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_rewrite_index_src_single_swizzle(midgard_instruction
*ins
, unsigned old
, unsigned new, unsigned *swizzle
)
44 for (unsigned i
= 0; i
< ARRAY_SIZE(ins
->src
); ++i
) {
45 if (ins
->src
[i
] != old
) continue;
48 mir_compose_swizzle(ins
->swizzle
[i
], swizzle
, ins
->swizzle
[i
]);
53 mir_rewrite_index_src(compiler_context
*ctx
, unsigned old
, unsigned new)
55 mir_foreach_instr_global(ctx
, ins
) {
56 mir_rewrite_index_src_single(ins
, old
, new);
61 mir_rewrite_index_src_swizzle(compiler_context
*ctx
, unsigned old
, unsigned new, unsigned *swizzle
)
63 mir_foreach_instr_global(ctx
, ins
) {
64 mir_rewrite_index_src_single_swizzle(ins
, old
, new, swizzle
);
69 mir_rewrite_index_dst(compiler_context
*ctx
, unsigned old
, unsigned new)
71 mir_foreach_instr_global(ctx
, ins
) {
72 mir_rewrite_index_dst_single(ins
, old
, new);
77 mir_rewrite_index(compiler_context
*ctx
, unsigned old
, unsigned new)
79 mir_rewrite_index_src(ctx
, old
, new);
80 mir_rewrite_index_dst(ctx
, old
, new);
84 mir_use_count(compiler_context
*ctx
, unsigned value
)
86 unsigned used_count
= 0;
88 mir_foreach_instr_global(ctx
, ins
) {
89 if (mir_has_arg(ins
, value
))
96 /* Checks if a value is used only once (or totally dead), which is an important
97 * heuristic to figure out if certain optimizations are Worth It (TM) */
100 mir_single_use(compiler_context
*ctx
, unsigned value
)
102 /* We can replicate constants in places so who cares */
103 if (value
== SSA_FIXED_REGISTER(REGISTER_CONSTANT
))
106 return mir_use_count(ctx
, value
) <= 1;
110 mir_nontrivial_raw_mod(midgard_vector_alu_src src
, bool is_int
)
113 return src
.mod
== midgard_int_shift
;
119 mir_nontrivial_mod(midgard_vector_alu_src src
, bool is_int
, unsigned mask
, unsigned *swizzle
)
121 if (mir_nontrivial_raw_mod(src
, is_int
)) return true;
123 /* size-conversion */
124 if (src
.half
) return true;
126 for (unsigned c
= 0; c
< 16; ++c
) {
127 if (!(mask
& (1 << c
))) continue;
128 if (swizzle
[c
] != c
) return true;
135 mir_nontrivial_source2_mod(midgard_instruction
*ins
)
137 bool is_int
= midgard_is_integer_op(ins
->alu
.op
);
139 midgard_vector_alu_src src2
=
140 vector_alu_from_unsigned(ins
->alu
.src2
);
142 return mir_nontrivial_mod(src2
, is_int
, ins
->mask
, ins
->swizzle
[1]);
146 mir_nontrivial_source2_mod_simple(midgard_instruction
*ins
)
148 bool is_int
= midgard_is_integer_op(ins
->alu
.op
);
150 midgard_vector_alu_src src2
=
151 vector_alu_from_unsigned(ins
->alu
.src2
);
153 return mir_nontrivial_raw_mod(src2
, is_int
) || src2
.half
;
157 mir_nontrivial_outmod(midgard_instruction
*ins
)
159 bool is_int
= midgard_is_integer_op(ins
->alu
.op
);
160 unsigned mod
= ins
->alu
.outmod
;
162 if (ins
->dest_type
!= ins
->src_types
[1])
166 return mod
!= midgard_outmod_int_wrap
;
168 return mod
!= midgard_outmod_none
;
172 mir_from_bytemask(uint16_t bytemask
, unsigned bits
)
175 unsigned count
= bits
/ 8;
177 for (unsigned c
= 0, d
= 0; c
< 16; c
+= count
, ++d
) {
178 bool a
= (bytemask
& (1 << c
)) != 0;
180 for (unsigned q
= c
; q
< count
; ++q
)
181 assert(((bytemask
& (1 << q
)) != 0) == a
);
189 /* Rounds up a bytemask to fill a given component count. Iterate each
190 * component, and check if any bytes in the component are masked on */
193 mir_round_bytemask_up(uint16_t mask
, unsigned bits
)
195 unsigned bytes
= bits
/ 8;
196 unsigned maxmask
= mask_of(bytes
);
197 unsigned channels
= 16 / bytes
;
199 for (unsigned c
= 0; c
< channels
; ++c
) {
200 unsigned submask
= maxmask
<< (c
* bytes
);
209 /* Grabs the per-byte mask of an instruction (as opposed to per-component) */
212 mir_bytemask(midgard_instruction
*ins
)
214 unsigned type_size
= nir_alu_type_get_type_size(ins
->dest_type
);
215 return pan_to_bytemask(type_size
, ins
->mask
);
219 mir_set_bytemask(midgard_instruction
