2 * Copyright (C) 2018 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
24 #include <sys/types.h>
33 #include "main/mtypes.h"
34 #include "compiler/glsl/glsl_to_nir.h"
35 #include "compiler/nir_types.h"
36 #include "main/imports.h"
37 #include "compiler/nir/nir_builder.h"
38 #include "util/half_float.h"
39 #include "util/register_allocate.h"
40 #include "util/u_debug.h"
41 #include "util/u_dynarray.h"
42 #include "util/list.h"
43 #include "main/mtypes.h"
46 #include "midgard_nir.h"
47 #include "midgard_compile.h"
50 #include "disassemble.h"
52 static const struct debug_named_value debug_options
[] = {
53 {"msgs", MIDGARD_DBG_MSGS
, "Print debug messages"},
54 {"shaders", MIDGARD_DBG_SHADERS
, "Dump shaders in NIR and MIR"},
58 DEBUG_GET_ONCE_FLAGS_OPTION(midgard_debug
, "MIDGARD_MESA_DEBUG", debug_options
, 0)
60 int midgard_debug
= 0;
62 #define DBG(fmt, ...) \
63 do { if (midgard_debug & MIDGARD_DBG_MSGS) \
64 fprintf(stderr, "%s:%d: "fmt, \
65 __FUNCTION__, __LINE__, ##__VA_ARGS__); } while (0)
67 /* Instruction arguments represented as block-local SSA indices, rather than
68 * registers. Negative values mean unused. */
75 /* src1 is -not- SSA but instead a 16-bit inline constant to be smudged
76 * in. Only valid for ALU ops. */
80 /* Forward declare so midgard_branch can reference */
83 /* Target types. Defaults to TARGET_GOTO (the type corresponding directly to
84 * the hardware), hence why that must be zero. TARGET_DISCARD signals this
85 * instruction is actually a discard op. */
88 #define TARGET_BREAK 1
89 #define TARGET_CONTINUE 2
90 #define TARGET_DISCARD 3
92 typedef struct midgard_branch
{
93 /* If conditional, the condition is specified in r31.w */
96 /* For conditionals, if this is true, we branch on FALSE. If false, we branch on TRUE. */
97 bool invert_conditional
;
99 /* Branch targets: the start of a block, the start of a loop (continue), the end of a loop (break). Value is one of TARGET_ */
100 unsigned target_type
;
102 /* The actual target */
111 midgard_is_branch_unit(unsigned unit
)
113 return (unit
== ALU_ENAB_BRANCH
) || (unit
== ALU_ENAB_BR_COMPACT
);
116 /* Generic in-memory data type repesenting a single logical instruction, rather
117 * than a single instruction group. This is the preferred form for code gen.
118 * Multiple midgard_insturctions will later be combined during scheduling,
119 * though this is not represented in this structure. Its format bridges
120 * the low-level binary representation with the higher level semantic meaning.
122 * Notably, it allows registers to be specified as block local SSA, for code
123 * emitted before the register allocation pass.
126 typedef struct midgard_instruction
{
127 /* Must be first for casting */
128 struct list_head link
;
130 unsigned type
; /* ALU, load/store, texture */
132 /* If the register allocator has not run yet... */
135 /* Special fields for an ALU instruction */
136 midgard_reg_info registers
;
138 /* I.e. (1 << alu_bit) */
143 uint16_t inline_constant
;
144 bool has_blend_constant
;
148 bool prepacked_branch
;
151 midgard_load_store_word load_store
;
152 midgard_vector_alu alu
;
153 midgard_texture_word texture
;
154 midgard_branch_extended branch_extended
;
157 /* General branch, rather than packed br_compact. Higher level
158 * than the other components */
159 midgard_branch branch
;
161 } midgard_instruction
;
163 typedef struct midgard_block
{
164 /* Link to next block. Must be first for mir_get_block */
165 struct list_head link
;
167 /* List of midgard_instructions emitted for the current block */
168 struct list_head instructions
;
172 /* List of midgard_bundles emitted (after the scheduler has run) */
173 struct util_dynarray bundles
;
175 /* Number of quadwords _actually_ emitted, as determined after scheduling */
176 unsigned quadword_count
;
178 /* Successors: always one forward (the block after us), maybe
179 * one backwards (for a backward branch). No need for a second
180 * forward, since graph traversal would get there eventually
182 struct midgard_block
*successors
[2];
183 unsigned nr_successors
;
185 /* The successors pointer form a graph, and in the case of
186 * complex control flow, this graph has a cycles. To aid
187 * traversal during liveness analysis, we have a visited?
188 * boolean for passes to use as they see fit, provided they
194 midgard_block_add_successor(midgard_block
*block
, midgard_block
*successor
)
196 block
->successors
[block
->nr_successors
++] = successor
;
197 assert(block
->nr_successors
<= ARRAY_SIZE(block
->successors
));
200 /* Helpers to generate midgard_instruction's using macro magic, since every
201 * driver seems to do it that way */
203 #define EMIT(op, ...) emit_mir_instruction(ctx, v_##op(__VA_ARGS__));
204 #define SWIZZLE_XYZW SWIZZLE(COMPONENT_X, COMPONENT_Y, COMPONENT_Z, COMPONENT_W)
206 #define M_LOAD_STORE(name, rname, uname) \
207 static midgard_instruction m_##name(unsigned ssa, unsigned address) { \
208 midgard_instruction i = { \
209 .type = TAG_LOAD_STORE_4, \
216 .op = midgard_op_##name, \
218 .swizzle = SWIZZLE_XYZW, \
226 #define M_LOAD(name) M_LOAD_STORE(name, dest, src0)
227 #define M_STORE(name) M_LOAD_STORE(name, src0, dest)
229 const midgard_vector_alu_src blank_alu_src
= {
230 .swizzle
= SWIZZLE(COMPONENT_X
, COMPONENT_Y
, COMPONENT_Z
, COMPONENT_W
),
233 const midgard_vector_alu_src blank_alu_src_xxxx
= {
234 .swizzle
= SWIZZLE(COMPONENT_X
, COMPONENT_X
, COMPONENT_X
, COMPONENT_X
),
237 const midgard_scalar_alu_src blank_scalar_alu_src
= {
241 /* Used for encoding the unused source of 1-op instructions */
242 const midgard_vector_alu_src zero_alu_src
= { 0 };
244 /* Coerce structs to integer */
247 vector_alu_srco_unsigned(midgard_vector_alu_src src
)
250 memcpy(&u
, &src
, sizeof(src
));
254 static midgard_vector_alu_src
255 vector_alu_from_unsigned(unsigned u
)
257 midgard_vector_alu_src s
;
258 memcpy(&s
, &u
, sizeof(s
));
262 /* Inputs a NIR ALU source, with modifiers attached if necessary, and outputs
263 * the corresponding Midgard source */
265 static midgard_vector_alu_src
266 vector_alu_modifiers(nir_alu_src
*src
, bool is_int
)
268 if (!src
) return blank_alu_src
;
270 midgard_vector_alu_src alu_src
= {
273 .half
= 0, /* TODO */
274 .swizzle
= SWIZZLE_FROM_ARRAY(src
->swizzle
)
278 /* TODO: sign-extend/zero-extend */
279 alu_src
.mod
= midgard_int_normal
;
281 /* These should have been lowered away */
282 assert(!(src
->abs
|| src
->negate
));
284 alu_src
.mod
= (src
->abs
<< 0) | (src
->negate
<< 1);
291 mir_nontrivial_mod(midgard_vector_alu_src src
, bool is_int
, unsigned mask
)
294 if (!is_int
&& src
.mod
) return true;
297 for (unsigned c
= 0; c
< 4; ++c
) {
298 if (!(mask
& (1 << c
))) continue;
299 if (((src
.swizzle
>> (2*c
)) & 3) != c
) return true;
305 /* 'Intrinsic' move for misc aliasing uses independent of actual NIR ALU code */
307 static midgard_instruction
308 v_fmov(unsigned src
, midgard_vector_alu_src mod
, unsigned dest
)
310 midgard_instruction ins
= {
313 .src0
= SSA_UNUSED_1
,
318 .op
= midgard_alu_op_fmov
,
319 .reg_mode
= midgard_reg_mode_full
,
320 .dest_override
= midgard_dest_override_none
,
322 .src1
= vector_alu_srco_unsigned(zero_alu_src
),
323 .src2
= vector_alu_srco_unsigned(mod
)
330 /* load/store instructions have both 32-bit and 16-bit variants, depending on
331 * whether we are using vectors composed of highp or mediump. At the moment, we
332 * don't support half-floats -- this requires changes in other parts of the
333 * compiler -- therefore the 16-bit versions are commented out. */
335 //M_LOAD(load_attr_16);
336 M_LOAD(load_attr_32
);
337 //M_LOAD(load_vary_16);
338 M_LOAD(load_vary_32
);
339 //M_LOAD(load_uniform_16);
340 M_LOAD(load_uniform_32
);
341 M_LOAD(load_color_buffer_8
);
342 //M_STORE(store_vary_16);
343 M_STORE(store_vary_32
);
344 M_STORE(store_cubemap_coords
);
346 static midgard_instruction
347 v_alu_br_compact_cond(midgard_jmp_writeout_op op
, unsigned tag
, signed offset
, unsigned cond
)
349 midgard_branch_cond branch
= {
357 memcpy(&compact
, &branch
, sizeof(branch
));
359 midgard_instruction ins
= {
361 .unit
= ALU_ENAB_BR_COMPACT
,
362 .prepacked_branch
= true,
363 .compact_branch
= true,
364 .br_compact
= compact
367 if (op
== midgard_jmp_writeout_op_writeout
)
373 static midgard_instruction
374 v_branch(bool conditional
, bool invert
)
376 midgard_instruction ins
= {
378 .unit
= ALU_ENAB_BRANCH
,
379 .compact_branch
= true,
381 .conditional
= conditional
,
382 .invert_conditional
= invert
389 static midgard_branch_extended
390 midgard_create_branch_extended( midgard_condition cond
,
391 midgard_jmp_writeout_op op
,
393 signed quadword_offset
)
395 /* For unclear reasons, the condition code is repeated 8 times */
396 uint16_t duplicated_cond
=
406 midgard_branch_extended branch
= {
408 .dest_tag
= dest_tag
,
409 .offset
= quadword_offset
,
410 .cond
= duplicated_cond
416 typedef struct midgard_bundle
{
417 /* Tag for the overall bundle */
420 /* Instructions contained by the bundle */
421 int instruction_count
;
422 midgard_instruction instructions
[5];
424 /* Bundle-wide ALU configuration */
427 bool has_embedded_constants
;
429 bool has_blend_constant
;
431 uint16_t register_words
[8];
432 int register_words_count
;
434 uint64_t body_words
[8];
436 int body_words_count
;
439 typedef struct compiler_context
{
441 gl_shader_stage stage
;
443 /* Is internally a blend shader? Depends on stage == FRAGMENT */
446 /* Tracking for blend constant patching */
447 int blend_constant_number
;
448 int blend_constant_offset
;
450 /* Current NIR function */
453 /* Unordered list of midgard_blocks */
455 struct list_head blocks
;
457 midgard_block
*initial_block
;
458 midgard_block
*previous_source_block
;
459 midgard_block
*final_block
;
461 /* List of midgard_instructions emitted for the current block */
462 midgard_block
*current_block
;
464 /* The current "depth" of the loop, for disambiguating breaks/continues
465 * when using nested loops */
466 int current_loop_depth
;
468 /* Constants which have been loaded, for later inlining */
469 struct hash_table_u64
*ssa_constants
;
471 /* SSA indices to be outputted to corresponding varying offset */
472 struct hash_table_u64
*ssa_varyings
;
474 /* SSA values / registers which have been aliased. Naively, these
475 * demand a fmov output; instead, we alias them in a later pass to
476 * avoid the wasted op.
478 * A note on encoding: to avoid dynamic memory management here, rather
479 * than ampping to a pointer, we map to the source index; the key
480 * itself is just the destination index. */
482 struct hash_table_u64
*ssa_to_alias
;
483 struct set
*leftover_ssa_to_alias
;
485 /* Actual SSA-to-register for RA */
486 struct hash_table_u64
*ssa_to_register
;
488 /* Mapping of hashes computed from NIR indices to the sequential temp indices ultimately used in MIR */
489 struct hash_table_u64
*hash_to_temp
;
493 /* Just the count of the max register used. Higher count => higher
494 * register pressure */
497 /* Used for cont/last hinting. Increase when a tex op is added.
