp_default
) {
1003 if (param
.interpolation
== midgard_interp_centroid
)
1004 printf(".centroid");
1006 printf(".interp%d", param
.interpolation
);
1009 if (param
.modifier
!= midgard_varying_mod_none
) {
1010 if (param
.modifier
== midgard_varying_mod_perspective_w
)
1011 printf(".perspectivew");
1012 else if (param
.modifier
== midgard_varying_mod_perspective_z
)
1013 printf(".perspectivez");
1015 printf(".mod%d", param
.modifier
);
1017 } else if (param
.flat
|| param
.interpolation
|| param
.modifier
) {
1018 printf(" /* is_varying not set but varying metadata attached */");
1021 if (param
.zero0
|| param
.zero1
|| param
.zero2
)
1022 printf(" /* zero tripped, %u %u %u */ ", param
.zero0
, param
.zero1
, param
.zero2
);
1026 is_op_varying(unsigned op
)
1029 case midgard_op_st_vary_16
:
1030 case midgard_op_st_vary_32
:
1031 case midgard_op_st_vary_32i
:
1032 case midgard_op_st_vary_32u
:
1033 case midgard_op_ld_vary_16
:
1034 case midgard_op_ld_vary_32
:
1035 case midgard_op_ld_vary_32i
:
1036 case midgard_op_ld_vary_32u
:
1044 is_op_attribute(unsigned op
)
1047 case midgard_op_ld_attr_16
:
1048 case midgard_op_ld_attr_32
:
1049 case midgard_op_ld_attr_32i
:
1050 case midgard_op_ld_attr_32u
:
1058 print_load_store_arg(uint8_t arg
, unsigned index
)
1060 /* Try to interpret as a register */
1061 midgard_ldst_register_select sel
;
1062 memcpy(&sel
, &arg
, sizeof(arg
));
1064 /* If unknown is set, we're not sure what this is or how to
1065 * interpret it. But if it's zero, we get it. */
1068 printf("0x%02X", arg
);
1072 unsigned reg
= REGISTER_LDST_BASE
+ sel
.select
;
1073 char comp
= components
[sel
.component
];
1075 printf("r%u.%c", reg
, comp
);
1077 /* Only print a shift if it's non-zero. Shifts only make sense for the
1078 * second index. For the first, we're not sure what it means yet */
1082 printf(" << %u", sel
.shift
);
1084 printf(" /* %X */", sel
.shift
);
1089 update_stats(signed *stat
, unsigned address
)
1092 *stat
= MAX2(*stat
, address
+ 1);
1096 print_load_store_instr(uint64_t data
,
1099 midgard_load_store_word
*word
= (midgard_load_store_word
*) &data
;
1101 print_ld_st_opcode(word
->op
);
1103 unsigned address
= word
->address
;
1105 if (is_op_varying(word
->op
)) {
1106 print_varying_parameters(word
);
1108 /* Do some analysis: check if direct cacess */
1110 if ((word
->arg_2
== 0x1E) && midg_stats
.varying_count
>= 0)
1111 update_stats(&midg_stats
.varying_count
, address
);
1113 midg_stats
.varying_count
= -16;
1114 } else if (is_op_attribute(word
->op
)) {
1115 if ((word
->arg_2
== 0x1E) && midg_stats
.attribute_count
>= 0)
1116 update_stats(&midg_stats
.attribute_count
, address
);
1118 midg_stats
.attribute_count
= -16;
1121 printf(" r%u", word
->reg
);
1122 print_mask_4(word
->mask
, false);
1124 if (!OP_IS_STORE(word
->op
))
1125 update_dest(word
->reg
);
1127 bool is_ubo
= OP_IS_UBO_READ(word
->op
);
1130 /* UBOs use their own addressing scheme */
1132 int lo
= word
->varying_parameters
>> 7;
1133 int hi
= word
->address
;
1135 /* TODO: Combine fields logically */
1136 address
= (hi
<< 3) | lo
;
1139 printf(", %u", address
);
1141 print_swizzle_vec4(word
->swizzle
, false, false);
1146 printf("ubo%u", word
->arg_1
);
1147 update_stats(&midg_stats
.uniform_buffer_count
, word
->arg_1
);
1149 print_load_store_arg(word
->arg_1
, 0);
