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 */
110 /* Generic in-memory data type repesenting a single logical instruction, rather
111 * than a single instruction group. This is the preferred form for code gen.
112 * Multiple midgard_insturctions will later be combined during scheduling,
113 * though this is not represented in this structure. Its format bridges
114 * the low-level binary representation with the higher level semantic meaning.
116 * Notably, it allows registers to be specified as block local SSA, for code
117 * emitted before the register allocation pass.
120 typedef struct midgard_instruction
{
121 /* Must be first for casting */
122 struct list_head link
;
124 unsigned type
; /* ALU, load/store, texture */
126 /* If the register allocator has not run yet... */
129 /* Special fields for an ALU instruction */
130 midgard_reg_info registers
;
132 /* I.e. (1 << alu_bit) */
137 uint16_t inline_constant
;
138 bool has_blend_constant
;
142 bool prepacked_branch
;
145 midgard_load_store_word load_store
;
146 midgard_vector_alu alu
;
147 midgard_texture_word texture
;
148 midgard_branch_extended branch_extended
;
151 /* General branch, rather than packed br_compact. Higher level
152 * than the other components */
153 midgard_branch branch
;
155 } midgard_instruction
;
157 typedef struct midgard_block
{
158 /* Link to next block. Must be first for mir_get_block */
159 struct list_head link
;
161 /* List of midgard_instructions emitted for the current block */
162 struct list_head instructions
;
166 /* List of midgard_bundles emitted (after the scheduler has run) */
167 struct util_dynarray bundles
;
169 /* Number of quadwords _actually_ emitted, as determined after scheduling */
170 unsigned quadword_count
;
172 struct midgard_block
*next_fallthrough
;
175 /* Helpers to generate midgard_instruction's using macro magic, since every
176 * driver seems to do it that way */
178 #define EMIT(op, ...) emit_mir_instruction(ctx, v_##op(__VA_ARGS__));
180 #define M_LOAD_STORE(name, rname, uname) \
181 static midgard_instruction m_##name(unsigned ssa, unsigned address) { \
182 midgard_instruction i = { \
183 .type = TAG_LOAD_STORE_4, \
190 .op = midgard_op_##name, \
192 .swizzle = SWIZZLE(COMPONENT_X, COMPONENT_Y, COMPONENT_Z, COMPONENT_W), \
200 #define M_LOAD(name) M_LOAD_STORE(name, dest, src0)
201 #define M_STORE(name) M_LOAD_STORE(name, src0, dest)
203 const midgard_vector_alu_src blank_alu_src
= {
204 .swizzle
= SWIZZLE(COMPONENT_X
, COMPONENT_Y
, COMPONENT_Z
, COMPONENT_W
),
207 const midgard_vector_alu_src blank_alu_src_xxxx
= {
208 .swizzle
= SWIZZLE(COMPONENT_X
, COMPONENT_X
, COMPONENT_X
, COMPONENT_X
),
211 const midgard_scalar_alu_src blank_scalar_alu_src
= {
215 /* Used for encoding the unused source of 1-op instructions */
216 const midgard_vector_alu_src zero_alu_src
= { 0 };
218 /* Coerce structs to integer */
221 vector_alu_srco_unsigned(midgard_vector_alu_src src
)
224 memcpy(&u
, &src
, sizeof(src
));
228 /* Inputs a NIR ALU source, with modifiers attached if necessary, and outputs
229 * the corresponding Midgard source */
231 static midgard_vector_alu_src
232 vector_alu_modifiers(nir_alu_src
*src
)
234 if (!src
) return blank_alu_src
;
236 midgard_vector_alu_src alu_src
= {
238 .negate
= src
->negate
,
241 .half
= 0, /* TODO */
242 .swizzle
= SWIZZLE_FROM_ARRAY(src
->swizzle
)
248 /* 'Intrinsic' move for misc aliasing uses independent of actual NIR ALU code */
250 static midgard_instruction
251 v_fmov(unsigned src
, midgard_vector_alu_src mod
, unsigned dest
)
253 midgard_instruction ins
= {
256 .src0
= SSA_UNUSED_1
,
261 .op
= midgard_alu_op_fmov
,
262 .reg_mode
= midgard_reg_mode_full
,
263 .dest_override
= midgard_dest_override_none
,
265 .src1
= vector_alu_srco_unsigned(zero_alu_src
),
266 .src2
= vector_alu_srco_unsigned(mod