*ins
, uint16_t bytemask
)
221 unsigned type_size
= nir_alu_type_get_type_size(ins
->dest_type
);
222 ins
->mask
= mir_from_bytemask(bytemask
, type_size
);
225 /* Checks if we should use an upper destination override, rather than the lower
226 * one in the IR. Returns zero if no, returns the bytes to shift otherwise */
229 mir_upper_override(midgard_instruction
*ins
)
231 /* If there is no override, there is no upper override, tautology */
232 if (ins
->alu
.dest_override
== midgard_dest_override_none
)
235 /* Make sure we didn't already lower somehow */
236 assert(ins
->alu
.dest_override
== midgard_dest_override_lower
);
238 /* There are 16 bytes per vector, so there are (16/bytes)
239 * components per vector. So the magic half is half of
240 * (16/bytes), which simplifies to 8/bytes = 8 / (bits / 8) = 64 / bits
243 unsigned type_size
= nir_alu_type_get_type_size(ins
->dest_type
);
244 unsigned threshold
= 64 / type_size
;
246 /* How many components did we shift over? */
247 unsigned zeroes
= __builtin_ctz(ins
->mask
);
249 /* Did we hit the threshold? */
250 return (zeroes
>= threshold
) ? threshold
: 0;
253 /* Creates a mask of the components of a node read by an instruction, by
254 * analyzing the swizzle with respect to the instruction's mask. E.g.:
256 * fadd r0.xz, r1.yyyy, r2.zwyx
258 * will return a mask of Z/Y for r2
262 mir_bytemask_of_read_components_single(unsigned *swizzle
, unsigned inmask
, unsigned bits
)
266 for (unsigned c
= 0; c
< MIR_VEC_COMPONENTS
; ++c
) {
267 if (!(inmask
& (1 << c
))) continue;
268 cmask
|= (1 << swizzle
[c
]);
271 return pan_to_bytemask(bits
, cmask
);
275 mir_bytemask_of_read_components_index(midgard_instruction
*ins
, unsigned i
)
277 if (ins
->compact_branch
&& ins
->writeout
&& (i
== 0)) {
278 /* Non-ZS writeout uses all components */
279 if (!ins
->writeout_depth
&& !ins
->writeout_stencil
)
282 /* For ZS-writeout, if both Z and S are written we need two
283 * components, otherwise we only need one.
285 if (ins
->writeout_depth
&& ins
->writeout_stencil
)
291 /* Conditional branches read one 32-bit component = 4 bytes (TODO: multi branch??) */
292 if (ins
->compact_branch
&& ins
->branch
.conditional
&& (i
== 0))
295 /* ALU ops act componentwise so we need to pay attention to
296 * their mask. Texture/ldst does not so we don't clamp source
297 * readmasks based on the writemask */
298 unsigned qmask
= (ins
->type
== TAG_ALU_4
) ? ins
->mask
: ~0;
300 /* Handle dot products and things */
301 if (ins
->type
== TAG_ALU_4
&& !ins
->compact_branch
) {
302 unsigned props
= alu_opcode_props
[ins
->alu
.op
].props
;
304 unsigned channel_override
= GET_CHANNEL_COUNT(props
);
306 if (channel_override
)
307 qmask
= mask_of(channel_override
);
310 return mir_bytemask_of_read_components_single(ins
->swizzle
[i
], qmask
,
311 nir_alu_type_get_type_size(ins
->src_types
[i
]));
315 mir_bytemask_of_read_components(midgard_instruction
*ins
, unsigned node
)
322 mir_foreach_src(ins
, i
) {
323 if (ins
->src
[i
] != node
) continue;
324 mask
|= mir_bytemask_of_read_components_index(ins
, i
);
330 /* Register allocation occurs after instruction scheduling, which is fine until
331 * we start needing to spill registers and therefore insert instructions into
332 * an already-scheduled program. We don't have to be terribly efficient about
333 * this, since spilling is already slow. So just semantically we need to insert
334 * the instruction into a new bundle before/after the bundle of the instruction
337 static midgard_bundle
338 mir_bundle_for_op(compiler_context
*ctx
, midgard_instruction ins
)
340 midgard_instruction
*u
= mir_upload_ins(ctx
, ins
);
342 midgard_bundle bundle
= {
344 .instruction_count
= 1,
345 .instructions
= { u
},
348 if (bundle
.tag
== TAG_ALU_4
) {
349 assert(OP_IS_MOVE(u
->alu
.op
));
352 size_t bytes_emitted
