2 * Copyright (C) 2019 Alyssa Rosenzweig <alyssa@rosenzweig.io>
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 FROM,
20 * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
25 #include "midgard_ops.h"
27 void mir_rewrite_index_src_single(midgard_instruction
*ins
, unsigned old
, unsigned new)
29 for (unsigned i
= 0; i
< ARRAY_SIZE(ins
->src
); ++i
) {
30 if (ins
->src
[i
] == old
)
35 void mir_rewrite_index_dst_single(midgard_instruction
*ins
, unsigned old
, unsigned new)
42 mir_get_swizzle(midgard_instruction
*ins
, unsigned idx
)
44 if (ins
->type
== TAG_ALU_4
) {
46 return ins
->csel_swizzle
;
48 unsigned b
= (idx
== 0) ? ins
->alu
.src1
: ins
->alu
.src2
;
50 midgard_vector_alu_src s
=
51 vector_alu_from_unsigned(b
);
54 } else if (ins
->type
== TAG_LOAD_STORE_4
) {
55 /* Main swizzle of a load is on the destination */
56 if (!OP_IS_STORE(ins
->load_store
.op
))
61 return ins
->load_store
.swizzle
;
65 (idx
== 2) ? ins
->load_store
.arg_2
: ins
->load_store
.arg_1
;
67 return component_to_swizzle(midgard_ldst_select(raw
).component
);
70 unreachable("Unknown load/store source");
72 } else if (ins
->type
== TAG_TEXTURE_4
) {
75 return ins
->texture
.in_reg_swizzle
;
77 /* Swizzle on bias doesn't make sense */
80 unreachable("Unknown texture source");
83 unreachable("Unknown type");
88 mir_set_swizzle(midgard_instruction
*ins
, unsigned idx
, unsigned new)
90 if (ins
->type
== TAG_ALU_4
) {
91 unsigned b
= (idx
== 0) ? ins
->alu
.src1
: ins
->alu
.src2
;
93 midgard_vector_alu_src s
=
94 vector_alu_from_unsigned(b
);
97 unsigned pack
= vector_alu_srco_unsigned(s
);
100 ins
->alu
.src1
= pack
;
102 ins
->alu
.src2
= pack
;
103 } else if (ins
->type
== TAG_LOAD_STORE_4
) {
104 /* Main swizzle of a load is on the destination */
105 if (!OP_IS_STORE(ins
->load_store
.op
))
110 ins
->load_store
.swizzle
= new;
115 (idx
== 2) ? ins
->load_store
.arg_2
: ins
->load_store
.arg_1
;
117 midgard_ldst_register_select sel
118 = midgard_ldst_select(raw
);
119 sel
.component
= swizzle_to_component(new);
120 uint8_t packed
= midgard_ldst_pack(sel
);
123 ins
->load_store
.arg_2
= packed
;
125 ins
->load_store
.arg_1
= packed
;
133 } else if (ins
->type
== TAG_TEXTURE_4
) {
136 ins
->texture
.in_reg_swizzle
= new;
143 unreachable("Unknown type");
148 mir_rewrite_index_src_single_swizzle(midgard_instruction
*ins
, unsigned old
, unsigned new, unsigned swizzle
)
150 for (unsigned i
= 0; i
< ARRAY_SIZE(ins
->src
); ++i
) {
151 if (ins
->src
[i
] != old
) continue;
155 mir_set_swizzle(ins
, i
,
156 pan_compose_swizzle(mir_get_swizzle(ins
, i
), swizzle
));
161 mir_rewrite_index_src(compiler_context
*ctx
, unsigned old
, unsigned new)
163 mir_foreach_instr_global(ctx
, ins
) {
164 mir_rewrite_index_src_single(ins
, old
, new);
169 mir_rewrite_index_src_swizzle(compiler_context
*ctx
, unsigned old
, unsigned new, unsigned swizzle
)
171 mir_foreach_instr_global(ctx
, ins
) {
172 mir_rewrite_index_src_single_swizzle(ins
, old
, new, swizzle
);
177 mir_rewrite_index_dst(compiler_context
*ctx
, unsigned old
, unsigned new)
179 mir_foreach_instr_global(ctx
, ins
) {
180 mir_rewrite_index_dst_single(ins
, old
, new);
185 mir_rewrite_index(compiler_context
*ctx
, unsigned old
, unsigned new)
187 mir_rewrite_index_src(ctx
, old
, new);
188 mir_rewrite_index_dst(ctx
, old
, new);
192 mir_use_count(compiler_context
*ctx
, unsigned value
)
194 unsigned used_count
= 0;
196 mir_foreach_instr_global(ctx
, ins
) {
197 if (mir_has_arg(ins
, value
))
204 /* Checks if a value is used only once (or totally dead), which is an important
205 * heuristic to figure out if certain optimizations are Worth It (TM) */
208 mir_single_use(compiler_context
*ctx
, unsigned value
)
210 /* We can replicate constants in places so who cares */
211 if (value