)
273 /* load/store instructions have both 32-bit and 16-bit variants, depending on
274 * whether we are using vectors composed of highp or mediump. At the moment, we
275 * don't support half-floats -- this requires changes in other parts of the
276 * compiler -- therefore the 16-bit versions are commented out. */
278 //M_LOAD(load_attr_16);
279 M_LOAD(load_attr_32
);
280 //M_LOAD(load_vary_16);
281 M_LOAD(load_vary_32
);
282 //M_LOAD(load_uniform_16);
283 M_LOAD(load_uniform_32
);
284 M_LOAD(load_color_buffer_8
);
285 //M_STORE(store_vary_16);
286 M_STORE(store_vary_32
);
288 static midgard_instruction
289 v_alu_br_compact_cond(midgard_jmp_writeout_op op
, unsigned tag
, signed offset
, unsigned cond
)
291 midgard_branch_cond branch
= {
299 memcpy(&compact
, &branch
, sizeof(branch
));
301 midgard_instruction ins
= {
303 .unit
= ALU_ENAB_BR_COMPACT
,
304 .prepacked_branch
= true,
305 .compact_branch
= true,
306 .br_compact
= compact
309 if (op
== midgard_jmp_writeout_op_writeout
)
315 static midgard_instruction
316 v_branch(bool conditional
, bool invert
)
318 midgard_instruction ins
= {
320 .unit
= ALU_ENAB_BRANCH
,
321 .compact_branch
= true,
323 .conditional
= conditional
,
324 .invert_conditional
= invert
331 static midgard_branch_extended
332 midgard_create_branch_extended( midgard_condition cond
,
333 midgard_jmp_writeout_op op
,
335 signed quadword_offset
)
337 /* For unclear reasons, the condition code is repeated 8 times */
338 uint16_t duplicated_cond
=
348 midgard_branch_extended branch
= {
350 .dest_tag
= dest_tag
,
351 .offset
= quadword_offset
,
352 .cond
= duplicated_cond
358 typedef struct midgard_bundle
{
359 /* Tag for the overall bundle */
362 /* Instructions contained by the bundle */
363 int instruction_count
;
364 midgard_instruction instructions
[5];
366 /* Bundle-wide ALU configuration */
369 bool has_embedded_constants
;
371 bool has_blend_constant
;
373 uint16_t register_words
[8];
374 int register_words_count
;
376 uint64_t body_words
[8];
378 int body_words_count
;
381 typedef struct compiler_context
{
383 gl_shader_stage stage
;
385 /* Is internally a blend shader? Depends on stage == FRAGMENT */
388 /* Tracking for blend constant patching */
389 int blend_constant_number
;
390 int blend_constant_offset
;
392 /* Current NIR function */
395 /* Unordered list of midgard_blocks */
397 struct list_head blocks
;
399 midgard_block
*initial_block
;
400 midgard_block
*previous_source_block
;
401 midgard_block
*final_block
;
403 /* List of midgard_instructions emitted for the current block */
404 midgard_block
*current_block
;
406 /* The index corresponding to the current loop, e.g. for breaks/contineus */
409 /* Constants which have been loaded, for later inlining */
410 struct hash_table_u64
*ssa_constants
;
412 /* SSA indices to be outputted to corresponding varying offset */
413 struct hash_table_u64
*ssa_varyings
;
415 /* SSA values / registers which have been aliased. Naively, these
416 * demand a fmov output; instead, we alias them in a later pass to
417 * avoid the wasted op.
419 * A note on encoding: to avoid dynamic memory management here, rather
420 * than ampping to a pointer, we map to the source index; the key
421 * itself is just the destination index. */
423 struct hash_table_u64
*ssa_to_alias
;
424 struct set
*leftover_ssa_to_alias
;
426 /* Actual SSA-to-register for RA */
427 struct hash_table_u64
*ssa_to_register
;
429 /* Mapping of hashes computed from NIR indices to the sequential temp indices ultimately used in MIR */
430 struct hash_table_u64
*hash_to_temp
;
434 /* Uniform IDs for mdg */
435 struct hash_table_u64
*uniform_nir_to_mdg
;
438 /* Just the count of the max register used. Higher count => higher
439 * register pressure */
442 /* Used for cont/last hinting. Increase when a tex op is added.