= sizeof(uint32_t) + sizeof(midgard_reg_info
) + sizeof(midgard_vector_alu
);
353 bundle
.padding
= ~(bytes_emitted
- 1) & 0xF;
354 bundle
.control
= ins
.type
| u
->unit
;
361 mir_bundle_idx_for_ins(midgard_instruction
*tag
, midgard_block
*block
)
363 midgard_bundle
*bundles
=
364 (midgard_bundle
*) block
->bundles
.data
;
366 size_t count
= (block
->bundles
.size
/ sizeof(midgard_bundle
));
368 for (unsigned i
= 0; i
< count
; ++i
) {
369 for (unsigned j
= 0; j
< bundles
[i
].instruction_count
; ++j
) {
370 if (bundles
[i
].instructions
[j
] == tag
)
375 mir_print_instruction(tag
);
376 unreachable("Instruction not scheduled in block");
380 mir_insert_instruction_before_scheduled(
381 compiler_context
*ctx
,
382 midgard_block
*block
,
383 midgard_instruction
*tag
,
384 midgard_instruction ins
)
386 unsigned before
= mir_bundle_idx_for_ins(tag
, block
);
387 size_t count
= util_dynarray_num_elements(&block
->bundles
, midgard_bundle
);
388 UNUSED
void *unused
= util_dynarray_grow(&block
->bundles
, midgard_bundle
, 1);
390 midgard_bundle
*bundles
= (midgard_bundle
*) block
->bundles
.data
;
391 memmove(bundles
+ before
+ 1, bundles
+ before
, (count
- before
) * sizeof(midgard_bundle
));
392 midgard_bundle
*before_bundle
= bundles
+ before
+ 1;
394 midgard_bundle
new = mir_bundle_for_op(ctx
, ins
);
395 memcpy(bundles
+ before
, &new, sizeof(new));
397 list_addtail(&new.instructions
[0]->link
, &before_bundle
->instructions
[0]->link
);
398 block
->quadword_count
+= midgard_tag_props
[new.tag
].size
;
402 mir_insert_instruction_after_scheduled(
403 compiler_context
*ctx
,
404 midgard_block
*block
,
405 midgard_instruction
*tag
,
406 midgard_instruction ins
)
408 /* We need to grow the bundles array to add our new bundle */
409 size_t count
= util_dynarray_num_elements(&block
->bundles
, midgard_bundle
);
410 UNUSED
void *unused
= util_dynarray_grow(&block
->bundles
, midgard_bundle
, 1);
412 /* Find the bundle that we want to insert after */
413 unsigned after
= mir_bundle_idx_for_ins(tag
, block
);
415 /* All the bundles after that one, we move ahead by one */
416 midgard_bundle
*bundles
= (midgard_bundle
*) block
->bundles
.data
;
417 memmove(bundles
+ after
+ 2, bundles
+ after
+ 1, (count
- after
- 1) * sizeof(midgard_bundle
));
418 midgard_bundle
*after_bundle
= bundles
+ after
;
420 midgard_bundle
new = mir_bundle_for_op(ctx
, ins
);
421 memcpy(bundles
+ after
+ 1, &new, sizeof(new));
422 list_add(&new.instructions
[0]->link
, &after_bundle
->instructions
[after_bundle
->instruction_count
- 1]->link
);
423 block
->quadword_count
+= midgard_tag_props
[new.tag
].size
;
426 /* Flip the first-two arguments of a (binary) op. Currently ALU
427 * only, no known uses for ldst/tex */
430 mir_flip(midgard_instruction
*ins
)
432 unsigned temp
= ins
->src
[0];
433 ins
->src
[0] = ins
->src
[1];
436 assert(ins
->type
== TAG_ALU_4
);
438 temp
= ins
->alu
.src1
;
439 ins
->alu
.src1
= ins
->alu
.src2
;
440 ins
->alu
.src2
= temp
;
442 temp
= ins
->src_types
[0];
443 ins
->src_types
[0] = ins
->src_types
[1];
444 ins
->src_types
[1] = temp
;
446 temp
= ins
->src_abs
[0];
447 ins
->src_abs
[0] = ins
->src_abs
[1];
448 ins
->src_abs
[1] = temp
;
450 temp
= ins
->src_neg
[0];
451 ins
->src_neg
[0] = ins
->src_neg
[1];
452 ins
->src_neg
[1] = temp
;
454 unsigned temp_swizzle
[16];
455 memcpy(temp_swizzle
, ins
->swizzle
[0], sizeof(ins
->swizzle
[0]));
456 memcpy(ins
->swizzle
[0], ins
->swizzle
[1], sizeof(ins
->swizzle
[0]));
457 memcpy(ins
->swizzle
[1], temp_swizzle
, sizeof(ins
->swizzle
[0]));
460 /* Before squashing, calculate ctx->temp_count just by observing the MIR */
463 mir_compute_temp_count(compiler_context
*ctx
)
468 unsigned max_dest
= 0;
470 mir_foreach_instr_global(ctx
, ins
) {
471 if (ins
->dest
< SSA_FIXED_MINIMUM
)
472 max_dest
= MAX2(max_dest
, ins
->dest
+ 1);
475 ctx
->temp_count
= max_dest
;