498 * Decrease when a tex op is removed. */
499 int texture_op_count
;
501 /* Mapping of texture register -> SSA index for unaliasing */
502 int texture_index
[2];
504 /* If any path hits a discard instruction */
507 /* The number of uniforms allowable for the fast path */
510 /* Count of instructions emitted from NIR overall, across all blocks */
511 int instruction_count
;
513 /* Alpha ref value passed in */
516 /* The index corresponding to the fragment output */
517 unsigned fragment_output
;
519 /* The mapping of sysvals to uniforms, the count, and the off-by-one inverse */
520 unsigned sysvals
[MAX_SYSVAL_COUNT
];
521 unsigned sysval_count
;
522 struct hash_table_u64
*sysval_to_id
;
525 /* Append instruction to end of current block */
527 static midgard_instruction
*
528 mir_upload_ins(struct midgard_instruction ins
)
530 midgard_instruction
*heap
= malloc(sizeof(ins
));
531 memcpy(heap
, &ins
, sizeof(ins
));
536 emit_mir_instruction(struct compiler_context
*ctx
, struct midgard_instruction ins
)
538 list_addtail(&(mir_upload_ins(ins
))->link
, &ctx
->current_block
->instructions
);
542 mir_insert_instruction_before(struct midgard_instruction
*tag
, struct midgard_instruction ins
)
544 list_addtail(&(mir_upload_ins(ins
))->link
, &tag
->link
);
548 mir_remove_instruction(struct midgard_instruction
*ins
)
550 list_del(&ins
->link
);
553 static midgard_instruction
*
554 mir_prev_op(struct midgard_instruction
*ins
)
556 return list_last_entry(&(ins
->link
), midgard_instruction
, link
);
559 static midgard_instruction
*
560 mir_next_op(struct midgard_instruction
*ins
)
562 return list_first_entry(&(ins
->link
), midgard_instruction
, link
);
565 #define mir_foreach_block(ctx, v) list_for_each_entry(struct midgard_block, v, &ctx->blocks, link)
566 #define mir_foreach_block_from(ctx, from, v) list_for_each_entry_from(struct midgard_block, v, from, &ctx->blocks, link)
568 #define mir_foreach_instr(ctx, v) list_for_each_entry(struct midgard_instruction, v, &ctx->current_block->instructions, link)
569 #define mir_foreach_instr_safe(ctx, v) list_for_each_entry_safe(struct midgard_instruction, v, &ctx->current_block->instructions, link)
570 #define mir_foreach_instr_in_block(block, v) list_for_each_entry(struct midgard_instruction, v, &block->instructions, link)
571 #define mir_foreach_instr_in_block_safe(block, v) list_for_each_entry_safe(struct midgard_instruction, v, &block->instructions, link)
572 #define mir_foreach_instr_in_block_safe_rev(block, v) list_for_each_entry_safe_rev(struct midgard_instruction, v, &block->instructions, link)
573 #define mir_foreach_instr_in_block_from(block, v, from) list_for_each_entry_from(struct midgard_instruction, v, from, &block->instructions, link)
574 #define mir_foreach_instr_in_block_from_rev(block, v, from) list_for_each_entry_from_rev(struct midgard_instruction, v, from, &block->instructions, link)
577 static midgard_instruction
*
578 mir_last_in_block(struct midgard_block
*block
)
580 return list_last_entry(&block
->instructions
, struct midgard_instruction
, link
);
583 static midgard_block
*
584 mir_get_block(compiler_context
*ctx
, int idx
)
586 struct list_head
*lst
= &ctx
->blocks
;
591 return (struct midgard_block
*) lst
;
594 /* Pretty printer for internal Midgard IR */
597 print_mir_source(int source
)
599 if (source
>= SSA_FIXED_MINIMUM
) {
600 /* Specific register */
601 int reg
= SSA_REG_FROM_FIXED(source
);
603 /* TODO: Moving threshold */
604 if (reg
> 16 && reg
< 24)
605 printf("u%d", 23 - reg
);
609 printf("%d", source
);
614 print_mir_instruction(midgard_instruction
*ins
)
620 midgard_alu_op op
= ins
->alu
.op
;
621 const char *name
= alu_opcode_props
[op
].name
;
624 printf("%d.", ins
->unit
);
626 printf("%s", name
? name
: "??");
630 case TAG_LOAD_STORE_4
: {
631 midgard_load_store_op op
= ins
->load_store
.op
;
632 const char *name
= load_store_opcode_names
[op
];
639 case TAG_TEXTURE_4
: {
648 ssa_args
*args
= &ins
->ssa_args
;
650 printf(" %d, ", args
->dest
);
652 print_mir_source(args
->src0
);
655 if (args
->inline_constant
)
656 printf("#%d", ins
->inline_constant
);
658 print_mir_source(args
->src1
);
660 if (ins
->has_constants
)
661 printf(" <%f, %f, %f, %f>", ins
->constants
[0], ins
->constants
[1], ins
->constants
[2], ins
->constants
[3]);
667 print_mir_block(midgard_block
*block
)
671 mir_foreach_instr_in_block(block
, ins
) {
672 print_mir_instruction(ins
);
679 attach_constants(compiler_context
*ctx
, midgard_instruction
*ins
, void *constants
, int name
)
681 ins
->has_constants
= true;
682 memcpy(&ins
->constants
, constants
, 16);
684 /* If this is the special blend constant, mark this instruction */
686 if (ctx
->is_blend
&& ctx
->blend_constant_number
== name
)
687 ins
->has_blend_constant
= true;
691 glsl_type_size(const struct glsl_type
*type
, bool bindless
)
693 return glsl_count_attribute_slots(type
, false);
696 /* Lower fdot2 to a vector multiplication followed by channel addition */
698 midgard_nir_lower_fdot2_body(nir_builder
*b
, nir_alu_instr
*alu
)
700 if (alu
->op
!= nir_op_fdot2
)
703 b
->cursor
= nir_before_instr(&alu
->instr
);
705 nir_ssa_def
*src0
= nir_ssa_for_alu_src(b
, alu
, 0);
706 nir_ssa_def
*src1
= nir_ssa_for_alu_src(b
, alu
, 1);
708 nir_ssa_def
*product
= nir_fmul(b
, src0
, src1
);
710 nir_ssa_def
*sum
= nir_fadd(b
,
711 nir_channel(b
, product
, 0),
712 nir_channel(b
, product
, 1));
714 /* Replace the fdot2 with this sum */
715 nir_ssa_def_rewrite_uses(&alu
->dest
.dest
.ssa
, nir_src_for_ssa(sum
));
719 midgard_nir_sysval_for_intrinsic(nir_intrinsic_instr
*instr
)
721 switch (instr
->intrinsic
) {
722 case nir_intrinsic_load_viewport_scale
:
723 return PAN_SYSVAL_VIEWPORT_SCALE
;
724 case nir_intrinsic_load_viewport_offset
:
725 return PAN_SYSVAL_VIEWPORT_OFFSET
;
732 midgard_nir_assign_sysval_body(compiler_context
*ctx
, nir_instr
*instr
)
736 if (instr
->type
== nir_instr_type_intrinsic
) {
737 nir_intrinsic_instr
*intr
= nir_instr_as_intrinsic(instr
);
738 sysval
= midgard_nir_sysval_for_intrinsic(intr
);
744 /* We have a sysval load; check if it's already been assigned */
746 if (_mesa_hash_table_u64_search(ctx
->sysval_to_id
, sysval
))
749 /* It hasn't -- so assign it now! */
751 unsigned id
= ctx
->sysval_count
++;
752 _mesa_hash_table_u64_insert(ctx
->sysval_to_id
, sysval
, (void *) ((uintptr_t) id
+ 1));
753 ctx
->sysvals
[id
] = sysval
;
757 midgard_nir_assign_sysvals(compiler_context
*ctx
, nir_shader
*shader
)
759 ctx
->sysval_count
= 0;
761 nir_foreach_function(function
, shader
) {
762 if (!function
->impl
) continue;
764 nir_foreach_block(block
, function
->impl
) {
765 nir_foreach_instr_safe(instr
, block
) {
766 midgard_nir_assign_sysval_body(ctx
, instr
);
773 midgard_nir_lower_fdot2(nir_shader
*shader
)
775 bool progress
= false;
777 nir_foreach_function(function
, shader
) {
778 if (!function
->impl
) continue;
781 nir_builder
*b
= &_b
;
782 nir_builder_init(b
, function
->impl
);
784 nir_foreach_block(block
, function
->impl
) {
785 nir_foreach_instr_safe(instr
, block
) {
786 if (instr
->type
!= nir_instr_type_alu
) continue;
788 nir_alu_instr
*alu
= nir_instr_as_alu(instr
);
789 midgard_nir_lower_fdot2_body(b
, alu
);
795 nir_metadata_preserve(function
->impl
, nir_metadata_block_index
| nir_metadata_dominance
);
803 optimise_nir(nir_shader
*nir
)
807 NIR_PASS(progress
, nir
, nir_lower_regs_to_ssa
);
808 NIR_PASS(progress
, nir
, midgard_nir_lower_fdot2
);
810 nir_lower_tex_options lower_tex_options
= {
814 NIR_PASS(progress
, nir
, nir_lower_tex
, &lower_tex_options
);
819 NIR_PASS(progress
, nir
, nir_lower_var_copies
);
820 NIR_PASS(progress
, nir
, nir_lower_vars_to_ssa
);
822 NIR_PASS(progress
, nir
, nir_copy_prop
);
823 NIR_PASS(progress
, nir
, nir_opt_dce
);
824 NIR_PASS(progress
, nir
, nir_opt_dead_cf
);
825 NIR_PASS(progress
, nir
, nir_opt_cse
);
826 NIR_PASS(progress
, nir
, nir_opt_peephole_select
, 64, false, true);
827 NIR_PASS(progress
, nir
, nir_opt_algebraic
);
828 NIR_PASS(progress
, nir
, nir_opt_constant_folding
);
829 NIR_PASS(progress
, nir
, nir_opt_undef
);
830 NIR_PASS(progress
, nir
, nir_opt_loop_unroll
,
833 nir_var_function_temp
);
835 /* TODO: Enable vectorize when merged upstream */
836 // NIR_PASS(progress, nir, nir_opt_vectorize);
839 /* Must be run at the end to prevent creation of fsin/fcos ops */
840 NIR_PASS(progress
, nir
, midgard_nir_scale_trig
);
845 NIR_PASS(progress
, nir
, nir_opt_dce
);
846 NIR_PASS(progress
, nir
, nir_opt_algebraic
);
847 NIR_PASS(progress
, nir
, nir_opt_constant_folding
);
848 NIR_PASS(progress
, nir
, nir_copy_prop
);
851 NIR_PASS(progress
, nir
, nir_opt_algebraic_late
);
852 NIR_PASS(progress
, nir
, midgard_nir_lower_algebraic_late
);
854 /* Lower mods for float ops only. Integer ops don't support modifiers
855 * (saturate doesn't make sense on integers, neg/abs require dedicated
858 NIR_PASS(progress
, nir
, nir_lower_to_source_mods
, nir_lower_float_source_mods
);
859 NIR_PASS(progress
, nir
, nir_copy_prop
);
860 NIR_PASS(progress
, nir
, nir_opt_dce
);
862 /* We implement booleans as 32-bit 0/~0 */
863 NIR_PASS(progress
, nir
, nir_lower_bool_to_int32
);
865 /* Take us out of SSA */
866 NIR_PASS(progress
, nir
, nir_lower_locals_to_regs
);
867 NIR_PASS(progress
, nir
, nir_convert_from_ssa
, true);
869 /* We are a vector architecture; write combine where possible */
870 NIR_PASS(progress
, nir
, nir_move_vec_src_uses_to_dest
);
871 NIR_PASS(progress
, nir
, nir_lower_vec_to_movs
);
873 NIR_PASS(progress
, nir
, nir_opt_dce
);
876 /* Front-half of aliasing the SSA slots, merely by inserting the flag in the
877 * appropriate hash table. Intentional off-by-one to avoid confusing NULL with
878 * r0. See the comments in compiler_context */
881 alias_ssa(compiler_context
*ctx
, int dest
, int src
)
883 _mesa_hash_table_u64_insert(ctx
->ssa_to_alias
, dest
+ 1, (void *) ((uintptr_t) src
+ 1));
884 _mesa_set_add(ctx
->leftover_ssa_to_alias
, (void *) (uintptr_t) (dest
+ 1));
887 /* ...or undo it, after which the original index will be used (dummy move should be emitted alongside this) */
890 unalias_ssa(compiler_context
*ctx
, int dest
)
892 _mesa_hash_table_u64_remove(ctx
->ssa_to_alias
, dest
+ 1);
893 /* TODO: Remove from leftover or no? */
897 midgard_pin_output(compiler_context
*ctx
, int index
, int reg
)
899 _mesa_hash_table_u64_insert(ctx
->ssa_to_register
, index
+ 1, (void *) ((uintptr_t) reg
+ 1));
903 midgard_is_pinned(compiler_context
*ctx
, int index
)
905 return _mesa_hash_table_u64_search(ctx
->ssa_to_register
, index
+ 1) != NULL
;
908 /* Do not actually emit a load; instead, cache the constant for inlining */
911 emit_load_const(compiler_context
*ctx
, nir_load_const_instr
*instr
)
913 nir_ssa_def def
= instr
->def
;
915 float *v
= ralloc_array(NULL
, float, 4);
916 nir_const_load_to_arr(v
, instr
, f32
);
917 _mesa_hash_table_u64_insert(ctx
->ssa_constants
, def
.index
+ 1, v
);
920 /* Duplicate bits to convert sane 4-bit writemask to obscure 8-bit format (or
924 expand_writemask(unsigned mask
)
928 for (int i
= 0; i
< 4; ++i
)
936 squeeze_writemask(unsigned mask
)
940 for (int i
= 0; i
< 4; ++i
)
941 if (mask
& (3 << (2 * i
)))
948 /* Determines effective writemask, taking quirks and expansion into account */
950 effective_writemask(midgard_vector_alu
*alu
)
952 /* Channel count is off-by-one to fit in two-bits (0 channel makes no
955 unsigned channel_count
= GET_CHANNEL_COUNT(alu_opcode_props
[alu
->op
].props
);
957 /* If there is a fixed channel count, construct the appropriate mask */
960 return (1 << channel_count
) - 1;
962 /* Otherwise, just squeeze the existing mask */
963 return squeeze_writemask(alu
->mask
);
967 find_or_allocate_temp(compiler_context
*ctx
, unsigned hash
)
969 if ((hash
< 0) || (hash
>= SSA_FIXED_MINIMUM
))
972 unsigned temp
= (uintptr_t) _mesa_hash_table_u64_search(ctx
->hash_to_temp
, hash
+ 1);
977 /* If no temp is find, allocate one */
978 temp
= ctx
->temp_count
++;
979 ctx
->max_hash
= MAX2(ctx
->max_hash
, hash
);
981 _mesa_hash_table_u64_insert(ctx
->hash_to_temp
, hash
+ 1, (void *) ((uintptr_t) temp
+ 1));
987 nir_src_index(compiler_context
*ctx
, nir_src
*src
)
990 return src
->ssa
->index
;
992 assert(!src
->reg
.indirect
);
993 return ctx
->func
->impl
->ssa_alloc
+ src
->reg
.reg
->index
;
998 nir_dest_index(compiler_context
*ctx
, nir_dest
*dst
)
1001 return dst
->ssa
.index
;
1003 assert(!dst
->reg
.indirect
);
1004 return ctx
->func
->impl
->ssa_alloc
+ dst
->reg
.reg
->index
;
1009 nir_alu_src_index(compiler_context
*ctx
, nir_alu_src
*src
)
1011 return nir_src_index(ctx
, &src
->src
);
1014 /* Midgard puts conditionals in r31.w; move an arbitrary source (the output of
1015 * a conditional test) into that register */
1018 emit_condition(compiler_context
*ctx
, nir_src
*src
, bool for_branch
, unsigned component
)
1020 int condition
= nir_src_index(ctx
, src
);
1022 /* Source to swizzle the desired component into w */
1024 const midgard_vector_alu_src alu_src
= {
1025 .swizzle
= SWIZZLE(component