443 * Decrease when a tex op is removed. */
444 int texture_op_count
;
446 /* Mapping of texture register -> SSA index for unaliasing */
447 int texture_index
[2];
449 /* Count of special uniforms (viewport, etc) in vec4 units */
450 int special_uniforms
;
452 /* If any path hits a discard instruction */
455 /* The number of uniforms allowable for the fast path */
458 /* Count of instructions emitted from NIR overall, across all blocks */
459 int instruction_count
;
461 /* Alpha ref value passed in */
464 /* The index corresponding to the fragment output */
465 unsigned fragment_output
;
468 /* Append instruction to end of current block */
470 static midgard_instruction
*
471 mir_upload_ins(struct midgard_instruction ins
)
473 midgard_instruction
*heap
= malloc(sizeof(ins
));
474 memcpy(heap
, &ins
, sizeof(ins
));
479 emit_mir_instruction(struct compiler_context
*ctx
, struct midgard_instruction ins
)
481 list_addtail(&(mir_upload_ins(ins
))->link
, &ctx
->current_block
->instructions
);
485 mir_insert_instruction_before(struct midgard_instruction
*tag
, struct midgard_instruction ins
)
487 list_addtail(&(mir_upload_ins(ins
))->link
, &tag
->link
);
491 mir_remove_instruction(struct midgard_instruction
*ins
)
493 list_del(&ins
->link
);
496 static midgard_instruction
*
497 mir_prev_op(struct midgard_instruction
*ins
)
499 return list_last_entry(&(ins
->link
), midgard_instruction
, link
);
502 static midgard_instruction
*
503 mir_next_op(struct midgard_instruction
*ins
)
505 return list_first_entry(&(ins
->link
), midgard_instruction
, link
);
508 static midgard_block
*
509 mir_next_block(struct midgard_block
*blk
)
511 return list_first_entry(&(blk
->link
), midgard_block
, link
);
515 #define mir_foreach_block(ctx, v) list_for_each_entry(struct midgard_block, v, &ctx->blocks, link)
516 #define mir_foreach_block_from(ctx, from, v) list_for_each_entry_from(struct midgard_block, v, from, &ctx->blocks, link)
518 #define mir_foreach_instr(ctx, v) list_for_each_entry(struct midgard_instruction, v, &ctx->current_block->instructions, link)
519 #define mir_foreach_instr_safe(ctx, v) list_for_each_entry_safe(struct midgard_instruction, v, &ctx->current_block->instructions, link)
520 #define mir_foreach_instr_in_block(block, v) list_for_each_entry(struct midgard_instruction, v, &block->instructions, link)
521 #define mir_foreach_instr_in_block_safe(block, v) list_for_each_entry_safe(struct midgard_instruction, v, &block->instructions, link)
522 #define mir_foreach_instr_in_block_safe_rev(block, v) list_for_each_entry_safe_rev(struct midgard_instruction, v, &block->instructions, link)
523 #define mir_foreach_instr_in_block_from(block, v, from) list_for_each_entry_from(struct midgard_instruction, v, from, &block->instructions, link)
526 static midgard_instruction
*
527 mir_last_in_block(struct midgard_block
*block
)
529 return list_last_entry(&block
->instructions
, struct midgard_instruction
, link
);
532 static midgard_block
*
533 mir_get_block(compiler_context
*ctx
, int idx
)
535 struct list_head
*lst
= &ctx
->blocks
;
540 return (struct midgard_block
*) lst
;
543 /* Pretty printer for internal Midgard IR */
546 print_mir_source(int source
)
548 if (source
>= SSA_FIXED_MINIMUM
) {
549 /* Specific register */
550 int reg
= SSA_REG_FROM_FIXED(source
);
552 /* TODO: Moving threshold */
553 if (reg
> 16 && reg
< 24)
554 printf("u%d", 23 - reg
);
558 printf("%d", source
);
563 print_mir_instruction(midgard_instruction
*ins
)
569 midgard_alu_op op
= ins
->alu
.op
;
570 const char *name
= alu_opcode_names
[op
];