1152 print_load_store_arg(word
->arg_2
, 1);
1153 printf(" /* %X */\n", word
->varying_parameters
);
1155 midg_stats
.instruction_count
++;
1159 print_load_store_word(uint32_t *word
, unsigned tabs
)
1161 midgard_load_store
*load_store
= (midgard_load_store
*) word
;
1163 if (load_store
->word1
!= 3) {
1164 print_load_store_instr(load_store
->word1
, tabs
);
1167 if (load_store
->word2
!= 3) {
1168 print_load_store_instr(load_store
->word2
, tabs
);
1173 print_texture_reg_select(uint8_t u
, unsigned base
)
1175 midgard_tex_register_select sel
;
1176 memcpy(&sel
, &u
, sizeof(u
));
1181 printf("r%u", base
+ sel
.select
);
1183 unsigned component
= sel
.component
;
1185 /* Use the upper half in half-reg mode */
1191 printf(".%c", components
[component
]);
1193 assert(sel
.zero
== 0);
1197 print_texture_format(int format
)
1199 /* Act like a modifier */
1203 DEFINE_CASE(MALI_TEX_1D
, "1d");
1204 DEFINE_CASE(MALI_TEX_2D
, "2d");
1205 DEFINE_CASE(MALI_TEX_3D
, "3d");
1206 DEFINE_CASE(MALI_TEX_CUBE
, "cube");
1209 unreachable("Bad format");
1214 midgard_op_has_helpers(unsigned op
, bool gather
)
1220 case TEXTURE_OP_NORMAL
:
1221 case TEXTURE_OP_DFDX
:
1222 case TEXTURE_OP_DFDY
:
1230 print_texture_op(unsigned op
, bool gather
)
1232 /* Act like a bare name, like ESSL functions */
1235 printf("textureGather");
1237 unsigned component
= op
>> 4;
1238 unsigned bottom
= op
& 0xF;
1241 printf("_unk%u", bottom
);
1243 printf(".%c", components
[component
]);
1248 DEFINE_CASE(TEXTURE_OP_NORMAL
, "texture");
1249 DEFINE_CASE(TEXTURE_OP_LOD
, "textureLod");
1250 DEFINE_CASE(TEXTURE_OP_TEXEL_FETCH
, "texelFetch");
1251 DEFINE_CASE(TEXTURE_OP_DFDX
, "dFdx");
1252 DEFINE_CASE(TEXTURE_OP_DFDY
, "dFdy");
1255 printf("tex_%X", op
);
1261 texture_op_takes_bias(unsigned op
)
1263 return op
== TEXTURE_OP_NORMAL
;
1267 sampler_type_name(enum mali_sampler_type t
)
1270 case MALI_SAMPLER_FLOAT
:
1272 case MALI_SAMPLER_UNSIGNED
:
1274 case MALI_SAMPLER_SIGNED
:
1285 print_texture_word(uint32_t *word
, unsigned tabs
, unsigned in_reg_base
, unsigned out_reg_base
)
1287 midgard_texture_word
*texture
= (midgard_texture_word
*) word
;
1289 midg_stats
.helper_invocations
|=
1290 midgard_op_has_helpers(texture
->op
, texture
->is_gather
);
1292 /* Broad category of texture operation in question */
1293 print_texture_op(texture
->op
, texture
->is_gather
);
1295 /* Specific format in question */
1296 print_texture_format(texture
->format
);
1298 /* Instruction "modifiers" parallel the ALU instructions. */
1300 if (texture
->shadow
)
1309 /* Output modifiers are always interpreted floatly */
1310 print_outmod(texture
->outmod
, false);
1312 printf(" %sr%d", texture
->out_full
? "" : "h",
1313 out_reg_base
+ texture
->out_reg_select
);
1314 print_mask_4(texture
->mask
, texture
->out_upper
);
1315 assert(!(texture
->out_full
&& texture
->out_upper
));
1318 /* Depending on whether we read from textures directly or indirectly,
1319 * we may be able to update our analysis */
1321 if (texture
->texture_register
) {
1323 print_texture_reg_select(texture
->texture_handle
, in_reg_base
);
1326 /* Indirect, tut tut */
1327 midg_stats
.texture_count
= -16;
1329 printf("texture%u, ", texture
->texture_handle
);
1330 update_stats(&midg_stats
.texture_count
, texture
->texture_handle
);
1333 /* Print the type, GL style */