, component
, component
, component
),
1028 /* There is no boolean move instruction. Instead, we simulate a move by
1029 * ANDing the condition with itself to get it into r31.w */
1031 midgard_instruction ins
= {
1033 .unit
= for_branch
? UNIT_SMUL
: UNIT_SADD
, /* TODO: DEDUCE THIS */
1037 .dest
= SSA_FIXED_REGISTER(31),
1040 .op
= midgard_alu_op_iand
,
1041 .reg_mode
= midgard_reg_mode_full
,
1042 .dest_override
= midgard_dest_override_none
,
1043 .mask
= (0x3 << 6), /* w */
1044 .src1
= vector_alu_srco_unsigned(alu_src
),
1045 .src2
= vector_alu_srco_unsigned(alu_src
)
1049 emit_mir_instruction(ctx
, ins
);
1052 /* Likewise, indirect offsets are put in r27.w. TODO: Allow componentwise
1053 * pinning to eliminate this move in all known cases */
1056 emit_indirect_offset(compiler_context
*ctx
, nir_src
*src
)
1058 int offset
= nir_src_index(ctx
, src
);
1060 midgard_instruction ins
= {
1063 .src0
= SSA_UNUSED_1
,
1065 .dest
= SSA_FIXED_REGISTER(REGISTER_OFFSET
),
1068 .op
= midgard_alu_op_imov
,
1069 .reg_mode
= midgard_reg_mode_full
,
1070 .dest_override
= midgard_dest_override_none
,
1071 .mask
= (0x3 << 6), /* w */
1072 .src1
= vector_alu_srco_unsigned(zero_alu_src
),
1073 .src2
= vector_alu_srco_unsigned(blank_alu_src_xxxx
)
1077 emit_mir_instruction(ctx
, ins
);
1080 #define ALU_CASE(nir, _op) \
1081 case nir_op_##nir: \
1082 op = midgard_alu_op_##_op; \
1086 emit_alu(compiler_context
*ctx
, nir_alu_instr
*instr
)
1088 bool is_ssa
= instr
->dest
.dest
.is_ssa
;
1090 unsigned dest
= nir_dest_index(ctx
, &instr
->dest
.dest
);
1091 unsigned nr_components
= is_ssa
? instr
->dest
.dest
.ssa
.num_components
: instr
->dest
.dest
.reg
.reg
->num_components
;
1092 unsigned nr_inputs
= nir_op_infos
[instr
->op
].num_inputs
;
1094 /* Most Midgard ALU ops have a 1:1 correspondance to NIR ops; these are
1095 * supported. A few do not and are commented for now. Also, there are a
1096 * number of NIR ops which Midgard does not support and need to be
1097 * lowered, also TODO. This switch block emits the opcode and calling
1098 * convention of the Midgard instruction; actual packing is done in
1103 switch (instr
->op
) {
1104 ALU_CASE(fadd
, fadd
);
1105 ALU_CASE(fmul
, fmul
);
1106 ALU_CASE(fmin
, fmin
);
1107 ALU_CASE(fmax
, fmax
);
1108 ALU_CASE(imin
, imin
);
1109 ALU_CASE(imax
, imax
);
1110 ALU_CASE(umin
, umin
);
1111 ALU_CASE(umax
, umax
);
1112 ALU_CASE(fmov
, fmov
);
1113 ALU_CASE(ffloor
, ffloor
);
1114 ALU_CASE(fround_even
, froundeven
);
1115 ALU_CASE(ftrunc
, ftrunc
);
1116 ALU_CASE(fceil
, fceil
);
1117 ALU_CASE(fdot3
, fdot3
);
1118 ALU_CASE(fdot4
, fdot4
);
1119 ALU_CASE(iadd
, iadd
);
1120 ALU_CASE(isub
, isub
);
1121 ALU_CASE(imul
, imul
);
1122 ALU_CASE(iabs
, iabs
);
1124 /* XXX: Use fmov, not imov, since imov was causing major
1125 * issues with texture precision? XXX research */
1126 ALU_CASE(imov
, imov
);
1128 ALU_CASE(feq32
, feq
);
1129 ALU_CASE(fne32
, fne
);
1130 ALU_CASE(flt32
, flt
);
1131 ALU_CASE(ieq32
, ieq
);
1132 ALU_CASE(ine32
, ine
);
1133 ALU_CASE(ilt32
, ilt
);
1134 ALU_CASE(ult32
, ult
);
1136 /* We don't have a native b2f32 instruction. Instead, like many
1137 * GPUs, we exploit booleans as 0/~0 for false/true, and
1138 * correspondingly AND
1139 * by 1.0 to do the type conversion. For the moment, prime us
1142 * iand [whatever], #0
1144 * At the end of emit_alu (as MIR), we'll fix-up the constant
1147 ALU_CASE(b2f32
, iand
);
1148 ALU_CASE(b2i32
, iand
);
1150 /* Likewise, we don't have a dedicated f2b32 instruction, but
1151 * we can do a "not equal to 0.0" test. */
1153 ALU_CASE(f2b32
, fne
);
1154 ALU_CASE(i2b32
, ine
);
1156 ALU_CASE(frcp
, frcp
);
1157 ALU_CASE(frsq
, frsqrt
);
1158 ALU_CASE(fsqrt
, fsqrt
);
1159 ALU_CASE(fexp2
, fexp2
);
1160 ALU_CASE(flog2
, flog2
);
1162 ALU_CASE(f2i32
, f2i
);
1163 ALU_CASE(f2u32
, f2u
);
1164 ALU_CASE(i2f32
, i2f
);
1165 ALU_CASE(u2f32
, u2f
);
1167 ALU_CASE(fsin
, fsin
);
1168 ALU_CASE(fcos
, fcos
);
1170 ALU_CASE(iand
, iand
);
1172 ALU_CASE(ixor
, ixor
);
1173 ALU_CASE(inot
, inot
);
1174 ALU_CASE(ishl
, ishl
);
1175 ALU_CASE(ishr
, iasr
);
1176 ALU_CASE(ushr
, ilsr
);
1178 ALU_CASE(b32all_fequal2
, fball_eq
);
1179 ALU_CASE(b32all_fequal3
, fball_eq
);
1180 ALU_CASE(b32all_fequal4
, fball_eq
);
1182 ALU_CASE(b32any_fnequal2
, fbany_neq
);
1183 ALU_CASE(b32any_fnequal3
, fbany_neq
);
1184 ALU_CASE(b32any_fnequal4
, fbany_neq
);
1186 ALU_CASE(b32all_iequal2
, iball_eq
);
1187 ALU_CASE(b32all_iequal3
, iball_eq
);
1188 ALU_CASE(b32all_iequal4
, iball_eq
);
1190 ALU_CASE(b32any_inequal2
, ibany_neq
);
1191 ALU_CASE(b32any_inequal3
, ibany_neq
);
1192 ALU_CASE(b32any_inequal4
, ibany_neq
);
1194 /* For greater-or-equal, we lower to less-or-equal and flip the
1200 case nir_op_uge32
: {
1202 instr
->op
== nir_op_fge
? midgard_alu_op_fle
:
1203 instr
->op
== nir_op_fge32
? midgard_alu_op_fle
:
1204 instr
->op
== nir_op_ige32
? midgard_alu_op_ile
:
1205 instr
->op
== nir_op_uge32
? midgard_alu_op_ule
:
1208 /* Swap via temporary */
1209 nir_alu_src temp
= instr
->src
[1];
1210 instr
->src
[1] = instr
->src
[0];
1211 instr
->src
[0] = temp
;
1216 case nir_op_b32csel
: {
1217 op
= midgard_alu_op_fcsel
;
1219 /* csel works as a two-arg in Midgard, since the condition is hardcoded in r31.w */
1222 /* Figure out which component the condition is in */
1224 unsigned comp
= instr
->src
[0].swizzle
[0];
1226 /* Make sure NIR isn't throwing a mixed condition at us */
1228 for (unsigned c
= 1; c
< nr_components
; ++c
)
1229 assert(instr
->src
[0].swizzle
[c
] == comp
);
1231 /* Emit the condition into r31.w */
1232 emit_condition(ctx
, &instr
->src
[0].src
, false, comp
);
1234 /* The condition is the first argument; move the other
1235 * arguments up one to be a binary instruction for
1238 memmove(instr
->src
, instr
->src
+ 1, 2 * sizeof(nir_alu_src
));
1243 DBG("Unhandled ALU op %s\n", nir_op_infos
[instr
->op
].name
);
1248 /* Fetch unit, quirks, etc information */
1249 unsigned opcode_props
= alu_opcode_props
[op
].props
;
1250 bool quirk_flipped_r24
= opcode_props
& QUIRK_FLIPPED_R24
;
1252 /* Initialise fields common between scalar/vector instructions */
1253 midgard_outmod outmod
= instr
->dest
.saturate
? midgard_outmod_sat
: midgard_outmod_none
;
1255 /* src0 will always exist afaik, but src1 will not for 1-argument
1256 * instructions. The latter can only be fetched if the instruction
1257 * needs it, or else we may segfault. */
1259 unsigned src0
= nir_alu_src_index(ctx
, &instr
->src
[0]);
1260 unsigned src1
= nr_inputs
== 2 ? nir_alu_src_index(ctx
, &instr
->src
[1]) : SSA_UNUSED_0
;
1262 /* Rather than use the instruction generation helpers, we do it
1263 * ourselves here to avoid the mess */
1265 midgard_instruction ins
= {
1268 .src0
= quirk_flipped_r24
? SSA_UNUSED_1
: src0
,
1269 .src1
= quirk_flipped_r24
? src0
: src1
,
1274 nir_alu_src
*nirmods
[2] = { NULL
};
1276 if (nr_inputs
== 2) {
1277 nirmods
[0] = &instr
->src
[0];
1278 nirmods
[1] = &instr
->src
[1];
1279 } else if (nr_inputs
== 1) {
1280 nirmods
[quirk_flipped_r24
] = &instr
->src
[0];
1285 bool is_int
= midgard_is_integer_op(op
);
1287 midgard_vector_alu alu
= {
1289 .reg_mode
= midgard_reg_mode_full
,
1290 .dest_override
= midgard_dest_override_none
,
1293 /* Writemask only valid for non-SSA NIR */
1294 .mask
= expand_writemask((1 << nr_components
) - 1),
1296 .src1
= vector_alu_srco_unsigned(vector_alu_modifiers(nirmods
[0], is_int
)),
1297 .src2
= vector_alu_srco_unsigned(vector_alu_modifiers(nirmods
[1], is_int
)),
1300 /* Apply writemask if non-SSA, keeping in mind that we can't write to components that don't exist */
1303 alu
.mask
&= expand_writemask(instr
->dest
.write_mask
);
1307 /* Late fixup for emulated instructions */
1309 if (instr
->op
== nir_op_b2f32
|| instr
->op
== nir_op_b2i32
) {
1310 /* Presently, our second argument is an inline #0 constant.
1311 * Switch over to an embedded 1.0 constant (that can't fit
1312 * inline, since we're 32-bit, not 16-bit like the inline
1315 ins
.ssa_args
.inline_constant
= false;
1316 ins
.ssa_args
.src1
= SSA_FIXED_REGISTER(REGISTER_CONSTANT
);
1317 ins
.has_constants
= true;
1319 if (instr
->op
== nir_op_b2f32
) {
1320 ins
.constants
[0] = 1.0f
;
1322 /* Type pun it into place */
1324 memcpy(&ins
.constants
[0], &one
, sizeof(uint32_t));
1327 ins
.alu
.src2
= vector_alu_srco_unsigned(blank_alu_src_xxxx
);
1328 } else if (instr
->op
== nir_op_f2b32
|| instr
->op
== nir_op_i2b32
) {
1329 ins
.ssa_args
.inline_constant
= false;
1330 ins
.ssa_args
.src1
= SSA_FIXED_REGISTER(REGISTER_CONSTANT
);
1331 ins
.has_constants
= true;
1332 ins
.constants
[0] = 0.0f
;
1333 ins
.alu
.src2
= vector_alu_srco_unsigned(blank_alu_src_xxxx
);
1336 if ((opcode_props
& UNITS_ALL
) == UNIT_VLUT
) {
1337 /* To avoid duplicating the lookup tables (probably), true LUT
1338 * instructions can only operate as if they were scalars. Lower
1339 * them here by changing the component. */
1341 uint8_t original_swizzle
[4];
1342 memcpy(original_swizzle
, nirmods
[0]->swizzle
, sizeof(nirmods
[0]->swizzle
));
1344 for (int i
= 0; i
< nr_components
; ++i
) {
1345 ins
.alu
.mask
= (0x3) << (2 * i
); /* Mask the associated component */
1347 for (int j
= 0; j
< 4; ++j
)
1348 nirmods
[0]->swizzle
[j
] = original_swizzle
[i
]; /* Pull from the correct component */
1350 ins
.alu
.src1
= vector_alu_srco_unsigned(vector_alu_modifiers(nirmods
[0], is_int
));
1351 emit_mir_instruction(ctx
, ins
);
1354 emit_mir_instruction(ctx
, ins
);
1361 emit_uniform_read(compiler_context
*ctx
, unsigned dest
, unsigned offset
, nir_src
*indirect_offset
)
1363 /* TODO: half-floats */
1365 if (!indirect_offset
&& offset
< ctx
->uniform_cutoff
) {
1366 /* Fast path: For the first 16 uniforms, direct accesses are
1367 * 0-cycle, since they're just a register fetch in the usual
1368 * case. So, we alias the registers while we're still in
1371 int reg_slot
= 23 - offset
;
1372 alias_ssa(ctx
, dest
, SSA_FIXED_REGISTER(reg_slot
));
1374 /* Otherwise, read from the 'special' UBO to access
1375 * higher-indexed uniforms, at a performance cost. More
1376 * generally, we're emitting a UBO read instruction. */
1378 midgard_instruction ins
= m_load_uniform_32(dest
, offset
);
1380 /* TODO: Don't split */
1381 ins
.load_store
.varying_parameters
= (offset
& 7) << 7;
1382 ins
.load_store
.address
= offset
>> 3;
1384 if (indirect_offset
) {
1385 emit_indirect_offset(ctx
, indirect_offset
);
1386 ins
.load_store
.unknown
= 0x8700; /* xxx: what is this? */
1388 ins
.load_store
.unknown
= 0x1E00; /* xxx: what is this? */
1391 emit_mir_instruction(ctx
, ins
);
1396 emit_sysval_read(compiler_context
*ctx
, nir_intrinsic_instr
*instr
)
1398 /* First, pull out the destination */
1399 unsigned dest
= nir_dest_index(ctx
, &instr
->dest
);
1401 /* Now, figure out which uniform this is */
1402 int sysval
= midgard_nir_sysval_for_intrinsic(instr
);
1403 void *val
= _mesa_hash_table_u64_search(ctx
->sysval_to_id
, sysval
);
1405 /* Sysvals are prefix uniforms */
1406 unsigned uniform
= ((uintptr_t) val
) - 1;
1408 /* Emit the read itself -- this is never indirect */
1409 emit_uniform_read(ctx
, dest
, uniform
, NULL
);
1413 emit_intrinsic(compiler_context
*ctx
, nir_intrinsic_instr
*instr
)
1415 unsigned offset
, reg
;
1417 switch (instr
->intrinsic
) {
1418 case nir_intrinsic_discard_if
:
1419 emit_condition(ctx
, &instr
->src
[0], true, COMPONENT_X
);
1423 case nir_intrinsic_discard
: {
1424 bool conditional
= instr
->intrinsic
== nir_intrinsic_discard_if
;
1425 struct midgard_instruction discard
= v_branch(conditional
, false);
1426 discard
.branch
.target_type
= TARGET_DISCARD
;
1427 emit_mir_instruction(ctx
, discard
);
1429 ctx
->can_discard
= true;
1433 case nir_intrinsic_load_uniform
:
1434 case nir_intrinsic_load_input
:
1435 offset
= nir_intrinsic_base(instr
);
1437 bool direct
= nir_src_is_const(instr
->src
[0]);
1440 offset
+= nir_src_as_uint(instr
->src
[0]);
1443 reg
= nir_dest_index(ctx
, &instr
->dest
);
1445 if (instr
->intrinsic
== nir_intrinsic_load_uniform
&& !ctx
->is_blend
) {
1446 emit_uniform_read(ctx
, reg
, ctx
->sysval_count
+ offset
, !direct
? &instr
->src
[0] : NULL
);
1447 } else if (ctx
->stage
== MESA_SHADER_FRAGMENT
&& !ctx
->is_blend
) {
1448 /* XXX: Half-floats? */
1449 /* TODO: swizzle, mask */
1451 midgard_instruction ins
= m_load_vary_32(reg
, offset
);
1453 midgard_varying_parameter p
= {
1455 .interpolation
= midgard_interp_default
,
1456 .flat
= /*var->data.interpolation == INTERP_MODE_FLAT*/ 0
1460 memcpy(&u
, &p
, sizeof(p
));
1461 ins
.load_store
.varying_parameters
= u
;
1464 /* We have the offset totally ready */
1465 ins
.load_store
.unknown
= 0x1e9e; /* xxx: what is this? */
1467 /* We have it partially ready, but we need to
1468 * add in the dynamic index, moved to r27.w */
1469 emit_indirect_offset(ctx
, &instr
->src
[0]);
1470 ins
.load_store
.unknown
= 0x79e; /* xxx: what is this? */
1473 emit_mir_instruction(ctx
, ins
);
1474 } else if (ctx
->is_blend
&& instr
->intrinsic
== nir_intrinsic_load_uniform