573 printf("%d.", ins
->unit
);
575 printf("%s", name
? name
: "??");
579 case TAG_LOAD_STORE_4
: {
580 midgard_load_store_op op
= ins
->load_store
.op
;
581 const char *name
= load_store_opcode_names
[op
];
588 case TAG_TEXTURE_4
: {
597 ssa_args
*args
= &ins
->ssa_args
;
599 printf(" %d, ", args
->dest
);
601 print_mir_source(args
->src0
);
604 if (args
->inline_constant
)
605 printf("#%d", ins
->inline_constant
);
607 print_mir_source(args
->src1
);
609 if (ins
->has_constants
)
610 printf(" <%f, %f, %f, %f>", ins
->constants
[0], ins
->constants
[1], ins
->constants
[2], ins
->constants
[3]);
616 print_mir_block(midgard_block
*block
)
620 mir_foreach_instr_in_block(block
, ins
) {
621 print_mir_instruction(ins
);
630 attach_constants(compiler_context
*ctx
, midgard_instruction
*ins
, void *constants
, int name
)
632 ins
->has_constants
= true;
633 memcpy(&ins
->constants
, constants
, 16);
635 /* If this is the special blend constant, mark this instruction */
637 if (ctx
->is_blend
&& ctx
->blend_constant_number
== name
)
638 ins
->has_blend_constant
= true;
642 glsl_type_size(const struct glsl_type
*type
)
644 return glsl_count_attribute_slots(type
, false);
647 /* Lower fdot2 to a vector multiplication followed by channel addition */
649 midgard_nir_lower_fdot2_body(nir_builder
*b
, nir_alu_instr
*alu
)
651 if (alu
->op
!= nir_op_fdot2
)
654 b
->cursor
= nir_before_instr(&alu
->instr
);
656 nir_ssa_def
*src0
= nir_ssa_for_alu_src(b
, alu
, 0);
657 nir_ssa_def
*src1
= nir_ssa_for_alu_src(b
, alu
, 1);
659 nir_ssa_def
*product
= nir_fmul(b
, src0
, src1
);
661 nir_ssa_def
*sum
= nir_fadd(b
,
662 nir_channel(b
, product
, 0),
663 nir_channel(b
, product
, 1));
665 /* Replace the fdot2 with this sum */
666 nir_ssa_def_rewrite_uses(&alu
->dest
.dest
.ssa
, nir_src_for_ssa(sum
));
670 midgard_nir_lower_fdot2(nir_shader
*shader
)
672 bool progress
= false;
674 nir_foreach_function(function
, shader
) {
675 if (!function
->impl
) continue;
678 nir_builder
*b
= &_b
;
679 nir_builder_init(b
, function
->impl
);
681 nir_foreach_block(block
, function
->impl
) {
682 nir_foreach_instr_safe(instr
, block
) {
683 if (instr
->type
!= nir_instr_type_alu
) continue;
685 nir_alu_instr
*alu
= nir_instr_as_alu(instr
);
686 midgard_nir_lower_fdot2_body(b
, alu
);
692 nir_metadata_preserve(function
->impl
, nir_metadata_block_index
| nir_metadata_dominance
);
700 optimise_nir(nir_shader
*nir
)
704 NIR_PASS(progress
, nir
, nir_lower_regs_to_ssa
);
705 NIR_PASS(progress
, nir
, midgard_nir_lower_fdot2
);
707 nir_lower_tex_options lower_tex_options
= {
711 NIR_PASS(progress
, nir
, nir_lower_tex
, &lower_tex_options
);
716 NIR_PASS(progress
, nir
, midgard_nir_lower_algebraic
);
717 NIR_PASS(progress
, nir
, nir_lower_io
, nir_var_all
, glsl_type_size
, 0);
718 NIR_PASS(progress
, nir
, nir_lower_var_copies
);
719 NIR_PASS(progress
, nir
, nir_lower_vars_to_ssa
);
721 NIR_PASS(progress
, nir
, nir_copy_prop
);
722 NIR_PASS(progress
, nir
, nir_opt_dce
);
723 NIR_PASS(progress
, nir
, nir_opt_dead_cf
);
724 NIR_PASS(progress
, nir
, nir_opt_cse
);
725 NIR_PASS(progress
, nir
, nir_opt_peephole_select
, 64, false, true);
726 NIR_PASS(progress
, nir
, nir_opt_algebraic
);
727 NIR_PASS(progress
, nir
, nir_opt_constant_folding
);
728 NIR_PASS(progress
, nir
, nir_opt_undef
);
729 NIR_PASS(progress
, nir
, nir_opt_loop_unroll
,
732 nir_var_function_temp
);
734 /* TODO: Enable vectorize when merged upstream */
735 // NIR_PASS(progress, nir, nir_opt_vectorize);
738 /* Must be run at the end to prevent creation of fsin/fcos ops */