== SSA_FIXED_REGISTER(REGISTER_CONSTANT
))
214 return mir_use_count(ctx
, value
) <= 1;
218 mir_nontrivial_raw_mod(midgard_vector_alu_src src
, bool is_int
)
221 return src
.mod
== midgard_int_shift
;
227 mir_nontrivial_mod(midgard_vector_alu_src src
, bool is_int
, unsigned mask
)
229 if (mir_nontrivial_raw_mod(src
, is_int
)) return true;
231 /* size-conversion */
232 if (src
.half
) return true;
235 for (unsigned c
= 0; c
< 4; ++c
) {
236 if (!(mask
& (1 << c
))) continue;
237 if (((src
.swizzle
>> (2*c
)) & 3) != c
) return true;
244 mir_nontrivial_source2_mod(midgard_instruction
*ins
)
246 bool is_int
= midgard_is_integer_op(ins
->alu
.op
);
248 midgard_vector_alu_src src2
=
249 vector_alu_from_unsigned(ins
->alu
.src2
);
251 return mir_nontrivial_mod(src2
, is_int
, ins
->mask
);
255 mir_nontrivial_source2_mod_simple(midgard_instruction
*ins
)
257 bool is_int
= midgard_is_integer_op(ins
->alu
.op
);
259 midgard_vector_alu_src src2
=
260 vector_alu_from_unsigned(ins
->alu
.src2
);
262 return mir_nontrivial_raw_mod(src2
, is_int
) || src2
.half
;
266 mir_nontrivial_outmod(midgard_instruction
*ins
)
268 bool is_int
= midgard_is_integer_op(ins
->alu
.op
);
269 unsigned mod
= ins
->alu
.outmod
;
275 /* Type conversion is a sort of outmod */
276 if (ins
->alu
.dest_override
!= midgard_dest_override_none
)
280 return mod
!= midgard_outmod_int_wrap
;
282 return mod
!= midgard_outmod_none
;
285 /* Checks if an index will be used as a special register -- basically, if we're
286 * used as the input to a non-ALU op */
289 mir_special_index(compiler_context
*ctx
, unsigned idx
)
291 mir_foreach_instr_global(ctx
, ins
) {
292 bool is_ldst
= ins
->type
== TAG_LOAD_STORE_4
;
293 bool is_tex
= ins
->type
== TAG_TEXTURE_4
;
294 bool is_writeout
= ins
->compact_branch
&& ins
->writeout
;
296 if (!(is_ldst
|| is_tex
|| is_writeout
))
299 if (mir_has_arg(ins
, idx
))
306 /* Is a node written before a given instruction? */
309 mir_is_written_before(compiler_context
*ctx
, midgard_instruction
*ins
, unsigned node
)
311 if (node
>= SSA_FIXED_MINIMUM
)
314 mir_foreach_instr_global(ctx
, q
) {
325 /* Creates a mask of the components of a node read by an instruction, by
326 * analyzing the swizzle with respect to the instruction's mask. E.g.:
328 * fadd r0.xz, r1.yyyy, r2.zwyx
330 * will return a mask of Z/Y for r2
334 mir_mask_of_read_components_single(unsigned swizzle
, unsigned outmask
)
338 for (unsigned c
= 0; c
< 4; ++c
) {
339 if (!(outmask
& (1 << c
))) continue;
341 unsigned comp
= (swizzle
>> (2*c
)) & 3;
349 mir_source_count(midgard_instruction
*ins
)
351 if (ins
->type
== TAG_ALU_4
) {
352 /* ALU is always binary, except csel */
353 return OP_IS_CSEL(ins
->alu
.op
) ? 3 : 2;
354 } else if (ins
->type
== TAG_LOAD_STORE_4
) {
355 bool load
= !OP_IS_STORE(ins
->load_store
.op
);
356 return (load
? 2 : 3);
357 } else if (ins
->type
== TAG_TEXTURE_4
) {
358 /* Coords, bias.. TODO: Offsets? */
361 unreachable("Invalid instruction type");
366 mir_component_count_implicit(midgard_instruction
*ins
, unsigned i
)
368 if (ins
->type
== TAG_LOAD_STORE_4
) {
369 switch (ins
->load_store
.op
) {
370 /* Address implicitly 64-bit */
371 case midgard_op_ld_int4
:
372 return (i
== 0) ? 1 : 0;
374 case midgard_op_st_int4
:
375 return (i
== 1) ? 1 : 0;
386 mir_mask_of_read_components(midgard_instruction
*ins
, unsigned node
)
390 for (unsigned i
= 0; i
< mir_source_count(ins
); ++i
) {
391 if (ins
->src
[i
] != node
) continue;
393 /* Branch writeout uses all components */
394 if (ins
->compact_branch
&& ins
->writeout
&& (i
== 0))
397 unsigned swizzle
= mir_get_swizzle(ins
, i
);
398 unsigned m
= mir_mask_of_read_components_single(swizzle
, ins
->mask
);
400 /* Sometimes multi-arg ops are passed implicitly */
401 unsigned implicit