) {
1475 /* Constant encoded as a pinned constant */
1477 midgard_instruction ins
= v_fmov(SSA_FIXED_REGISTER(REGISTER_CONSTANT
), blank_alu_src
, reg
);
1478 ins
.has_constants
= true;
1479 ins
.has_blend_constant
= true;
1480 emit_mir_instruction(ctx
, ins
);
1481 } else if (ctx
->is_blend
) {
1482 /* For blend shaders, a load might be
1483 * translated various ways depending on what
1484 * we're loading. Figure out how this is used */
1486 nir_variable
*out
= NULL
;
1488 nir_foreach_variable(var
, &ctx
->nir
->inputs
) {
1489 int drvloc
= var
->data
.driver_location
;
1491 if (nir_intrinsic_base(instr
) == drvloc
) {
1499 if (out
->data
.location
== VARYING_SLOT_COL0
) {
1500 /* Source color preloaded to r0 */
1502 midgard_pin_output(ctx
, reg
, 0);
1503 } else if (out
->data
.location
== VARYING_SLOT_COL1
) {
1504 /* Destination color must be read from framebuffer */
1506 midgard_instruction ins
= m_load_color_buffer_8(reg
, 0);
1507 ins
.load_store
.swizzle
= 0; /* xxxx */
1509 /* Read each component sequentially */
1511 for (int c
= 0; c
< 4; ++c
) {
1512 ins
.load_store
.mask
= (1 << c
);
1513 ins
.load_store
.unknown
= c
;
1514 emit_mir_instruction(ctx
, ins
);
1517 /* vadd.u2f hr2, zext(hr2), #0 */
1519 midgard_vector_alu_src alu_src
= blank_alu_src
;
1520 alu_src
.mod
= midgard_int_zero_extend
;
1521 alu_src
.half
= true;
1523 midgard_instruction u2f
= {
1527 .src1
= SSA_UNUSED_0
,
1529 .inline_constant
= true
1532 .op
= midgard_alu_op_u2f
,
1533 .reg_mode
= midgard_reg_mode_half
,
1534 .dest_override
= midgard_dest_override_none
,
1536 .src1
= vector_alu_srco_unsigned(alu_src
),
1537 .src2
= vector_alu_srco_unsigned(blank_alu_src
),
1541 emit_mir_instruction(ctx
, u2f
);
1543 /* vmul.fmul.sat r1, hr2, #0.00392151 */
1547 midgard_instruction fmul
= {
1549 .inline_constant
= _mesa_float_to_half(1.0 / 255.0),
1553 .src1
= SSA_UNUSED_0
,
1554 .inline_constant
= true
1557 .op
= midgard_alu_op_fmul
,
1558 .reg_mode
= midgard_reg_mode_full
,
1559 .dest_override
= midgard_dest_override_none
,
1560 .outmod
= midgard_outmod_sat
,
1562 .src1
= vector_alu_srco_unsigned(alu_src
),
1563 .src2
= vector_alu_srco_unsigned(blank_alu_src
),
1567 emit_mir_instruction(ctx
, fmul
);
1569 DBG("Unknown input in blend shader\n");
1572 } else if (ctx
->stage
== MESA_SHADER_VERTEX
) {
1573 midgard_instruction ins
= m_load_attr_32(reg
, offset
);
1574 ins
.load_store
.unknown
= 0x1E1E; /* XXX: What is this? */
1575 ins
.load_store
.mask
= (1 << instr
->num_components
) - 1;
1576 emit_mir_instruction(ctx
, ins
);
1578 DBG("Unknown load\n");
1584 case nir_intrinsic_store_output
:
1585 assert(nir_src_is_const(instr
->src
[1]) && "no indirect outputs");
1587 offset
= nir_intrinsic_base(instr
) + nir_src_as_uint(instr
->src
[1]);
1589 reg
= nir_src_index(ctx
, &instr
->src
[0]);
1591 if (ctx
->stage
== MESA_SHADER_FRAGMENT
) {
1592 /* gl_FragColor is not emitted with load/store
1593 * instructions. Instead, it gets plonked into
1594 * r0 at the end of the shader and we do the
1595 * framebuffer writeout dance. TODO: Defer
1598 midgard_pin_output(ctx
, reg
, 0);
1600 /* Save the index we're writing to for later reference
1601 * in the epilogue */
1603 ctx
->fragment_output
= reg
;
1604 } else if (ctx
->stage
== MESA_SHADER_VERTEX
) {
1605 /* Varyings are written into one of two special
1606 * varying register, r26 or r27. The register itself is selected as the register
1607 * in the st_vary instruction, minus the base of 26. E.g. write into r27 and then call st_vary(1)
1609 * Normally emitting fmov's is frowned upon,
1610 * but due to unique constraints of
1611 * REGISTER_VARYING, fmov emission + a
1612 * dedicated cleanup pass is the only way to
1613 * guarantee correctness when considering some
1614 * (common) edge cases XXX: FIXME */
1616 /* If this varying corresponds to a constant (why?!),
1617 * emit that now since it won't get picked up by
1618 * hoisting (since there is no corresponding move
1619 * emitted otherwise) */
1621 void *constant_value
= _mesa_hash_table_u64_search(ctx
->ssa_constants
, reg
+ 1);
1623 if (constant_value
) {
1624 /* Special case: emit the varying write
1625 * directly to r26 (looks funny in asm but it's
1626 * fine) and emit the store _now_. Possibly
1627 * slightly slower, but this is a really stupid
1628 * special case anyway (why on earth would you
1629 * have a constant varying? Your own fault for
1630 * slightly worse perf :P) */
1632 midgard_instruction ins
= v_fmov(SSA_FIXED_REGISTER(REGISTER_CONSTANT
), blank_alu_src
, SSA_FIXED_REGISTER(26));
1633 attach_constants(ctx
, &ins
, constant_value
, reg
+ 1);
1634 emit_mir_instruction(ctx
, ins
);
1636 midgard_instruction st
= m_store_vary_32(SSA_FIXED_REGISTER(0), offset
);
1637 st
.load_store
.unknown
= 0x1E9E; /* XXX: What is this? */
1638 emit_mir_instruction(ctx
, st
);
1640 /* Do not emit the varying yet -- instead, just mark down that we need to later */
1642 _mesa_hash_table_u64_insert(ctx
->ssa_varyings
, reg
+ 1, (void *) ((uintptr_t) (offset
+ 1)));
1645 DBG("Unknown store\n");
1651 case nir_intrinsic_load_alpha_ref_float
:
1652 assert(instr
->dest
.is_ssa
);
1654 float ref_value
= ctx
->alpha_ref
;
1656 float *v
= ralloc_array(NULL
, float, 4);
1657 memcpy(v
, &ref_value
, sizeof(float));
1658 _mesa_hash_table_u64_insert(ctx
->ssa_constants
, instr
->dest
.ssa
.index
+ 1, v
);
1661 case nir_intrinsic_load_viewport_scale
:
1662 case nir_intrinsic_load_viewport_offset
:
1663 emit_sysval_read(ctx
, instr
);
1667 printf ("Unhandled intrinsic\n");
1674 midgard_tex_format(enum glsl_sampler_dim dim
)
1677 case GLSL_SAMPLER_DIM_2D
:
1678 case GLSL_SAMPLER_DIM_EXTERNAL
:
1681 case GLSL_SAMPLER_DIM_3D
:
1684 case GLSL_SAMPLER_DIM_CUBE
:
1685 return TEXTURE_CUBE
;
1688 DBG("Unknown sampler dim type\n");
1695 emit_tex(compiler_context
*ctx
, nir_tex_instr
*instr
)
1698 //assert (!instr->sampler);
1699 //assert (!instr->texture_array_size);
1700 assert (instr
->op
== nir_texop_tex
);
1702 /* Allocate registers via a round robin scheme to alternate between the two registers */
1703 int reg
= ctx
->texture_op_count
& 1;
1704 int in_reg
= reg
, out_reg
= reg
;
1706 /* Make room for the reg */
1708 if (ctx
->texture_index
[reg
] > -1)
1709 unalias_ssa(ctx
, ctx
->texture_index
[reg
]);
1711 int texture_index
= instr
->texture_index
;
1712 int sampler_index
= texture_index
;
1714 for (unsigned i
= 0; i
< instr
->num_srcs
; ++i
) {
1715 switch (instr
->src
[i
].src_type
) {
1716 case nir_tex_src_coord
: {
1717 int index
= nir_src_index(ctx
, &instr
->src
[i
].src
);
1719 midgard_vector_alu_src alu_src
= blank_alu_src
;
1721 int reg
= SSA_FIXED_REGISTER(REGISTER_TEXTURE_BASE
+ in_reg
);
1723 if (instr
->sampler_dim
== GLSL_SAMPLER_DIM_CUBE
) {
1724 /* For cubemaps, we need to load coords into
1725 * special r27, and then use a special ld/st op
1726 * to copy into the texture register */
1728 alu_src
.swizzle
= SWIZZLE(COMPONENT_X
, COMPONENT_Y
, COMPONENT_Z
, COMPONENT_X
);
1730 midgard_instruction move
= v_fmov(index
, alu_src
, SSA_FIXED_REGISTER(27));
1731 emit_mir_instruction(ctx
, move
);
1733 midgard_instruction st
= m_store_cubemap_coords(reg
, 0);
1734 st
.load_store
.unknown
= 0x24; /* XXX: What is this? */
1735 st
.load_store
.mask
= 0x3; /* xy? */
1736 st
.load_store
.swizzle
= alu_src
.swizzle
;
1737 emit_mir_instruction(ctx
, st
);
1740 alu_src
.swizzle
= SWIZZLE(COMPONENT_X
, COMPONENT_Y
, COMPONENT_X
, COMPONENT_X
);
1742 midgard_instruction ins
= v_fmov(index
, alu_src
, reg
);
1743 emit_mir_instruction(ctx
, ins
);
1746 //midgard_pin_output(ctx, index, REGISTER_TEXTURE_BASE + in_reg);
1752 DBG("Unknown source type\n");
1759 /* No helper to build texture words -- we do it all here */
1760 midgard_instruction ins
= {
1761 .type
= TAG_TEXTURE_4
,
1763 .op
= TEXTURE_OP_NORMAL
,
1764 .format
= midgard_tex_format(instr
->sampler_dim
),
1765 .texture_handle
= texture_index
,
1766 .sampler_handle
= sampler_index
,
1768 /* TODO: Don't force xyzw */
1769 .swizzle
= SWIZZLE(COMPONENT_X
, COMPONENT_Y
, COMPONENT_Z
, COMPONENT_W
),
1781 /* Assume we can continue; hint it out later */
1786 /* Set registers to read and write from the same place */
1787 ins
.texture
.in_reg_select
= in_reg
;
1788 ins
.texture
.out_reg_select
= out_reg
;
1790 /* TODO: Dynamic swizzle input selection, half-swizzles? */
1791 if (instr
->sampler_dim
== GLSL_SAMPLER_DIM_3D
) {
1792 ins
.texture
.in_reg_swizzle_right
= COMPONENT_X
;
1793 ins
.texture
.in_reg_swizzle_left
= COMPONENT_Y
;
1794 //ins.texture.in_reg_swizzle_third = COMPONENT_Z;
1796 ins
.texture
.in_reg_swizzle_left
= COMPONENT_X
;
1797 ins
.texture
.in_reg_swizzle_right
= COMPONENT_Y
;
1798 //ins.texture.in_reg_swizzle_third = COMPONENT_X;
1801 emit_mir_instruction(ctx
, ins
);
1803 /* Simultaneously alias the destination and emit a move for it. The move will be eliminated if possible */
1805 int o_reg
= REGISTER_TEXTURE_BASE
+ out_reg
, o_index
= nir_dest_index(ctx
, &instr
->dest
);
1806 alias_ssa(ctx
, o_index
, SSA_FIXED_REGISTER(o_reg
));
1807 ctx
->texture_index
[reg
] = o_index
;
1809 midgard_instruction ins2
= v_fmov(SSA_FIXED_REGISTER(o_reg
), blank_alu_src
, o_index
);
1810 emit_mir_instruction(ctx
, ins2
);
1812 /* Used for .cont and .last hinting */
1813 ctx
->texture_op_count
++;
1817 emit_jump(compiler_context
*ctx
, nir_jump_instr
*instr
)
1819 switch (instr
->type
) {
1820 case nir_jump_break
: {
1821 /* Emit a branch out of the loop */
1822 struct midgard_instruction br
= v_branch(false, false);
1823 br
.branch
.target_type
= TARGET_BREAK
;
1824 br
.branch
.target_break
= ctx
->current_loop_depth
;
1825 emit_mir_instruction(ctx
, br
);
1832 DBG("Unknown jump type %d\n", instr
->type
);
1838 emit_instr(compiler_context
*ctx
, struct nir_instr
*instr
)
1840 switch (instr
->type
) {
1841 case nir_instr_type_load_const
:
1842 emit_load_const(ctx
, nir_instr_as_load_const(instr
));
1845 case nir_instr_type_intrinsic
:
1846 emit_intrinsic(ctx
, nir_instr_as_intrinsic(instr
));
1849 case nir_instr_type_alu
:
1850 emit_alu(ctx
, nir_instr_as_alu(instr
));
1853 case nir_instr_type_tex
:
1854 emit_tex(ctx
, nir_instr_as_tex(instr
));
1857 case nir_instr_type_jump
:
1858 emit_jump(ctx
, nir_instr_as_jump(instr
));
1861 case nir_instr_type_ssa_undef
:
1866 DBG("Unhandled instruction type\n");
1871 /* Determine the actual hardware from the index based on the RA results or special values */
1874 dealias_register(compiler_context
*ctx
, struct ra_graph
*g
, int reg
, int maxreg
)
1876 if (reg
>= SSA_FIXED_MINIMUM
)
1877 return SSA_REG_FROM_FIXED(reg
);
1880 assert(reg
< maxreg
);
1881 int r
= ra_get_node_reg(g
, reg
);
1882 ctx
->work_registers
= MAX2(ctx
->work_registers
, r
);
1887 /* fmov style unused */
1889 return REGISTER_UNUSED
;
1891 /* lut style unused */
1893 return REGISTER_UNUSED
;
1896 DBG("Unknown SSA register alias %d\n", reg
);
1903 midgard_ra_select_callback(struct ra_graph
*g
, BITSET_WORD
*regs
, void *data
)
1905 /* Choose the first available register to minimise reported register pressure */
1907 for (int i
= 0; i
< 16; ++i
) {
1908 if (BITSET_TEST(regs
, i
)) {
1918 midgard_is_live_in_instr(midgard_instruction
*ins
, int src
)
1920 if (ins
->ssa_args
.src0
== src
) return true;
1921 if (ins
->ssa_args
.src1
== src
) return true;
1926 /* Determine if a variable is live in the successors of a block */
1928 is_live_after_successors(compiler_context
*ctx
, midgard_block
*bl
, int src
)
1930 for (unsigned i
= 0; i
< bl
->nr_successors
; ++i
) {
1931 midgard_block
*succ
= bl
->successors
[i
];
1933 /* If we already visited, the value we're seeking
1934 * isn't down this path (or we would have short
1937 if (succ
->visited
) continue;
1939 /* Otherwise (it's visited *now*), check the block */
1941 succ
->visited
= true;
1943 mir_foreach_instr_in_block(succ
, ins
) {
1944 if (midgard_is_live_in_instr(ins
, src
))
1948 /* ...and also, check *its* successors */
1949 if (is_live_after_successors(ctx
, succ
, src
))
1954 /* Welp. We're really not live. */
1960 is_live_after(compiler_context
*ctx
, midgard_block
*block
, midgard_instruction
*start
, int src
)
1962 /* Check the rest of the block for liveness */
1964 mir_foreach_instr_in_block_from(block
, ins
, mir_next_op(start
)) {
1965 if (midgard_is_live_in_instr(ins
, src
))
1969 /* Check the rest of the blocks for liveness recursively */
1971 bool succ
= is_live_after_successors(ctx
, block
, src
);
1973 mir_foreach_block(ctx
, block
) {
1974 block
->visited
= false;
1981 allocate_registers(compiler_context
*ctx
)
1983 /* First, initialize the RA */
1984 struct ra_regs
*regs
= ra_alloc_reg_set(NULL
, 32, true);
1986 /* Create a primary (general purpose) class, as well as special purpose
1987 * pipeline register classes */
1989 int primary_class
= ra_alloc_reg_class(regs
);
1990 int varying_class
= ra_alloc_reg_class(regs
);
1992 /* Add the full set of work registers */
1993 int work_count
= 16 - MAX2((ctx
->uniform_cutoff
- 8), 0);
1994 for (int i
= 0; i
< work_count
; ++i
)
1995 ra_class_add_reg(regs