1334 printf("%csampler", sampler_type_name(texture
->sampler_type
));
1336 if (texture
->sampler_register
) {
1338 print_texture_reg_select(texture
->sampler_handle
, in_reg_base
);
1341 midg_stats
.sampler_count
= -16;
1343 printf("%u", texture
->sampler_handle
);
1344 update_stats(&midg_stats
.sampler_count
, texture
->sampler_handle
);
1347 print_swizzle_vec4(texture
->swizzle
, false, false);
1348 printf(", %sr%d", texture
->in_reg_full
? "" : "h", in_reg_base
+ texture
->in_reg_select
);
1349 assert(!(texture
->in_reg_full
&& texture
->in_reg_upper
));
1351 /* TODO: integrate with swizzle */
1352 if (texture
->in_reg_upper
)
1355 print_swizzle_vec4(texture
->in_reg_swizzle
, false, false);
1357 /* There is *always* an offset attached. Of
1358 * course, that offset is just immediate #0 for a
1359 * GLES call that doesn't take an offset. If there
1360 * is a non-negative non-zero offset, this is
1361 * specified in immediate offset mode, with the
1362 * values in the offset_* fields as immediates. If
1363 * this is a negative offset, we instead switch to
1364 * a register offset mode, where the offset_*
1365 * fields become register triplets */
1367 if (texture
->offset_register
) {
1370 bool full
= texture
->offset_x
& 1;
1371 bool select
= texture
->offset_x
& 2;
1372 bool upper
= texture
->offset_x
& 4;
1374 printf("%sr%d", full
? "" : "h", in_reg_base
+ select
);
1375 assert(!(texture
->out_full
&& texture
->out_upper
));
1377 /* TODO: integrate with swizzle */
1381 /* The less questions you ask, the better. */
1383 unsigned swizzle_lo
, swizzle_hi
;
1384 unsigned orig_y
= texture
->offset_y
;
1385 unsigned orig_z
= texture
->offset_z
;
1387 memcpy(&swizzle_lo
, &orig_y
, sizeof(unsigned));
1388 memcpy(&swizzle_hi
, &orig_z
, sizeof(unsigned));
1390 /* Duplicate hi swizzle over */
1391 assert(swizzle_hi
< 4);
1392 swizzle_hi
= (swizzle_hi
<< 2) | swizzle_hi
;
1394 unsigned swiz
= (swizzle_lo
<< 4) | swizzle_hi
;
1395 unsigned reversed
= util_bitreverse(swiz
) >> 24;
1396 print_swizzle_vec4(reversed
, false, false);
1399 } else if (texture
->offset_x
|| texture
->offset_y
|| texture
->offset_z
) {
1400 /* Only select ops allow negative immediate offsets, verify */
1402 bool neg_x
= texture
->offset_x
< 0;
1403 bool neg_y
= texture
->offset_y
< 0;
1404 bool neg_z
= texture
->offset_z
< 0;
1405 bool any_neg
= neg_x
|| neg_y
|| neg_z
;
1407 if (any_neg
&& texture
->op
!= TEXTURE_OP_TEXEL_FETCH
)
1408 printf("/* invalid negative */ ");
1410 /* Regardless, just print the immediate offset */
1412 printf(" + <%d, %d, %d>, ",
1420 char lod_operand
= texture_op_takes_bias(texture
->op
) ? '+' : '=';
1422 if (texture
->lod_register
) {
1423 printf("lod %c ", lod_operand
);
1424 print_texture_reg_select(texture
->bias
, in_reg_base
);
1427 if (texture
->bias_int
)
1428 printf(" /* bias_int = 0x%X */", texture
->bias_int
);
1429 } else if (texture
->op
== TEXTURE_OP_TEXEL_FETCH
) {
1430 /* For texel fetch, the int LOD is in the fractional place and
1431 * there is no fraction / possibility of bias. We *always* have
1432 * an explicit LOD, even if it's zero. */
1434 if (texture
->bias_int
)
1435 printf(" /* bias_int = 0x%X */ ", texture
->bias_int
);
1437 printf("lod = %u, ", texture
->bias
);
1438 } else if (texture
->bias
|| texture
->bias_int
) {