739 NIR_PASS(progress
, nir
, midgard_nir_scale_trig
);
744 NIR_PASS(progress
, nir
, nir_opt_dce
);
745 NIR_PASS(progress
, nir
, nir_opt_algebraic
);
746 NIR_PASS(progress
, nir
, nir_opt_constant_folding
);
747 NIR_PASS(progress
, nir
, nir_copy_prop
);
750 NIR_PASS(progress
, nir
, nir_opt_algebraic_late
);
753 NIR_PASS(progress
, nir
, nir_lower_to_source_mods
, nir_lower_all_source_mods
);
754 NIR_PASS(progress
, nir
, nir_copy_prop
);
755 NIR_PASS(progress
, nir
, nir_opt_dce
);
757 /* We implement booleans as 32-bit 0/~0 */
758 NIR_PASS(progress
, nir
, nir_lower_bool_to_int32
);
760 /* Take us out of SSA */
761 NIR_PASS(progress
, nir
, nir_lower_locals_to_regs
);
762 NIR_PASS(progress
, nir
, nir_convert_from_ssa
, true);
764 /* We are a vector architecture; write combine where possible */
765 NIR_PASS(progress
, nir
, nir_move_vec_src_uses_to_dest
);
766 NIR_PASS(progress
, nir
, nir_lower_vec_to_movs
);
768 NIR_PASS(progress
, nir
, nir_opt_dce
);
771 /* Front-half of aliasing the SSA slots, merely by inserting the flag in the
772 * appropriate hash table. Intentional off-by-one to avoid confusing NULL with
773 * r0. See the comments in compiler_context */
776 alias_ssa(compiler_context
*ctx
, int dest
, int src
)
778 _mesa_hash_table_u64_insert(ctx
->ssa_to_alias
, dest
+ 1, (void *) ((uintptr_t) src
+ 1));
779 _mesa_set_add(ctx
->leftover_ssa_to_alias
, (void *) (uintptr_t) (dest
+ 1));
782 /* ...or undo it, after which the original index will be used (dummy move should be emitted alongside this) */
785 unalias_ssa(compiler_context
*ctx
, int dest
)
787 _mesa_hash_table_u64_remove(ctx
->ssa_to_alias
, dest
+ 1);
788 /* TODO: Remove from leftover or no? */
792 midgard_pin_output(compiler_context
*ctx
, int index
, int reg
)
794 _mesa_hash_table_u64_insert(ctx
->ssa_to_register
, index
+ 1, (void *) ((uintptr_t) reg
+ 1));
798 midgard_is_pinned(compiler_context
*ctx
, int index
)
800 return _mesa_hash_table_u64_search(ctx
->ssa_to_register
, index
+ 1) != NULL
;
803 /* Do not actually emit a load; instead, cache the constant for inlining */
806 emit_load_const(compiler_context
*ctx
, nir_load_const_instr
*instr
)
808 nir_ssa_def def
= instr
->def
;
810 float *v
= ralloc_array(NULL
, float, 4);
811 memcpy(v
, &instr
->value
.f32
, 4 * sizeof(float));
812 _mesa_hash_table_u64_insert(ctx
->ssa_constants
, def
.index
+ 1, v
);
815 /* Duplicate bits to convert sane 4-bit writemask to obscure 8-bit format (or
819 expand_writemask(unsigned mask
)
823 for (int i
= 0; i
< 4; ++i
)
831 squeeze_writemask(unsigned mask
)
835 for (int i
= 0; i
< 4; ++i
)
836 if (mask
& (3 << (2 * i
)))
843 /* Determines effective writemask, taking quirks and expansion into account */
845 effective_writemask(midgard_vector_alu
*alu
)
847 /* Channel count is off-by-one to fit in two-bits (0 channel makes no
850 unsigned channel_count
= GET_CHANNEL_COUNT(alu_opcode_props
[alu
->op
]);
852 /* If there is a fixed channel count, construct the appropriate mask */
855 return (1 << channel_count
) - 1;
857 /* Otherwise, just squeeze the existing mask */
858 return squeeze_writemask(alu
->mask
);
862 find_or_allocate_temp(compiler_context
*ctx
, unsigned hash
)
864 if ((hash
< 0) || (hash
>= SSA_FIXED_MINIMUM
))
867 unsigned temp
= (uintptr_t) _mesa_hash_table_u64_search(ctx
->hash_to_temp
, hash
+ 1);
872 /* If no temp is find, allocate one */
873 temp
= ctx
->temp_count
++;
874 ctx
->max_hash
= MAX2(ctx
->max_hash
, hash
);
876 _mesa_hash_table_u64_insert(ctx
->hash_to_temp
, hash
+ 1, (void *) ((uintptr_t) temp
+ 1));