= mir_component_count_implicit(ins
, i
);
402 assert(implicit
< 2);
404 /* Extend the mask */
406 /* Ensure it's a single bit currently */
407 assert((m
>> __builtin_ctz(m
)) == 0x1);
409 /* Set the next bit to extend one*/
413 /* Handle dot products and things */
414 if (ins
->type
== TAG_ALU_4
&& !ins
->compact_branch
) {
415 unsigned channel_override
=
416 GET_CHANNEL_COUNT(alu_opcode_props
[ins
->alu
.op
].props
);
418 if (channel_override
)
419 m
= mask_of(channel_override
);
429 mir_ubo_shift(midgard_load_store_op op
)
432 case midgard_op_ld_ubo_char
:
434 case midgard_op_ld_ubo_char2
:
436 case midgard_op_ld_ubo_char4
:
438 case midgard_op_ld_ubo_short4
:
440 case midgard_op_ld_ubo_int4
:
443 unreachable("Invalid op");
447 /* Register allocation occurs after instruction scheduling, which is fine until
448 * we start needing to spill registers and therefore insert instructions into
449 * an already-scheduled program. We don't have to be terribly efficient about
450 * this, since spilling is already slow. So just semantically we need to insert
451 * the instruction into a new bundle before/after the bundle of the instruction
454 static midgard_bundle
455 mir_bundle_for_op(compiler_context
*ctx
, midgard_instruction ins
)
457 midgard_instruction
*u
= mir_upload_ins(ctx
, ins
);
459 midgard_bundle bundle
= {
461 .instruction_count
= 1,
462 .instructions
= { u
},
465 if (bundle
.tag
== TAG_ALU_4
) {
466 assert(OP_IS_MOVE(u
->alu
.op
));
469 size_t bytes_emitted
= sizeof(uint32_t) + sizeof(midgard_reg_info
) + sizeof(midgard_vector_alu
);
470 bundle
.padding
= ~(bytes_emitted
- 1) & 0xF;
471 bundle
.control
= ins
.type
| u
->unit
;
478 mir_bundle_idx_for_ins(midgard_instruction
*tag
, midgard_block
*block
)
480 midgard_bundle
*bundles
=
481 (midgard_bundle
*) block
->bundles
.data
;
483 size_t count
= (block
->bundles
.size
/ sizeof(midgard_bundle
));
485 for (unsigned i
= 0; i
< count
; ++i
) {
486 for (unsigned j
= 0; j
< bundles
[i
].instruction_count
; ++j
) {
487 if (bundles
[i
].instructions
[j
] == tag
)
492 mir_print_instruction(tag
);
493 unreachable("Instruction not scheduled in block");
497 mir_insert_instruction_before_scheduled(
498 compiler_context
*ctx
,
499 midgard_block
*block
,
500 midgard_instruction
*tag
,
501 midgard_instruction ins
)
503 unsigned before
= mir_bundle_idx_for_ins(tag
, block
);
504 size_t count
= util_dynarray_num_elements(&block
->bundles
, midgard_bundle
);
505 UNUSED
void *unused
= util_dynarray_grow(&block
->bundles
, midgard_bundle
, 1);
507 midgard_bundle
*bundles
= (midgard_bundle
*) block
->bundles
.data
;
508 memmove(bundles
+ before
+ 1, bundles
+ before
, (count
- before
) * sizeof(midgard_bundle
));
509 midgard_bundle
*before_bundle
= bundles
+ before
+ 1;
511 midgard_bundle
new = mir_bundle_for_op(ctx
, ins
);
512 memcpy(bundles
+ before
, &new, sizeof(new));
514 list_addtail(&new.instructions
[0]->link
, &before_bundle
->instructions
[0]->link
);
518 mir_insert_instruction_after_scheduled(
519 compiler_context
*ctx
,
520 midgard_block
*block
,
521 midgard_instruction
*tag
,
522 midgard_instruction ins
)
524 unsigned after
= mir_bundle_idx_for_ins(tag
, block
);
525 size_t count
= util_dynarray_num_elements(&block
->bundles
, midgard_bundle
);
526 UNUSED
void *unused
= util_dynarray_grow(&block
->bundles
, midgard_bundle
, 1);
528 midgard_bundle
*bundles
= (midgard_bundle
*) block
->bundles
.data
;
529 memmove(bundles
+ after
+ 2, bundles
+ after
+ 1, (count
- after
- 1) * sizeof(midgard_bundle
));
530 midgard_bundle
*after_bundle_1
= bundles
+ after
+ 2;
532 midgard_bundle
new = mir_bundle_for_op(ctx
, ins
);
533 memcpy(bundles
+ after
+ 1, &new, sizeof(new));
534 list_addtail(&new.instructions
[0]->link
, &after_bundle_1
->instructions
[0]->link
);