, primary_class
, i
);
1997 /* Add special registers */
1998 ra_class_add_reg(regs
, varying_class
, REGISTER_VARYING_BASE
);
1999 ra_class_add_reg(regs
, varying_class
, REGISTER_VARYING_BASE
+ 1);
2001 /* We're done setting up */
2002 ra_set_finalize(regs
, NULL
);
2004 /* Transform the MIR into squeezed index form */
2005 mir_foreach_block(ctx
, block
) {
2006 mir_foreach_instr_in_block(block
, ins
) {
2007 if (ins
->compact_branch
) continue;
2009 ins
->ssa_args
.src0
= find_or_allocate_temp(ctx
, ins
->ssa_args
.src0
);
2010 ins
->ssa_args
.src1
= find_or_allocate_temp(ctx
, ins
->ssa_args
.src1
);
2011 ins
->ssa_args
.dest
= find_or_allocate_temp(ctx
, ins
->ssa_args
.dest
);
2013 if (midgard_debug
& MIDGARD_DBG_SHADERS
)
2014 print_mir_block(block
);
2017 /* Let's actually do register allocation */
2018 int nodes
= ctx
->temp_count
;
2019 struct ra_graph
*g
= ra_alloc_interference_graph(regs
, nodes
);
2021 /* Set everything to the work register class, unless it has somewhere
2024 mir_foreach_block(ctx
, block
) {
2025 mir_foreach_instr_in_block(block
, ins
) {
2026 if (ins
->compact_branch
) continue;
2028 if (ins
->ssa_args
.dest
< 0) continue;
2030 if (ins
->ssa_args
.dest
>= SSA_FIXED_MINIMUM
) continue;
2032 int class = primary_class
;
2034 ra_set_node_class(g
, ins
->ssa_args
.dest
, class);
2038 for (int index
= 0; index
<= ctx
->max_hash
; ++index
) {
2039 unsigned temp
= (uintptr_t) _mesa_hash_table_u64_search(ctx
->ssa_to_register
, index
+ 1);
2042 unsigned reg
= temp
- 1;
2043 int t
= find_or_allocate_temp(ctx
, index
);
2044 ra_set_node_reg(g
, t
, reg
);
2048 /* Determine liveness */
2050 int *live_start
= malloc(nodes
* sizeof(int));
2051 int *live_end
= malloc(nodes
* sizeof(int));
2053 /* Initialize as non-existent */
2055 for (int i
= 0; i
< nodes
; ++i
) {
2056 live_start
[i
] = live_end
[i
] = -1;
2061 mir_foreach_block(ctx
, block
) {
2062 mir_foreach_instr_in_block(block
, ins
) {
2063 if (ins
->compact_branch
) continue;
2065 if (ins
->ssa_args
.dest
< SSA_FIXED_MINIMUM
) {
2066 /* If this destination is not yet live, it is now since we just wrote it */
2068 int dest
= ins
->ssa_args
.dest
;
2070 if (live_start
[dest
] == -1)
2071 live_start
[dest
] = d
;
2074 /* Since we just used a source, the source might be
2075 * dead now. Scan the rest of the block for
2076 * invocations, and if there are none, the source dies
2079 int sources
[2] = { ins
->ssa_args
.src0
, ins
->ssa_args
.src1
};
2081 for (int src
= 0; src
< 2; ++src
) {
2082 int s
= sources
[src
];
2084 if (s
< 0) continue;
2086 if (s
>= SSA_FIXED_MINIMUM
) continue;
2088 if (!is_live_after(ctx
, block
, ins
, s
)) {
2097 /* If a node still hasn't been killed, kill it now */
2099 for (int i
= 0; i
< nodes
; ++i
) {
2100 /* live_start == -1 most likely indicates a pinned output */
2102 if (live_end
[i
] == -1)
2106 /* Setup interference between nodes that are live at the same time */
2108 for (int i
= 0; i
< nodes
; ++i
) {
2109 for (int j
= i
+ 1; j
< nodes
; ++j
) {
2110 if (!(live_start
[i
] >= live_end
[j
] || live_start
[j
] >= live_end
[i
]))
2111 ra_add_node_interference(g
, i
, j
);
2115 ra_set_select_reg_callback(g
, midgard_ra_select_callback
, NULL
);
2117 if (!ra_allocate(g
)) {
2118 DBG("Error allocating registers\n");
2126 mir_foreach_block(ctx
, block
) {
2127 mir_foreach_instr_in_block(block
, ins
) {
2128 if (ins
->compact_branch
) continue;
2130 ssa_args args
= ins
->ssa_args
;
2132 switch (ins
->type
) {
2134 ins
->registers
.src1_reg
= dealias_register(ctx
, g
, args
.src0
, nodes
);
2136 ins
->registers
.src2_imm
= args
.inline_constant
;
2138 if (args
.inline_constant
) {
2139 /* Encode inline 16-bit constant as a vector by default */
2141 ins
->registers
.src2_reg
= ins
->inline_constant
>> 11;
2143 int lower_11
= ins
->inline_constant
& ((1 << 12) - 1);
2145 uint16_t imm
= ((lower_11
>> 8) & 0x7) | ((lower_11
& 0xFF) << 3);
2146 ins
->alu
.src2
= imm
<< 2;
2148 ins
->registers
.src2_reg
= dealias_register(ctx
, g
, args
.src1
, nodes
);
2151 ins
->registers
.out_reg
= dealias_register(ctx
, g
, args
.dest
, nodes
);
2155 case TAG_LOAD_STORE_4
: {
2156 if (OP_IS_STORE_VARY(ins
->load_store
.op
)) {
2157 /* TODO: use ssa_args for store_vary */
2158 ins
->load_store
.reg
= 0;
2160 bool has_dest
= args
.dest
>= 0;
2161 int ssa_arg
= has_dest
? args
.dest
: args
.src0
;
2163 ins
->load_store
.reg
= dealias_register(ctx
, g
, ssa_arg
, nodes
);
2176 /* Midgard IR only knows vector ALU types, but we sometimes need to actually
2177 * use scalar ALU instructions, for functional or performance reasons. To do
2178 * this, we just demote vector ALU payloads to scalar. */
2181 component_from_mask(unsigned mask
)
2183 for (int c
= 0; c
< 4; ++c
) {
2184 if (mask
& (3 << (2 * c
)))
2193 is_single_component_mask(unsigned mask
)
2197 for (int c
= 0; c
< 4; ++c
)
2198 if (mask
& (3 << (2 * c
)))
2201 return components
== 1;
2204 /* Create a mask of accessed components from a swizzle to figure out vector
2208 swizzle_to_access_mask(unsigned swizzle
)
2210 unsigned component_mask
= 0;
2212 for (int i
= 0; i
< 4; ++i
) {
2213 unsigned c
= (swizzle
>> (2 * i
)) & 3;
2214 component_mask
|= (1 << c
);
2217 return component_mask
;
2221 vector_to_scalar_source(unsigned u
, bool is_int
)
2223 midgard_vector_alu_src v
;
2224 memcpy(&v
, &u
, sizeof(v
));
2226 /* TODO: Integers */
2228 midgard_scalar_alu_src s
= {
2230 .component
= (v
.swizzle
& 3) << 1
2236 s
.abs
= v
.mod
& MIDGARD_FLOAT_MOD_ABS
;
2237 s
.negate
= v
.mod
& MIDGARD_FLOAT_MOD_NEG
;
2241 memcpy(&o
, &s
, sizeof(s
));
2243 return o
& ((1 << 6) - 1);
2246 static midgard_scalar_alu
2247 vector_to_scalar_alu(midgard_vector_alu v
, midgard_instruction
*ins
)
2249 bool is_int
= midgard_is_integer_op(v
.op
);
2251 /* The output component is from the mask */
2252 midgard_scalar_alu s
= {
2254 .src1
= vector_to_scalar_source(v
.src1
, is_int
),
2255 .src2
= vector_to_scalar_source(v
.src2
, is_int
),
2258 .output_full
= 1, /* TODO: Half */
2259 .output_component
= component_from_mask(v
.mask
) << 1,
2262 /* Inline constant is passed along rather than trying to extract it
2265 if (ins
->ssa_args
.inline_constant
) {
2267 int lower_11
= ins
->inline_constant
& ((1 << 12) - 1);
2268 imm
|= (lower_11
>> 9) & 3;
2269 imm
|= (lower_11
>> 6) & 4;
2270 imm
|= (lower_11
>> 2) & 0x38;
2271 imm
|= (lower_11
& 63) << 6;
2279 /* Midgard prefetches instruction types, so during emission we need to
2280 * lookahead too. Unless this is the last instruction, in which we return 1. Or
2281 * if this is the second to last and the last is an ALU, then it's also 1... */
2283 #define IS_ALU(tag) (tag == TAG_ALU_4 || tag == TAG_ALU_8 || \
2284 tag == TAG_ALU_12 || tag == TAG_ALU_16)
2286 #define EMIT_AND_COUNT(type, val) util_dynarray_append(emission, type, val); \
2287 bytes_emitted += sizeof(type)
2290 emit_binary_vector_instruction(midgard_instruction
*ains
,
2291 uint16_t *register_words
, int *register_words_count
,
2292 uint64_t *body_words
, size_t *body_size
, int *body_words_count
,
2293 size_t *bytes_emitted
)
2295 memcpy(®ister_words
[(*register_words_count
)++], &ains
->registers
, sizeof(ains
->registers
));
2296 *bytes_emitted
+= sizeof(midgard_reg_info
);
2298 body_size
[*body_words_count
] = sizeof(midgard_vector_alu
);
2299 memcpy(&body_words
[(*body_words_count
)++], &ains
->alu
, sizeof(ains
->alu
));
2300 *bytes_emitted
+= sizeof(midgard_vector_alu
);
2303 /* Checks for an SSA data hazard between two adjacent instructions, keeping in
2304 * mind that we are a vector architecture and we can write to different
2305 * components simultaneously */
2308 can_run_concurrent_ssa(midgard_instruction
*first
, midgard_instruction
*second
)
2310 /* Each instruction reads some registers and writes to a register. See
2311 * where the first writes */
2313 /* Figure out where exactly we wrote to */
2314 int source
= first
->ssa_args
.dest
;
2315 int source_mask
= first
->type
== TAG_ALU_4
? squeeze_writemask(first
->alu
.mask
) : 0xF;
2317 /* As long as the second doesn't read from the first, we're okay */
2318 if (second
->ssa_args
.src0
== source
) {
2319 if (first
->type
== TAG_ALU_4
) {
2320 /* Figure out which components we just read from */
2322 int q
= second
->alu
.src1
;
2323 midgard_vector_alu_src
*m
= (midgard_vector_alu_src
*) &q
;
2325 /* Check if there are components in common, and fail if so */
2326 if (swizzle_to_access_mask(m
->swizzle
) & source_mask
)
2333 if (second
->ssa_args
.src1
== source
)
2336 /* Otherwise, it's safe in that regard. Another data hazard is both
2337 * writing to the same place, of course */
2339 if (second
->ssa_args
.dest
== source
) {
2340 /* ...but only if the components overlap */
2341 int dest_mask
= second
->type
== TAG_ALU_4
? squeeze_writemask(second
->alu
.mask
) : 0xF;
2343 if (dest_mask
& source_mask
)
2353 midgard_instruction
**segment
, unsigned segment_size
,
2354 midgard_instruction
*ains
)
2356 for (int s
= 0; s
< segment_size
; ++s
)
2357 if (!can_run_concurrent_ssa(segment
[s
], ains
))
2365 /* Schedules, but does not emit, a single basic block. After scheduling, the
2366 * final tag and size of the block are known, which are necessary for branching
2369 static midgard_bundle
2370 schedule_bundle(compiler_context
*ctx
, midgard_block
*block
, midgard_instruction
*ins
, int *skip
)
2372 int instructions_emitted
= 0, instructions_consumed
= -1;
2373 midgard_bundle bundle
= { 0 };
2375 uint8_t tag
= ins
->type
;
2377 /* Default to the instruction's tag */
2380 switch (ins
->type
) {
2382 uint32_t control
= 0;
2383 size_t bytes_emitted
= sizeof(control
);
2385 /* TODO: Constant combining */
2386 int index
= 0, last_unit
= 0;
2388 /* Previous instructions, for the purpose of parallelism */
2389 midgard_instruction
*segment
[4] = {0};
2390 int segment_size
= 0;
2392 instructions_emitted
= -1;
2393 midgard_instruction
*pins
= ins
;
2396 midgard_instruction
*ains
= pins
;
2398 /* Advance instruction pointer */
2400 ains
= mir_next_op(pins
);
2404 /* Out-of-work condition */
2405 if ((struct list_head
*) ains
== &block
->instructions
)
2408 /* Ensure that the chain can continue */
2409 if (ains
->type
!= TAG_ALU_4
) break;
2411 /* According to the presentation "The ARM
2412 * Mali-T880 Mobile GPU" from HotChips 27,
2413 * there are two pipeline stages. Branching
2414 * position determined experimentally. Lines
2415 * are executed in parallel:
2418 * [ VADD ] [ SMUL ] [ LUT ] [ BRANCH ]
2420 * Verify that there are no ordering dependencies here.
2422 * TODO: Allow for parallelism!!!
2425 /* Pick a unit for it if it doesn't force a particular unit */
2427 int unit
= ains
->unit
;
2430 int op
= ains
->alu
.op
;
2431 int units
= alu_opcode_props
[op
].props
;
2433 /* TODO: Promotion of scalars to vectors */
2434 int vector
= ((!is_single_component_mask(ains
->alu
.mask
)) || ((units
& UNITS_SCALAR
) == 0)) && (units
& UNITS_ANY_VECTOR
);
2437 assert(units
& UNITS_SCALAR
);
2440 if (last_unit
>= UNIT_VADD
) {
2441 if (units
& UNIT_VLUT
)
2446 if ((units
& UNIT_VMUL
) && !(control
& UNIT_VMUL
))
2448 else if ((units
& UNIT_VADD
) && !(control
& UNIT_VADD
))
2450 else if (units
& UNIT_VLUT
)
2456 if (last_unit
>= UNIT_VADD
) {
2457 if ((units
& UNIT_SMUL
) && !(control
& UNIT_SMUL
))
2459 else if (units
& UNIT_VLUT
)
2464 if ((units
& UNIT_SADD
) && !(control
& UNIT_SADD
) && !midgard_has_hazard(segment
, segment_size
, ains
))
2466 else if (units
& UNIT_SMUL
)
2467 unit
= ((units
& UNIT_VMUL
) && !(control
& UNIT_VMUL
)) ? UNIT_VMUL
: UNIT_SMUL
;
2468 else if ((units
& UNIT_VADD
) && !(control
& UNIT_VADD
))
2475 assert(unit
& units
);
2478 /* Late unit check, this time for encoding (not parallelism) */
2479 if (unit
<= last_unit
) break;
2481 /* Clear the segment */
2482 if (last_unit
< UNIT_VADD
&& unit
>= UNIT_VADD
)
2485 if (midgard_has_hazard(segment
, segment_size
, ains
))
2488 /* We're good to go -- emit the instruction */
2491 segment
[segment_size
++] = ains
;
2493 /* Only one set of embedded constants per
2494 * bundle possible; if we have more, we must
2495 * break the chain early, unfortunately */
2497 if (ains
->has_constants
) {
2498 if (bundle
.has_embedded_constants
) {
2499 /* ...but if there are already
2500 * constants but these are the
2501 * *same* constants, we let it
2504 if (memcmp(bundle
.constants
, ains
->constants
, sizeof(bundle
.constants
)))
2507 bundle
.has_embedded_constants
= true;
2508 memcpy(bundle
.constants
, ains
->constants
, sizeof(bundle
.constants
));
2510 /* If this is a blend shader special constant, track it for patching */
2511 if (ains
->has_blend_constant
)
2512 bundle
.has_blend_constant
= true;
2516 if (ains
->unit
& UNITS_ANY_VECTOR
) {
2517 emit_binary_vector_instruction(ains
, bundle
.register_words
,
2518 &bundle
.register_words_count
, bundle
.body_words
,
2519 bundle
.body_size
, &bundle
.body_words_count
, &bytes_emitted
);
2520 } else if (ains
->compact_branch
) {
2521 /* All of r0 has to be written out
2522 * along with the branch writeout.