882 nir_src_index(compiler_context
*ctx
, nir_src
*src
)
885 return src
->ssa
->index
;
887 return ctx
->func
->impl
->ssa_alloc
+ src
->reg
.reg
->index
;
891 nir_dest_index(compiler_context
*ctx
, nir_dest
*dst
)
894 return dst
->ssa
.index
;
896 return ctx
->func
->impl
->ssa_alloc
+ dst
->reg
.reg
->index
;
900 nir_alu_src_index(compiler_context
*ctx
, nir_alu_src
*src
)
902 return nir_src_index(ctx
, &src
->src
);
905 /* Midgard puts conditionals in r31.w; move an arbitrary source (the output of
906 * a conditional test) into that register */
909 emit_condition(compiler_context
*ctx
, nir_src
*src
, bool for_branch
)
911 /* XXX: Force component correct */
912 int condition
= nir_src_index(ctx
, src
);
914 /* There is no boolean move instruction. Instead, we simulate a move by
915 * ANDing the condition with itself to get it into r31.w */
917 midgard_instruction ins
= {
919 .unit
= for_branch
? UNIT_SMUL
: UNIT_SADD
, /* TODO: DEDUCE THIS */
923 .dest
= SSA_FIXED_REGISTER(31),
926 .op
= midgard_alu_op_iand
,
927 .reg_mode
= midgard_reg_mode_full
,
928 .dest_override
= midgard_dest_override_none
,
929 .mask
= (0x3 << 6), /* w */
930 .src1
= vector_alu_srco_unsigned(blank_alu_src_xxxx
),
931 .src2
= vector_alu_srco_unsigned(blank_alu_src_xxxx
)
935 emit_mir_instruction(ctx
, ins
);
938 #define ALU_CASE(nir, _op) \
940 op = midgard_alu_op_##_op; \
944 emit_alu(compiler_context
*ctx
, nir_alu_instr
*instr
)
946 bool is_ssa
= instr
->dest
.dest
.is_ssa
;
948 unsigned dest
= nir_dest_index(ctx
, &instr
->dest
.dest
);
949 unsigned nr_components
= is_ssa
? instr
->dest
.dest
.ssa
.num_components
: instr
->dest
.dest
.reg
.reg
->num_components
;
950 unsigned nr_inputs
= nir_op_infos
[instr
->op
].num_inputs
;
952 /* Most Midgard ALU ops have a 1:1 correspondance to NIR ops; these are
953 * supported. A few do not and are commented for now. Also, there are a
954 * number of NIR ops which Midgard does not support and need to be
955 * lowered, also TODO. This switch block emits the opcode and calling
956 * convention of the Midgard instruction; actual packing is done in
962 ALU_CASE(fadd
, fadd
);
963 ALU_CASE(fmul
, fmul
);
964 ALU_CASE(fmin
, fmin
);
965 ALU_CASE(fmax
, fmax
);
966 ALU_CASE(imin
, imin
);
967 ALU_CASE(imax
, imax
);
968 ALU_CASE(fmov
, fmov
);
969 ALU_CASE(ffloor
, ffloor
);
970 ALU_CASE(fround_even
, froundeven
);
971 ALU_CASE(ftrunc
, ftrunc
);
972 ALU_CASE(fceil
, fceil
);
973 ALU_CASE(fdot3
, fdot3
);
974 ALU_CASE(fdot4
, fdot4
);
975 ALU_CASE(iadd
, iadd
);
976 ALU_CASE(isub
, isub
);
977 ALU_CASE(imul
, imul
);
979 /* XXX: Use fmov, not imov, since imov was causing major
980 * issues with texture precision? XXX research */
981 ALU_CASE(imov
, fmov
);
983 ALU_CASE(feq32
, feq
);
984 ALU_CASE(fne32
, fne
);
985 ALU_CASE(flt32
, flt
);
986 ALU_CASE(ieq32
, ieq
);
987 ALU_CASE(ine32
, ine
);
988 ALU_CASE(ilt32
, ilt
);
990 /* Likewise, we don't have a dedicated f2b32 instruction, but
991 * we can do a "not equal to 0.0" test. Since an inline
992 * constant vec4(0.0) is the default, we don't need to do any
993 * special lowering */
995 ALU_CASE(f2b32
, fne
);
997 ALU_CASE(frcp
, frcp
);
998 ALU_CASE(frsq
, frsqrt
);
999 ALU_CASE(fsqrt
, fsqrt
);
1000 ALU_CASE(fpow
, fpow
);
1001 ALU_CASE(fexp2
, fexp2
);
1002 ALU_CASE(flog2
, flog2
);
1004 ALU_CASE(f2i32
, f2i
);
1005 ALU_CASE(f2u32
, f2u
);
1006 ALU_CASE(i2f32
, i2f
);
1007 ALU_CASE(u2f32
, u2f
);
1009 ALU_CASE(fsin
, fsin
);
1010 ALU_CASE(fcos
, fcos
);
1012 ALU_CASE(iand
, iand
);
1014 ALU_CASE(ixor
, ixor
);
1015 ALU_CASE(inot
, inot
);
1016 ALU_CASE(ishl