1439 signed bias_int
= texture
->bias_int
;
1440 float bias_frac
= texture
->bias
/ 256.0f
;
1441 float bias
= bias_int
+ bias_frac
;
1443 bool is_bias
= texture_op_takes_bias(texture
->op
);
1444 char sign
= (bias
>= 0.0) ? '+' : '-';
1445 char operand
= is_bias
? sign
: '=';
1447 printf("lod %c %f, ", operand
, fabsf(bias
));
1452 /* While not zero in general, for these simple instructions the
1453 * following unknowns are zero, so we don't include them */
1455 if (texture
->unknown4
||
1456 texture
->unknownA
||
1457 texture
->unknown8
) {
1458 printf("// unknown4 = 0x%x\n", texture
->unknown4
);
1459 printf("// unknownA = 0x%x\n", texture
->unknownA
);
1460 printf("// unknown8 = 0x%x\n", texture
->unknown8
);
1463 midg_stats
.instruction_count
++;
1466 struct midgard_disasm_stats
1467 disassemble_midgard(uint8_t *code
, size_t size
, unsigned gpu_id
, gl_shader_stage stage
)
1469 uint32_t *words
= (uint32_t *) code
;
1470 unsigned num_words
= size
/ 4;
1473 bool prefetch_flag
= false;
1475 int last_next_tag
= -1;
1479 midg_tags
= calloc(sizeof(midg_tags
[0]), num_words
);
1481 /* Stats for shader-db */
1482 memset(&midg_stats
, 0, sizeof(midg_stats
));
1483 midg_ever_written
= 0;
1485 while (i
< num_words
) {
1486 unsigned tag
= words
[i
] & 0xF;
1487 unsigned next_tag
= (words
[i
] >> 4) & 0xF;
1488 unsigned num_quad_words
= midgard_word_size
[tag
];
1490 if (midg_tags
[i
] && midg_tags
[i
] != tag
) {
1491 printf("\t/* XXX: TAG ERROR branch, got ");
1492 print_tag_short(tag
);
1493 printf(" expected ");
1494 print_tag_short(midg_tags
[i
]);
1501 if (last_next_tag
> 1) {
1502 if (last_next_tag
!= tag
) {
1503 printf("\t/* XXX: TAG ERROR sequence, got ");
1504 print_tag_short(tag
);
1505 printf(" expected ");
1506 print_tag_short(last_next_tag
);
1510 /* TODO: Check ALU case */
1513 last_next_tag
= next_tag
;
1515 switch (midgard_word_types
[tag
]) {
1516 case midgard_word_type_texture
: {
1517 /* Texturing uses ldst/work space on T720 */
1518 bool has_texture_pipeline
= gpu_id
!= 0x0720;
1519 print_texture_word(&words
[i
], tabs
,
1520 has_texture_pipeline
? REG_TEX_BASE
: 0,
1521 has_texture_pipeline
? REG_TEX_BASE
: REGISTER_LDST_BASE
);
1525 case midgard_word_type_load_store
:
1526 print_load_store_word(&words
[i
], tabs
);
1529 case midgard_word_type_alu
:
1530 print_alu_word(&words
[i
], num_quad_words
, tabs
, i
+ 4*num_quad_words
);
1532 /* Reset word static analysis state */
1533 is_embedded_constant_half
= false;
1534 is_embedded_constant_int
= false;
1539 printf("Unknown word type %u:\n", words
[i
] & 0xF);
1541 print_quad_word(&words
[i
], tabs
);
1546 if (prefetch_flag
&& midgard_word_types
[tag
] == midgard_word_type_alu
)
1551 unsigned next
= (words
[i
] & 0xF0) >> 4;
1553 /* We are parsing per bundle anyway */
1554 midg_stats
.bundle_count
++;
1555 midg_stats
.quadword_count
+= num_quad_words
;
1557 /* Break based on instruction prefetch flag */
1559 if (i
< num_words
&& next
== 1) {
1560 prefetch_flag
= true;
1562 if (midgard_word_types
[words
[i
] & 0xF] != midgard_word_type_alu
)
1566 i
+= 4 * num_quad_words
;
1569 /* We computed work_count as max_work_registers, so add one to get the
1570 * count. If no work registers are written, you still have one work
1571 * reported, which is exactly what the hardware expects */
1573 midg_stats
.work_count
++;