2525 if (ains
->writeout
) {
2527 midgard_instruction ins
= v_fmov(0, blank_alu_src
, SSA_FIXED_REGISTER(0));
2528 ins
.unit
= UNIT_VMUL
;
2530 control
|= ins
.unit
;
2532 emit_binary_vector_instruction(&ins
, bundle
.register_words
,
2533 &bundle
.register_words_count
, bundle
.body_words
,
2534 bundle
.body_size
, &bundle
.body_words_count
, &bytes_emitted
);
2536 /* Analyse the group to see if r0 is written in full, on-time, without hanging dependencies*/
2537 bool written_late
= false;
2538 bool components
[4] = { 0 };
2539 uint16_t register_dep_mask
= 0;
2540 uint16_t written_mask
= 0;
2542 midgard_instruction
*qins
= ins
;
2543 for (int t
= 0; t
< index
; ++t
) {
2544 if (qins
->registers
.out_reg
!= 0) {
2545 /* Mark down writes */
2547 written_mask
|= (1 << qins
->registers
.out_reg
);
2549 /* Mark down the register dependencies for errata check */
2551 if (qins
->registers
.src1_reg
< 16)
2552 register_dep_mask
|= (1 << qins
->registers
.src1_reg
);
2554 if (qins
->registers
.src2_reg
< 16)
2555 register_dep_mask
|= (1 << qins
->registers
.src2_reg
);
2557 int mask
= qins
->alu
.mask
;
2559 for (int c
= 0; c
< 4; ++c
)
2560 if (mask
& (0x3 << (2 * c
)))
2561 components
[c
] = true;
2563 /* ..but if the writeout is too late, we have to break up anyway... for some reason */
2565 if (qins
->unit
== UNIT_VLUT
)
2566 written_late
= true;
2569 /* Advance instruction pointer */
2570 qins
= mir_next_op(qins
);
2574 /* ERRATA (?): In a bundle ending in a fragment writeout, the register dependencies of r0 cannot be written within this bundle (discovered in -bshading:shading=phong) */
2575 if (register_dep_mask
& written_mask
) {
2576 DBG("ERRATA WORKAROUND: Breakup for writeout dependency masks %X vs %X (common %X)\n", register_dep_mask
, written_mask
, register_dep_mask
& written_mask
);
2583 /* If even a single component is not written, break it up (conservative check). */
2584 bool breakup
= false;
2586 for (int c
= 0; c
< 4; ++c
)
2593 /* Otherwise, we're free to proceed */
2597 if (ains
->unit
== ALU_ENAB_BRANCH
) {
2598 bundle
.body_size
[bundle
.body_words_count
] = sizeof(midgard_branch_extended
);
2599 memcpy(&bundle
.body_words
[bundle
.body_words_count
++], &ains
->branch_extended
, sizeof(midgard_branch_extended
));
2600 bytes_emitted
+= sizeof(midgard_branch_extended
);
2602 bundle
.body_size
[bundle
.body_words_count
] = sizeof(ains
->br_compact
);
2603 memcpy(&bundle
.body_words
[bundle
.body_words_count
++], &ains
->br_compact
, sizeof(ains
->br_compact
));
2604 bytes_emitted
+= sizeof(ains
->br_compact
);
2607 memcpy(&bundle
.register_words
[bundle
.register_words_count
++], &ains
->registers
, sizeof(ains
->registers
));
2608 bytes_emitted
+= sizeof(midgard_reg_info
);
2610 bundle
.body_size
[bundle
.body_words_count
] = sizeof(midgard_scalar_alu
);
2611 bundle
.body_words_count
++;
2612 bytes_emitted
+= sizeof(midgard_scalar_alu
);
2615 /* Defer marking until after writing to allow for break */
2616 control
|= ains
->unit
;
2617 last_unit
= ains
->unit
;
2618 ++instructions_emitted
;
2622 /* Bubble up the number of instructions for skipping */
2623 instructions_consumed
= index
- 1;
2627 /* Pad ALU op to nearest word */
2629 if (bytes_emitted
& 15) {
2630 padding
= 16 - (bytes_emitted
& 15);
2631 bytes_emitted
+= padding
;
2634 /* Constants must always be quadwords */
2635 if (bundle
.has_embedded_constants
)
2636 bytes_emitted
+= 16;
2638 /* Size ALU instruction for tag */
2639 bundle
.tag
= (TAG_ALU_4
) + (bytes_emitted
/ 16) - 1;
2640 bundle
.padding
= padding
;
2641 bundle
.control
= bundle
.tag
| control
;
2646 case TAG_LOAD_STORE_4
: {
2647 /* Load store instructions have two words at once. If
2648 * we only have one queued up, we need to NOP pad.
2649 * Otherwise, we store both in succession to save space
2650 * and cycles -- letting them go in parallel -- skip
2651 * the next. The usefulness of this optimisation is
2652 * greatly dependent on the quality of the instruction
2656 midgard_instruction
*next_op
= mir_next_op(ins
);
2658 if ((struct list_head
*) next_op
!= &block
->instructions
&& next_op
->type
== TAG_LOAD_STORE_4
) {
2659 /* As the two operate concurrently, make sure
2660 * they are not dependent */
2662 if (can_run_concurrent_ssa(ins
, next_op
) || true) {
2663 /* Skip ahead, since it's redundant with the pair */
2664 instructions_consumed
= 1 + (instructions_emitted
++);
2672 /* Texture ops default to single-op-per-bundle scheduling */
2676 /* Copy the instructions into the bundle */
2677 bundle
.instruction_count
= instructions_emitted
+ 1;
2681 midgard_instruction
*uins
= ins
;
2682 for (int i
= 0; used_idx
< bundle
.instruction_count
; ++i
) {
2683 bundle
.instructions
[used_idx
++] = *uins
;
2684 uins
= mir_next_op(uins
);
2687 *skip
= (instructions_consumed
== -1) ? instructions_emitted
: instructions_consumed
;
2693 quadword_size(int tag
)
2708 case TAG_LOAD_STORE_4
:
2720 /* Schedule a single block by iterating its instruction to create bundles.
2721 * While we go, tally about the bundle sizes to compute the block size. */
2724 schedule_block(compiler_context
*ctx
, midgard_block
*block
)
2726 util_dynarray_init(&block
->bundles
, NULL
);
2728 block
->quadword_count
= 0;
2730 mir_foreach_instr_in_block(block
, ins
) {
2732 midgard_bundle bundle
= schedule_bundle(ctx
, block
, ins
, &skip
);
2733 util_dynarray_append(&block
->bundles
, midgard_bundle
, bundle
);
2735 if (bundle
.has_blend_constant
) {
2736 /* TODO: Multiblock? */
2737 int quadwords_within_block
= block
->quadword_count
+ quadword_size(bundle
.tag
) - 1;
2738 ctx
->blend_constant_offset
= quadwords_within_block
* 0x10;
2742 ins
= mir_next_op(ins
);
2744 block
->quadword_count
+= quadword_size(bundle
.tag
);
2747 block
->is_scheduled
= true;
2751 schedule_program(compiler_context
*ctx
)
2753 allocate_registers(ctx
);
2755 mir_foreach_block(ctx
, block
) {
2756 schedule_block(ctx
, block
);
2760 /* After everything is scheduled, emit whole bundles at a time */
2763 emit_binary_bundle(compiler_context
*ctx
, midgard_bundle
*bundle
, struct util_dynarray
*emission
, int next_tag
)
2765 int lookahead
= next_tag
<< 4;
2767 switch (bundle
->tag
) {
2772 /* Actually emit each component */
2773 util_dynarray_append(emission
, uint32_t, bundle
->control
| lookahead
);
2775 for (int i
= 0; i
< bundle
->register_words_count
; ++i
)
2776 util_dynarray_append(emission
, uint16_t, bundle
->register_words
[i
]);
2778 /* Emit body words based on the instructions bundled */
2779 for (int i
= 0; i
< bundle
->instruction_count
; ++i
) {
2780 midgard_instruction
*ins
= &bundle
->instructions
[i
];
2782 if (ins
->unit
& UNITS_ANY_VECTOR
) {
2783 memcpy(util_dynarray_grow(emission
, sizeof(midgard_vector_alu
)), &ins
->alu
, sizeof(midgard_vector_alu
));
2784 } else if (ins
->compact_branch
) {
2785 /* Dummy move, XXX DRY */
2786 if ((i
== 0) && ins
->writeout
) {
2787 midgard_instruction ins
= v_fmov(0, blank_alu_src
, SSA_FIXED_REGISTER(0));
2788 memcpy(util_dynarray_grow(emission
, sizeof(midgard_vector_alu
)), &ins
.alu
, sizeof(midgard_vector_alu
));
2791 if (ins
->unit
== ALU_ENAB_BR_COMPACT
) {
2792 memcpy(util_dynarray_grow(emission
, sizeof(ins
->br_compact
)), &ins
->br_compact
, sizeof(ins
->br_compact
));
2794 memcpy(util_dynarray_grow(emission
, sizeof(ins
->branch_extended
)), &ins
->branch_extended
, sizeof(ins
->branch_extended
));
2798 midgard_scalar_alu scalarised
= vector_to_scalar_alu(ins
->alu
, ins
);
2799 memcpy(util_dynarray_grow(emission
, sizeof(scalarised
)), &scalarised
, sizeof(scalarised
));
2803 /* Emit padding (all zero) */
2804 memset(util_dynarray_grow(emission
, bundle
->padding
), 0, bundle
->padding
);
2806 /* Tack on constants */
2808 if (bundle
->has_embedded_constants
) {
2809 util_dynarray_append(emission
, float, bundle
->constants
[0]);
2810 util_dynarray_append(emission
, float, bundle
->constants
[1]);
2811 util_dynarray_append(emission
, float, bundle
->constants
[2]);
2812 util_dynarray_append(emission
, float, bundle
->constants
[3]);
2818 case TAG_LOAD_STORE_4
: {
2819 /* One or two composing instructions */
2821 uint64_t current64
, next64
= LDST_NOP
;
2823 memcpy(¤t64
, &bundle
->instructions
[0].load_store
, sizeof(current64
));
2825 if (bundle
->instruction_count
== 2)
2826 memcpy(&next64
, &bundle
->instructions
[1].load_store
, sizeof(next64
));
2828 midgard_load_store instruction
= {
2829 .type
= bundle
->tag
,
2830 .next_type
= next_tag
,
2835 util_dynarray_append(emission
, midgard_load_store
, instruction
);
2840 case TAG_TEXTURE_4
: {
2841 /* Texture instructions are easy, since there is no
2842 * pipelining nor VLIW to worry about. We may need to set the .last flag */
2844 midgard_instruction
*ins
= &bundle
->instructions
[0];
2846 ins
->texture
.type
= TAG_TEXTURE_4
;
2847 ins
->texture
.next_type
= next_tag
;
2849 ctx
->texture_op_count
--;
2851 if (!ctx
->texture_op_count
) {
2852 ins
->texture
.cont
= 0;
2853 ins
->texture
.last
= 1;
2856 util_dynarray_append(emission
, midgard_texture_word
, ins
->texture
);
2861 DBG("Unknown midgard instruction type\n");
2868 /* ALU instructions can inline or embed constants, which decreases register
2869 * pressure and saves space. */
2871 #define CONDITIONAL_ATTACH(src) { \
2872 void *entry = _mesa_hash_table_u64_search(ctx->ssa_constants, alu->ssa_args.src + 1); \
2875 attach_constants(ctx, alu, entry, alu->ssa_args.src + 1); \
2876 alu->ssa_args.src = SSA_FIXED_REGISTER(REGISTER_CONSTANT); \
2881 inline_alu_constants(compiler_context
*ctx
)
2883 mir_foreach_instr(ctx
, alu
) {
2884 /* Other instructions cannot inline constants */
2885 if (alu
->type
!= TAG_ALU_4
) continue;
2887 /* If there is already a constant here, we can do nothing */
2888 if (alu
->has_constants
) continue;
2890 /* It makes no sense to inline constants on a branch */
2891 if (alu
->compact_branch
|| alu
->prepacked_branch
) continue;
2893 CONDITIONAL_ATTACH(src0
);
2895 if (!alu
->has_constants
) {
2896 CONDITIONAL_ATTACH(src1
)
2897 } else if (!alu
->inline_constant
) {
2898 /* Corner case: _two_ vec4 constants, for instance with a
2899 * csel. For this case, we can only use a constant
2900 * register for one, we'll have to emit a move for the
2901 * other. Note, if both arguments are constants, then
2902 * necessarily neither argument depends on the value of
2903 * any particular register. As the destination register
2904 * will be wiped, that means we can spill the constant
2905 * to the destination register.