, ishl
);
1017 ALU_CASE(ishr
, iasr
);
1018 ALU_CASE(ushr
, ilsr
);
1020 ALU_CASE(b32all_fequal2
, fball_eq
);
1021 ALU_CASE(b32all_fequal3
, fball_eq
);
1022 ALU_CASE(b32all_fequal4
, fball_eq
);
1024 ALU_CASE(b32any_fnequal2
, fbany_neq
);
1025 ALU_CASE(b32any_fnequal3
, fbany_neq
);
1026 ALU_CASE(b32any_fnequal4
, fbany_neq
);
1028 ALU_CASE(b32all_iequal2
, iball_eq
);
1029 ALU_CASE(b32all_iequal3
, iball_eq
);
1030 ALU_CASE(b32all_iequal4
, iball_eq
);
1032 ALU_CASE(b32any_inequal2
, ibany_neq
);
1033 ALU_CASE(b32any_inequal3
, ibany_neq
);
1034 ALU_CASE(b32any_inequal4
, ibany_neq
);
1036 /* For greater-or-equal, we use less-or-equal and flip the
1039 case nir_op_ige32
: {
1040 op
= midgard_alu_op_ile
;
1042 /* Swap via temporary */
1043 nir_alu_src temp
= instr
->src
[1];
1044 instr
->src
[1] = instr
->src
[0];
1045 instr
->src
[0] = temp
;
1050 case nir_op_b32csel
: {
1051 op
= midgard_alu_op_fcsel
;
1053 /* csel works as a two-arg in Midgard, since the condition is hardcoded in r31.w */
1056 emit_condition(ctx
, &instr
->src
[0].src
, false);
1058 /* The condition is the first argument; move the other
1059 * arguments up one to be a binary instruction for
1062 memmove(instr
->src
, instr
->src
+ 1, 2 * sizeof(nir_alu_src
));
1066 /* We don't have a native b2f32 instruction. Instead, like many GPUs,
1067 * we exploit booleans as 0/~0 for false/true, and correspondingly AND
1068 * by 1.0 to do the type conversion. For the moment, prime us to emit:
1070 * iand [whatever], #0
1072 * At the end of emit_alu (as MIR), we'll fix-up the constant */
1074 case nir_op_b2f32
: {
1075 op
= midgard_alu_op_iand
;
1080 DBG("Unhandled ALU op %s\n", nir_op_infos
[instr
->op
].name
);
1085 /* Fetch unit, quirks, etc information */
1086 unsigned opcode_props
= alu_opcode_props
[op
];
1087 bool quirk_flipped_r24
= opcode_props
& QUIRK_FLIPPED_R24
;
1089 /* Initialise fields common between scalar/vector instructions */
1090 midgard_outmod outmod
= instr
->dest
.saturate
? midgard_outmod_sat
: midgard_outmod_none
;
1092 /* src0 will always exist afaik, but src1 will not for 1-argument
1093 * instructions. The latter can only be fetched if the instruction
1094 * needs it, or else we may segfault. */
1096 unsigned src0
= nir_alu_src_index(ctx
, &instr
->src
[0]);
1097 unsigned src1
= nr_inputs
== 2 ? nir_alu_src_index(ctx
, &instr
->src
[1]) : SSA_UNUSED_0
;
1099 /* Rather than use the instruction generation helpers, we do it
1100 * ourselves here to avoid the mess */
1102 midgard_instruction ins
= {
1105 .src0
= quirk_flipped_r24
? SSA_UNUSED_1
: src0
,
1106 .src1
= quirk_flipped_r24
? src0
: src1
,
1111 nir_alu_src
*nirmods
[2] = { NULL
};
1113 if (nr_inputs
== 2) {
1114 nirmods
[0] = &instr
->src
[0];
1115 nirmods
[1] = &instr
->src
[1];
1116 } else if (nr_inputs
== 1) {
1117 nirmods
[quirk_flipped_r24
] = &instr
->src
[0];
1122 midgard_vector_alu alu
= {
1124 .reg_mode
= midgard_reg_mode_full
,
1125 .dest_override
= midgard_dest_override_none
,
1128 /* Writemask only valid for non-SSA NIR */
1129 .mask
= expand_writemask((1 << nr_components
) - 1),
1131 .src1
= vector_alu_srco_unsigned(vector_alu_modifiers(nirmods
[0])),
1132 .src2
= vector_alu_srco_unsigned(vector_alu_modifiers(nirmods
[1])),
1135 /* Apply writemask if non-SSA, keeping in mind that we can't write to components that don't exist */
1138 alu
.mask
&= expand_writemask(instr
->dest
.write_mask
);
1142 /* Late fixup for emulated instructions */
1144 if (instr
->op
== nir_op_b2f32
) {
1145 /* Presently, our second argument is an inline #0 constant.