2908 void *entry
= _mesa_hash_table_u64_search(ctx
->ssa_constants
, alu
->ssa_args
.src1
+ 1);
2909 unsigned scratch
= alu
->ssa_args
.dest
;
2912 midgard_instruction ins
= v_fmov(SSA_FIXED_REGISTER(REGISTER_CONSTANT
), blank_alu_src
, scratch
);
2913 attach_constants(ctx
, &ins
, entry
, alu
->ssa_args
.src1
+ 1);
2915 /* Force a break XXX Defer r31 writes */
2916 ins
.unit
= UNIT_VLUT
;
2918 /* Set the source */
2919 alu
->ssa_args
.src1
= scratch
;
2921 /* Inject us -before- the last instruction which set r31 */
2922 mir_insert_instruction_before(mir_prev_op(alu
), ins
);
2928 /* Midgard supports two types of constants, embedded constants (128-bit) and
2929 * inline constants (16-bit). Sometimes, especially with scalar ops, embedded
2930 * constants can be demoted to inline constants, for space savings and
2931 * sometimes a performance boost */
2934 embedded_to_inline_constant(compiler_context
*ctx
)
2936 mir_foreach_instr(ctx
, ins
) {
2937 if (!ins
->has_constants
) continue;
2939 if (ins
->ssa_args
.inline_constant
) continue;
2941 /* Blend constants must not be inlined by definition */
2942 if (ins
->has_blend_constant
) continue;
2944 /* src1 cannot be an inline constant due to encoding
2945 * restrictions. So, if possible we try to flip the arguments
2948 int op
= ins
->alu
.op
;
2950 if (ins
->ssa_args
.src0
== SSA_FIXED_REGISTER(REGISTER_CONSTANT
)) {
2951 /* Flip based on op. Fallthrough intentional */
2954 /* These ops require an operational change to flip their arguments TODO */
2955 case midgard_alu_op_flt
:
2956 case midgard_alu_op_fle
:
2957 case midgard_alu_op_ilt
:
2958 case midgard_alu_op_ile
:
2959 case midgard_alu_op_fcsel
:
2960 case midgard_alu_op_icsel
:
2961 case midgard_alu_op_isub
:
2962 DBG("Missed non-commutative flip (%s)\n", alu_opcode_props
[op
].name
);
2965 /* These ops are commutative and Just Flip */
2966 case midgard_alu_op_fne
:
2967 case midgard_alu_op_fadd
:
2968 case midgard_alu_op_fmul
:
2969 case midgard_alu_op_fmin
:
2970 case midgard_alu_op_fmax
:
2971 case midgard_alu_op_iadd
:
2972 case midgard_alu_op_imul
:
2973 case midgard_alu_op_feq
:
2974 case midgard_alu_op_ieq
:
2975 case midgard_alu_op_ine
:
2976 case midgard_alu_op_iand
:
2977 case midgard_alu_op_ior
:
2978 case midgard_alu_op_ixor
:
2979 /* Flip the SSA numbers */
2980 ins
->ssa_args
.src0
= ins
->ssa_args
.src1
;
2981 ins
->ssa_args
.src1
= SSA_FIXED_REGISTER(REGISTER_CONSTANT
);
2983 /* And flip the modifiers */
2987 src_temp
= ins
->alu
.src2
;
2988 ins
->alu
.src2
= ins
->alu
.src1
;
2989 ins
->alu
.src1
= src_temp
;
2996 if (ins
->ssa_args
.src1
== SSA_FIXED_REGISTER(REGISTER_CONSTANT
)) {
2997 /* Extract the source information */
2999 midgard_vector_alu_src
*src
;
3000 int q
= ins
->alu
.src2
;
3001 midgard_vector_alu_src
*m
= (midgard_vector_alu_src
*) &q
;
3004 /* Component is from the swizzle, e.g. r26.w -> w component. TODO: What if x is masked out? */
3005 int component
= src
->swizzle
& 3;
3007 /* Scale constant appropriately, if we can legally */
3008 uint16_t scaled_constant
= 0;
3010 /* XXX: Check legality */
3011 if (midgard_is_integer_op(op
)) {
3012 /* TODO: Inline integer */
3015 unsigned int *iconstants
= (unsigned int *) ins
->constants
;
3016 scaled_constant
= (uint16_t) iconstants
[component
];
3018 /* Constant overflow after resize */
3019 if (scaled_constant
!= iconstants
[component
])
3022 scaled_constant
= _mesa_float_to_half((float) ins
->constants
[component
]);
3025 /* We don't know how to handle these with a constant */
3027 if (src
->mod
|| src
->half
|| src
->rep_low
|| src
->rep_high
) {
3028 DBG("Bailing inline constant...\n");
3032 /* Make sure that the constant is not itself a
3033 * vector by checking if all accessed values
3034 * (by the swizzle) are the same. */
3036 uint32_t *cons
= (uint32_t *) ins
->constants
;
3037 uint32_t value
= cons
[component
];
3039 bool is_vector
= false;
3040 unsigned mask
= effective_writemask(&ins
->alu
);
3042 for (int c
= 1; c
< 4; ++c
) {
3043 /* We only care if this component is actually used */
3044 if (!(mask
& (1 << c
)))
3047 uint32_t test
= cons
[(src
->swizzle
>> (2 * c
)) & 3];
3049 if (test
!= value
) {
3058 /* Get rid of the embedded constant */
3059 ins
->has_constants
= false;
3060 ins
->ssa_args
.src1
= SSA_UNUSED_0
;
3061 ins
->ssa_args
.inline_constant
= true;
3062 ins
->inline_constant
= scaled_constant
;
3067 /* Map normal SSA sources to other SSA sources / fixed registers (like
3071 map_ssa_to_alias(compiler_context
*ctx
, int *ref
)
3073 unsigned int alias
= (uintptr_t) _mesa_hash_table_u64_search(ctx
->ssa_to_alias
, *ref
+ 1);
3076 /* Remove entry in leftovers to avoid a redunant fmov */
3078 struct set_entry
*leftover
= _mesa_set_search(ctx
->leftover_ssa_to_alias
, ((void *) (uintptr_t) (*ref
+ 1)));
3081 _mesa_set_remove(ctx
->leftover_ssa_to_alias
, leftover
);
3083 /* Assign the alias map */
3089 /* Basic dead code elimination on the MIR itself, which cleans up e.g. the
3090 * texture pipeline */
3093 midgard_opt_dead_code_eliminate(compiler_context
*ctx
, midgard_block
*block
)
3095 bool progress
= false;
3097 mir_foreach_instr_in_block_safe(block
, ins
) {
3098 if (ins
->type
!= TAG_ALU_4
) continue;
3099 if (ins
->compact_branch
) continue;
3101 if (ins
->ssa_args
.dest
>= SSA_FIXED_MINIMUM
) continue;
3102 if (midgard_is_pinned(ctx
, ins
->ssa_args
.dest
)) continue;
3103 if (is_live_after(ctx
, block
, ins
, ins
->ssa_args
.dest
)) continue;
3105 mir_remove_instruction(ins
);
3113 midgard_opt_copy_prop(compiler_context
*ctx
, midgard_block
*block
)
3115 bool progress
= false;
3117 mir_foreach_instr_in_block_safe(block
, ins
) {
3118 if (ins
->type
!= TAG_ALU_4
) continue;
3119 if (!OP_IS_MOVE(ins
->alu
.op
)) continue;
3121 unsigned from
= ins
->ssa_args
.src1
;
3122 unsigned to
= ins
->ssa_args
.dest
;
3124 /* We only work on pure SSA */
3126 if (to
>= SSA_FIXED_MINIMUM
) continue;
3127 if (from
>= SSA_FIXED_MINIMUM
) continue;
3129 /* Also, if the move has side effects, we're helpless */
3131 midgard_vector_alu_src src
=
3132 vector_alu_from_unsigned(ins
->alu
.src2
);
3133 unsigned mask
= squeeze_writemask(ins
->alu
.mask
);
3134 bool is_int
= midgard_is_integer_op(ins
->alu
.op
);
3136 if (mir_nontrivial_mod(src
, is_int
, mask
)) continue;
3138 mir_foreach_instr_in_block_from(block
, v
, mir_next_op(ins
)) {
3139 if (v
->ssa_args
.src0
== to
) {
3140 v
->ssa_args
.src0
= from
;
3144 if (v
->ssa_args
.src1
== to
&& !v
->ssa_args
.inline_constant
) {
3145 v
->ssa_args
.src1
= from
;
3154 /* The following passes reorder MIR instructions to enable better scheduling */
3157 midgard_pair_load_store(compiler_context
*ctx
, midgard_block
*block
)
3159 mir_foreach_instr_in_block_safe(block
, ins
) {
3160 if (ins
->type
!= TAG_LOAD_STORE_4
) continue;
3162 /* We've found a load/store op. Check if next is also load/store. */
3163 midgard_instruction
*next_op
= mir_next_op(ins
);
3164 if (&next_op
->link
!= &block
->instructions
) {
3165 if (next_op
->type
== TAG_LOAD_STORE_4
) {
3166 /* If so, we're done since we're a pair */
3167 ins
= mir_next_op(ins
);
3171 /* Maximum search distance to pair, to avoid register pressure disasters */
3172 int search_distance
= 8;
3174 /* Otherwise, we have an orphaned load/store -- search for another load */
3175 mir_foreach_instr_in_block_from(block
, c
, mir_next_op(ins
)) {
3176 /* Terminate search if necessary */
3177 if (!(search_distance
--)) break;
3179 if (c
->type
!= TAG_LOAD_STORE_4
) continue;
3181 /* Stores cannot be reordered, since they have
3182 * dependencies. For the same reason, indirect
3183 * loads cannot be reordered as their index is
3184 * loaded in r27.w */
3186 if (OP_IS_STORE(c
->load_store
.op
)) continue;
3188 /* It appears the 0x800 bit is set whenever a
3189 * load is direct, unset when it is indirect.
3190 * Skip indirect loads. */
3192 if (!(c
->load_store
.unknown
& 0x800)) continue;
3194 /* We found one! Move it up to pair and remove it from the old location */
3196 mir_insert_instruction_before(ins
, *c
);
3197 mir_remove_instruction(c
);
3205 /* Emit varying stores late */
3208 midgard_emit_store(compiler_context
*ctx
, midgard_block
*block
) {
3209 /* Iterate in reverse to get the final write, rather than the first */
3211 mir_foreach_instr_in_block_safe_rev(block
, ins
) {
3212 /* Check if what we just wrote needs a store */
3213 int idx
= ins
->ssa_args
.dest
;
3214 uintptr_t varying
= ((uintptr_t) _mesa_hash_table_u64_search(ctx
->ssa_varyings
, idx
+ 1));
3216 if (!varying
) continue;
3220 /* We need to store to the appropriate varying, so emit the
3223 /* TODO: Integrate with special purpose RA (and scheduler?) */
3224 bool high_varying_register
= false;
3226 midgard_instruction mov
= v_fmov(idx
, blank_alu_src
, SSA_FIXED_REGISTER(REGISTER_VARYING_BASE
+ high_varying_register
));
3228 midgard_instruction st
= m_store_vary_32(SSA_FIXED_REGISTER(high_varying_register
), varying
);
3229 st
.load_store
.unknown
= 0x1E9E; /* XXX: What is this? */
3231 mir_insert_instruction_before(mir_next_op(ins
), st
);
3232 mir_insert_instruction_before(mir_next_op(ins
), mov
);
3234 /* We no longer need to store this varying */
3235 _mesa_hash_table_u64_remove(ctx
->ssa_varyings
, idx
+ 1);
3239 /* If there are leftovers after the below pass, emit actual fmov
3240 * instructions for the slow-but-correct path */
3243 emit_leftover_move(compiler_context
*ctx
)
3245 set_foreach(ctx
->leftover_ssa_to_alias
, leftover
) {
3246 int base
= ((uintptr_t) leftover
->key
) - 1;
3249 map_ssa_to_alias(ctx
, &mapped
);
3250 EMIT(fmov
, mapped
, blank_alu_src
, base
);
3255 actualise_ssa_to_alias(compiler_context
*ctx
)
3257 mir_foreach_instr(ctx
, ins
) {
3258 map_ssa_to_alias(ctx
, &ins
->ssa_args
.src0
);
3259 map_ssa_to_alias(ctx
, &ins
->ssa_args
.src1
);
3262 emit_leftover_move(ctx
);
3266 emit_fragment_epilogue(compiler_context
*ctx
)
3268 /* Special case: writing out constants requires us to include the move
3269 * explicitly now, so shove it into r0 */
3271 void *constant_value
= _mesa_hash_table_u64_search(ctx
->ssa_constants
, ctx
->fragment_output
+ 1);
3273 if (constant_value
) {
3274 midgard_instruction ins
= v_fmov(SSA_FIXED_REGISTER(REGISTER_CONSTANT
), blank_alu_src
, SSA_FIXED_REGISTER(0));
3275 attach_constants(ctx
, &ins
, constant_value
, ctx
->fragment_output
+ 1);
3276 emit_mir_instruction(ctx
, ins
);
3279 /* Perform the actual fragment writeout. We have two writeout/branch
3280 * instructions, forming a loop until writeout is successful as per the
3281 * docs. TODO: gl_FragDepth */
3283 EMIT(alu_br_compact_cond
, midgard_jmp_writeout_op_writeout
, TAG_ALU_4
, 0, midgard_condition_always
);
3284 EMIT(alu_br_compact_cond
, midgard_jmp_writeout_op_writeout
, TAG_ALU_4
, -1, midgard_condition_always
);
3287 /* For the blend epilogue, we need to convert the blended fragment vec4 (stored
3288 * in r0) to a RGBA8888 value by scaling and type converting. We then output it
3289 * with the int8 analogue to the fragment epilogue */
3292 emit_blend_epilogue(compiler_context
*ctx
)
3294 /* vmul.fmul.none.fulllow hr48, r0, #255 */
3296 midgard_instruction scale
= {
3299 .inline_constant
= _mesa_float_to_half(255.0),
3301 .src0
= SSA_FIXED_REGISTER(0),
3302 .src1
= SSA_UNUSED_0
,
3303 .dest
= SSA_FIXED_REGISTER(24),
3304 .inline_constant
= true
3307 .op
= midgard_alu_op_fmul
,
3308 .reg_mode
= midgard_reg_mode_full
,
3309 .dest_override
= midgard_dest_override_lower
,
3311 .src1
= vector_alu_srco_unsigned(blank_alu_src
),
3312 .src2
= vector_alu_srco_unsigned(blank_alu_src
),
3316 emit_mir_instruction(ctx
, scale
);
3318 /* vadd.f2u8.pos.low hr0, hr48, #0 */
3320 midgard_vector_alu_src alu_src
= blank_alu_src
;
3321 alu_src
.half
= true;
3323 midgard_instruction f2u8
= {
3326 .src0
= SSA_FIXED_REGISTER(24),
3327 .src1
= SSA_UNUSED_0
,
3328 .dest
= SSA_FIXED_REGISTER(0),
3329 .inline_constant
= true
3332 .op
= midgard_alu_op_f2u8
,
3333 .reg_mode
= midgard_reg_mode_half
,
3334 .dest_override
= midgard_dest_override_lower
,
3335 .outmod
= midgard_outmod_pos
,
3337 .src1
= vector_alu_srco_unsigned(alu_src
),
3338 .src2
= vector_alu_srco_unsigned(blank_alu_src
),
3342 emit_mir_instruction(ctx
, f2u8
);
3344 /* vmul.imov.quarter r0, r0, r0 */
3346 midgard_instruction imov_8
= {
3349 .src0
= SSA_UNUSED_1
,
3350 .src1
= SSA_FIXED_REGISTER(0),
3351 .dest
= SSA_FIXED_REGISTER(0),
3354 .op
= midgard_alu_op_imov
,
3355 .reg_mode
= midgard_reg_mode_quarter
,
3356 .dest_override
= midgard_dest_override_none
,
3358 .src1
= vector_alu_srco_unsigned(blank_alu_src
),
3359 .src2
= vector_alu_srco_unsigned(blank_alu_src
),
3363 /* Emit branch epilogue with the 8-bit move as the source */
3365 emit_mir_instruction(ctx
, imov_8
);
3366 EMIT(alu_br_compact_cond
, midgard_jmp_writeout_op_writeout
, TAG_ALU_4
, 0, midgard_condition_always
);
3368 emit_mir_instruction(ctx
, imov_8
);
3369 EMIT(alu_br_compact_cond
, midgard_jmp_writeout_op_writeout
, TAG_ALU_4
, -1, midgard_condition_always
);
3372 static midgard_block
*
3373 emit_block(compiler_context
*ctx
, nir_block
*block
)
3375 midgard_block
*this_block
= calloc(sizeof(midgard_block
), 1);
3376 list_addtail(&this_block
->link
, &ctx
->blocks
);
3378 this_block
->is_scheduled
= false;
3381 ctx
->texture_index
[0] = -1;
3382 ctx
->texture_index
[1] = -1;
3384 /* Add us as a successor to the block we are following */
3385 if (ctx
->current_block
)
3386 midgard_block_add_successor(ctx
->current_block
, this_block
);
3388 /* Set up current block */
3389 list_inithead(&this_block
->instructions
);
3390 ctx
->current_block
= this_block
;
3392 nir_foreach_instr(instr
, block
) {
3393 emit_instr(ctx
, instr
);
3394 ++ctx
->instruction_count
;
3397 inline_alu_constants(ctx
);
3398 embedded_to_inline_constant(ctx
);
3400 /* Perform heavylifting for aliasing */
3401 actualise_ssa_to_alias(ctx
);
3403 midgard_emit_store(ctx
, this_block
);
3404 midgard_pair_load_store(ctx
, this_block
);
3406 /* Append fragment shader epilogue (value writeout) */
3407 if (ctx
->stage
== MESA_SHADER_FRAGMENT
) {
3408 if (block
== nir_impl_last_block(ctx
->func
->impl
)) {
3410 emit_blend_epilogue(ctx
);
3412 emit_fragment_epilogue(ctx
);
3416 if (block
== nir_start_block(ctx
->func
->impl
))
3417 ctx
->initial_block
= this_block
;
3419 if (block
== nir_impl_last_block(ctx
->func
->impl
))
3420 ctx
->final_block
= this_block
;
3422 /* Allow the next control flow to access us retroactively, for
3424 ctx
->current_block
= this_block
;
3426 /* Document the fallthrough chain */
3427 ctx
->previous_source_block
= this_block
;
3432 static midgard_block
*emit_cf_list(struct compiler_context
*ctx
, struct exec_list
*list
);
3435 emit_if(struct compiler_context
*ctx
, nir_if
*nif
)
3437 /* Conditional branches expect the condition in r31.w; emit a move for
3438 * that in the _previous_ block (which is the current block). */
3439 emit_condition(ctx
, &nif
->condition
, true, COMPONENT_X
);
3441 /* Speculatively emit the branch, but we can't fill it in until later */
3442 EMIT(branch
, true, true);
3443 midgard_instruction
*then_branch
= mir_last_in_block(ctx
->current_block
);
3445 /* Emit the two subblocks */
3446 midgard_block
*then_block
= emit_cf_list(ctx
, &nif
->then_list
);
3448 /* Emit a jump from the end of the then block to the end of the else */
3449 EMIT(branch
, false, false);
3450 midgard_instruction
*then_exit
= mir_last_in_block(ctx
->current_block
);
3452 /* Emit second block, and check if it's empty */
3454 int else_idx
= ctx
->block_count
;
3455 int count_in
= ctx
->instruction_count
;
3456 midgard_block
*else_block
= emit_cf_list(ctx
, &nif
->else_list
);
3457 int after_else_idx
= ctx
->block_count
;
3459 /* Now that we have the subblocks emitted, fix up the branches */
3464 if (ctx
->instruction_count
== count_in
) {
3465 /* The else block is empty, so don't emit an exit jump */
3466 mir_remove_instruction(then_exit
);
3467 then_branch
->branch
.target_block
= after_else_idx
;
3469 then_branch
->branch
.target_block
= else_idx
;
3470 then_exit
->branch
.target_block
= after_else_idx
;
3475 emit_loop(struct compiler_context
*ctx
, nir_loop
*nloop
)
3477 /* Remember where we are */
3478 midgard_block
*start_block
= ctx
->current_block
;
3480 /* Allocate a loop number, growing the current inner loop depth */
3481 int loop_idx
= ++ctx
->current_loop_depth
;
3483 /* Get index from before the body so we can loop back later */
3484 int start_idx
= ctx
->block_count
;
3486 /* Emit the body itself */
3487 emit_cf_list(ctx
, &nloop
->body
);
3489 /* Branch back to loop back */
3490 struct midgard_instruction br_back
= v_branch(false, false);
3491 br_back
.branch
.target_block
= start_idx
;
3492 emit_mir_instruction(ctx
, br_back
);
3494 /* Mark down that branch in the graph. Note that we're really branching
3495 * to the block *after* we started in. TODO: Why doesn't the branch
3496 * itself have an off-by-one then...? */
3497 midgard_block_add_successor(ctx
->current_block
, start_block
->successors
[0]);
3499 /* Find the index of the block about to follow us (note: we don't add
3500 * one; blocks are 0-indexed so we get a fencepost problem) */
3501 int break_block_idx
= ctx
->block_count
;
3503 /* Fix up the break statements we emitted to point to the right place,
3504 * now that we can allocate a block number for them */
3506 list_for_each_entry_from(struct midgard_block
, block
, start_block
, &ctx
->blocks
, link
) {
3507 mir_foreach_instr_in_block(block
, ins
) {
3508 if (ins
->type
!= TAG_ALU_4
) continue;
3509 if (!ins
->compact_branch
) continue;
3510 if (ins
->prepacked_branch
) continue;
3512 /* We found a branch -- check the type to see if we need to do anything */
3513 if (ins
->branch
.target_type
!= TARGET_BREAK
) continue;
3515 /* It's a break! Check if it's our break */
3516 if (ins
->branch
.target_break
!= loop_idx
) continue;
3518 /* Okay, cool, we're breaking out of this loop.
3519 * Rewrite from a break to a goto */
3521 ins
->branch
.target_type
= TARGET_GOTO
;
3522 ins
->branch
.target_block
= break_block_idx
;
3526 /* Now that we've finished emitting the loop, free up the depth again
3527 * so we play nice with recursion amid nested loops */
3528 --ctx
->current_loop_depth
;
3531 static midgard_block
*
3532 emit_cf_list(struct compiler_context
*ctx
, struct exec_list
*list
)
3534 midgard_block
*start_block
= NULL
;
3536 foreach_list_typed(nir_cf_node
, node
, node
, list
) {
3537 switch (node
->type
) {
3538 case nir_cf_node_block
: {
3539 midgard_block
*block
= emit_block(ctx
, nir_cf_node_as_block(node
));
3542 start_block
= block
;
3547 case nir_cf_node_if
:
3548 emit_if(ctx
, nir_cf_node_as_if(node
));
3551 case nir_cf_node_loop
:
3552 emit_loop(ctx
, nir_cf_node_as_loop(node
));
3555 case nir_cf_node_function
:
3564 /* Due to lookahead, we need to report the first tag executed in the command
3565 * stream and in branch targets. An initial block might be empty, so iterate
3566 * until we find one that 'works' */
3569 midgard_get_first_tag_from_block(compiler_context
*ctx
, unsigned block_idx
)
3571 midgard_block
*initial_block
= mir_get_block(ctx
, block_idx
);
3573 unsigned first_tag
= 0;
3576 midgard_bundle
*initial_bundle
= util_dynarray_element(&initial_block
->bundles
, midgard_bundle
, 0);
3578 if (initial_bundle
) {
3579 first_tag
= initial_bundle
->tag
;
3583 /* Initial block is empty, try the next block */
3584 initial_block
= list_first_entry(&(initial_block
->link
), midgard_block
, link
);
3585 } while(initial_block
!= NULL
);
3592 midgard_compile_shader_nir(nir_shader
*nir
, midgard_program
*program
, bool is_blend
)
3594 struct util_dynarray
*compiled
= &program
->compiled
;
3596 midgard_debug
= debug_get_option_midgard_debug();
3598 compiler_context ictx
= {
3600 .stage
= nir
->info
.stage
,
3602 .is_blend
= is_blend
,
3603 .blend_constant_offset
= -1,
3605 .alpha_ref
= program
->alpha_ref
3608 compiler_context
*ctx
= &ictx
;
3610 /* TODO: Decide this at runtime */
3611 ctx
->uniform_cutoff
= 8;
3613 /* Assign var locations early, so the epilogue can use them if necessary */
3615 nir_assign_var_locations(&nir
->outputs
, &nir
->num_outputs
, glsl_type_size
);
3616 nir_assign_var_locations(&nir
->inputs
, &nir
->num_inputs
, glsl_type_size
);
3617 nir_assign_var_locations(&nir
->uniforms
, &nir
->num_uniforms
, glsl_type_size
);
3619 /* Initialize at a global (not block) level hash tables */
3621 ctx
->ssa_constants
= _mesa_hash_table_u64_create(NULL
);
3622 ctx
->ssa_varyings
= _mesa_hash_table_u64_create(NULL
);
3623 ctx
->ssa_to_alias
= _mesa_hash_table_u64_create(NULL
);
3624 ctx
->ssa_to_register
= _mesa_hash_table_u64_create(NULL
);
3625 ctx
->hash_to_temp
= _mesa_hash_table_u64_create(NULL
);
3626 ctx
->sysval_to_id
= _mesa_hash_table_u64_create(NULL
);
3627 ctx
->leftover_ssa_to_alias
= _mesa_set_create(NULL
, _mesa_hash_pointer
, _mesa_key_pointer_equal
);
3629 /* Record the varying mapping for the command stream's bookkeeping */
3631 struct exec_list
*varyings
=
3632 ctx
->stage
== MESA_SHADER_VERTEX
? &nir
->outputs
: &nir
->inputs
;
3634 nir_foreach_variable(var
, varyings
) {
3635 unsigned loc
= var
->data
.driver_location
;
3636 unsigned sz
= glsl_type_size(var
->type
, FALSE
);
3638 for (int c
= 0; c
< sz
; ++c
) {
3639 program
->varyings
[loc
+ c
] = var
->data
.location
;
3643 /* Lower gl_Position pre-optimisation */
3645 if (ctx
->stage
== MESA_SHADER_VERTEX
)
3646 NIR_PASS_V(nir
, nir_lower_viewport_transform
);
3648 NIR_PASS_V(nir
, nir_lower_var_copies
);
3649 NIR_PASS_V(nir
, nir_lower_vars_to_ssa
);
3650 NIR_PASS_V(nir
, nir_split_var_copies
);
3651 NIR_PASS_V(nir
, nir_lower_var_copies
);
3652 NIR_PASS_V(nir
, nir_lower_global_vars_to_local
);
3653 NIR_PASS_V(nir
, nir_lower_var_copies
);
3654 NIR_PASS_V(nir
, nir_lower_vars_to_ssa
);
3656 NIR_PASS_V(nir
, nir_lower_io
, nir_var_all
, glsl_type_size
, 0);
3658 /* Optimisation passes */
3662 if (midgard_debug
& MIDGARD_DBG_SHADERS
) {
3663 nir_print_shader(nir
, stdout
);
3666 /* Assign sysvals and counts, now that we're sure
3667 * (post-optimisation) */
3669 midgard_nir_assign_sysvals(ctx
, nir
);
3671 program
->uniform_count
= nir
->num_uniforms
;
3672 program
->sysval_count
= ctx
->sysval_count
;
3673 memcpy(program
->sysvals
, ctx
->sysvals
, sizeof(ctx
->sysvals
[0]) * ctx
->sysval_count
);
3675 program
->attribute_count
= (ctx
->stage
== MESA_SHADER_VERTEX
) ? nir
->num_inputs
: 0;
3676 program
->varying_count
= (ctx
->stage
== MESA_SHADER_VERTEX
) ? nir
->num_outputs
: ((ctx
->stage
== MESA_SHADER_FRAGMENT
) ? nir
->num_inputs
: 0);
3678 nir_foreach_function(func
, nir
) {
3682 list_inithead(&ctx
->blocks
);
3683 ctx
->block_count
= 0;
3686 emit_cf_list(ctx
, &func
->impl
->body
);
3687 emit_block(ctx
, func
->impl
->end_block
);
3689 break; /* TODO: Multi-function shaders */
3692 util_dynarray_init(compiled
, NULL
);
3694 /* MIR-level optimizations */
3696 bool progress
= false;
3701 mir_foreach_block(ctx
, block
) {
3702 progress
|= midgard_opt_copy_prop(ctx
, block
);
3703 progress
|= midgard_opt_dead_code_eliminate(ctx
, block
);
3708 schedule_program(ctx
);
3710 /* Now that all the bundles are scheduled and we can calculate block
3711 * sizes, emit actual branch instructions rather than placeholders */
3713 int br_block_idx
= 0;
3715 mir_foreach_block(ctx
, block
) {
3716 util_dynarray_foreach(&block
->bundles
, midgard_bundle
, bundle
) {
3717 for (int c
= 0; c
< bundle
->instruction_count
; ++c
) {
3718 midgard_instruction
*ins
= &bundle
->instructions
[c
];
3720 if (!midgard_is_branch_unit(ins
->unit
)) continue;
3722 if (ins
->prepacked_branch
) continue;
3724 /* Parse some basic branch info */
3725 bool is_compact
= ins
->unit
== ALU_ENAB_BR_COMPACT
;
3726 bool is_conditional
= ins
->branch
.conditional
;
3727 bool is_inverted
= ins
->branch
.invert_conditional
;
3728 bool is_discard
= ins
->branch
.target_type
== TARGET_DISCARD
;
3730 /* Determine the block we're jumping to */
3731 int target_number
= ins
->branch
.target_block
;
3733 /* Report the destination tag. Discards don't need this */
3734 int dest_tag
= is_discard
? 0 : midgard_get_first_tag_from_block(ctx
, target_number
);
3736 /* Count up the number of quadwords we're jumping over. That is, the number of quadwords in each of the blocks between (br_block_idx, target_number) */
3737 int quadword_offset
= 0;
3740 /* Jump to the end of the shader. We
3741 * need to include not only the
3742 * following blocks, but also the
3743 * contents of our current block (since
3744 * discard can come in the middle of
3747 midgard_block
*blk
= mir_get_block(ctx
, br_block_idx
+ 1);
3749 for (midgard_bundle
*bun
= bundle
+ 1; bun
< (midgard_bundle
*)((char*) block
->bundles
.data
+ block
->bundles
.size
); ++bun
) {
3750 quadword_offset
+= quadword_size(bun
->tag
);
3753 mir_foreach_block_from(ctx
, blk
, b
) {
3754 quadword_offset
+= b
->quadword_count
;
3757 } else if (target_number
> br_block_idx
) {
3760 for (int idx
= br_block_idx
+ 1; idx
< target_number
; ++idx
) {
3761 midgard_block
*blk
= mir_get_block(ctx
, idx
);
3764 quadword_offset
+= blk
->quadword_count
;
3767 /* Jump backwards */
3769 for (int idx
= br_block_idx
; idx
>= target_number
; --idx
) {
3770 midgard_block
*blk
= mir_get_block(ctx
, idx
);
3773 quadword_offset
-= blk
->quadword_count
;
3777 /* Unconditional extended branches (far jumps)
3778 * have issues, so we always use a conditional
3779 * branch, setting the condition to always for
3780 * unconditional. For compact unconditional
3781 * branches, cond isn't used so it doesn't
3782 * matter what we pick. */
3784 midgard_condition cond
=
3785 !is_conditional
? midgard_condition_always
:
3786 is_inverted
? midgard_condition_false
:
3787 midgard_condition_true
;
3789 midgard_jmp_writeout_op op
=
3790 is_discard
? midgard_jmp_writeout_op_discard
:
3791 (is_compact
&& !is_conditional
) ? midgard_jmp_writeout_op_branch_uncond
:
3792 midgard_jmp_writeout_op_branch_cond
;
3795 midgard_branch_extended branch
=
3796 midgard_create_branch_extended(
3801 memcpy(&ins
->branch_extended
, &branch
, sizeof(branch
));
3802 } else if (is_conditional
|| is_discard
) {
3803 midgard_branch_cond branch
= {
3805 .dest_tag
= dest_tag
,
3806 .offset
= quadword_offset
,
3810 assert(branch
.offset
== quadword_offset
);
3812 memcpy(&ins
->br_compact
, &branch
, sizeof(branch
));
3814 assert(op
== midgard_jmp_writeout_op_branch_uncond
);
3816 midgard_branch_uncond branch
= {
3818 .dest_tag
= dest_tag
,
3819 .offset
= quadword_offset
,
3823 assert(branch
.offset
== quadword_offset
);
3825 memcpy(&ins
->br_compact
, &branch
, sizeof(branch
));
3833 /* Emit flat binary from the instruction arrays. Iterate each block in
3834 * sequence. Save instruction boundaries such that lookahead tags can
3835 * be assigned easily */
3837 /* Cache _all_ bundles in source order for lookahead across failed branches */
3839 int bundle_count
= 0;
3840 mir_foreach_block(ctx
, block
) {
3841 bundle_count
+= block
->bundles
.size
/ sizeof(midgard_bundle
);
3843 midgard_bundle
**source_order_bundles
= malloc(sizeof(midgard_bundle
*) * bundle_count
);
3845 mir_foreach_block(ctx
, block
) {
3846 util_dynarray_foreach(&block
->bundles
, midgard_bundle
, bundle
) {
3847 source_order_bundles
[bundle_idx
++] = bundle
;
3851 int current_bundle
= 0;
3853 mir_foreach_block(ctx
, block
) {
3854 util_dynarray_foreach(&block
->bundles
, midgard_bundle
, bundle
) {
3857 if (current_bundle
+ 1 < bundle_count
) {
3858 uint8_t next
= source_order_bundles
[current_bundle
+ 1]->tag
;
3860 if (!(current_bundle
+ 2 < bundle_count
) && IS_ALU(next
)) {
3867 emit_binary_bundle(ctx
, bundle
, compiled
, lookahead
);
3871 /* TODO: Free deeper */
3872 //util_dynarray_fini(&block->instructions);
3875 free(source_order_bundles
);
3877 /* Report the very first tag executed */
3878 program
->first_tag
= midgard_get_first_tag_from_block(ctx
, 0);
3880 /* Deal with off-by-one related to the fencepost problem */
3881 program
->work_register_count
= ctx
->work_registers
+ 1;
3883 program
->can_discard
= ctx
->can_discard
;
3884 program
->uniform_cutoff
= ctx
->uniform_cutoff
;
3886 program
->blend_patch_offset
= ctx
->blend_constant_offset
;
3888 if (midgard_debug
& MIDGARD_DBG_SHADERS
)
3889 disassemble_midgard(program
->compiled
.data
, program
->compiled
.size
);