1146 * Switch over to an embedded 1.0 constant (that can't fit
1147 * inline, since we're 32-bit, not 16-bit like the inline
1150 ins
.ssa_args
.inline_constant
= false;
1151 ins
.ssa_args
.src1
= SSA_FIXED_REGISTER(REGISTER_CONSTANT
);
1152 ins
.has_constants
= true;
1153 ins
.constants
[0] = 1.0;
1155 ins
.alu
.src2
= vector_alu_srco_unsigned(blank_alu_src_xxxx
);
1158 if ((opcode_props
& UNITS_ALL
) == UNIT_VLUT
) {
1159 /* To avoid duplicating the lookup tables (probably), true LUT
1160 * instructions can only operate as if they were scalars. Lower
1161 * them here by changing the component. */
1163 uint8_t original_swizzle
[4];
1164 memcpy(original_swizzle
, nirmods
[0]->swizzle
, sizeof(nirmods
[0]->swizzle
));
1166 for (int i
= 0; i
< nr_components
; ++i
) {
1167 ins
.alu
.mask
= (0x3) << (2 * i
); /* Mask the associated component */
1169 for (int j
= 0; j
< 4; ++j
)
1170 nirmods
[0]->swizzle
[j
] = original_swizzle
[i
]; /* Pull from the correct component */
1172 ins
.alu
.src1
= vector_alu_srco_unsigned(vector_alu_modifiers(nirmods
[0]));
1173 emit_mir_instruction(ctx
, ins
);
1176 emit_mir_instruction(ctx
, ins
);
1183 emit_intrinsic(compiler_context
*ctx
, nir_intrinsic_instr
*instr
)
1185 nir_const_value
*const_offset
;
1186 unsigned offset
, reg
;
1188 switch (instr
->intrinsic
) {
1189 case nir_intrinsic_discard_if
:
1190 emit_condition(ctx
, &instr
->src
[0], true);
1194 case nir_intrinsic_discard
: {
1195 bool conditional
= instr
->intrinsic
== nir_intrinsic_discard_if
;
1196 struct midgard_instruction discard
= v_branch(conditional
, false);
1197 discard
.branch
.target_type
= TARGET_DISCARD
;
1198 emit_mir_instruction(ctx
, discard
);
1200 ctx
->can_discard
= true;
1204 case nir_intrinsic_load_uniform
:
1205 case nir_intrinsic_load_input
:
1206 const_offset
= nir_src_as_const_value(instr
->src
[0]);
1207 assert (const_offset
&& "no indirect inputs");
1209 offset
= nir_intrinsic_base(instr
) + const_offset
->u32
[0];
1211 reg
= nir_dest_index(ctx
, &instr
->dest
);
1213 if (instr
->intrinsic
== nir_intrinsic_load_uniform
&& !ctx
->is_blend
) {
1214 /* TODO: half-floats */
1216 int uniform_offset
= 0;
1218 if (offset
>= SPECIAL_UNIFORM_BASE
) {
1219 /* XXX: Resolve which uniform */
1222 /* Offset away from the special
1225 void *entry
= _mesa_hash_table_u64_search(ctx
->uniform_nir_to_mdg
, offset
+ 1);
1229 DBG("WARNING: Unknown uniform %d\n", offset
);
1233 uniform_offset
= (uintptr_t) (entry
) - 1;
1234 uniform_offset
+= ctx
->special_uniforms
;
1237 if (uniform_offset
< ctx
->uniform_cutoff
) {
1238 /* Fast path: For the first 16 uniform,
1239 * accesses are 0-cycle, since they're
1240 * just a register fetch in the usual
1241 * case. So, we alias the registers
1242 * while we're still in SSA-space */
1244 int reg_slot
= 23 - uniform_offset
;
1245 alias_ssa(ctx
, reg
, SSA_FIXED_REGISTER(reg_slot
));
1247 /* Otherwise, read from the 'special'
1248 * UBO to access higher-indexed
1249 * uniforms, at a performance cost */
1251 midgard_instruction ins
= m_load_uniform_32(reg
, uniform_offset
);
1253 /* TODO: Don't split */
1254 ins
.load_store
.varying_parameters
= (uniform_offset
& 7) << 7;
1255 ins
.load_store
.address
= uniform_offset
>> 3;
1257 ins
.load_store
.unknown
= 0x1E00; /* xxx: what is this? */
1258 emit_mir_instruction(ctx
, ins
);
1260 } else if (ctx
->stage
== MESA_SHADER_FRAGMENT
&& !ctx
->is_blend
) {
1261 /* XXX: Half-floats? */
1262 /* TODO: swizzle, mask */
1264 midgard_instruction ins
= m_load_vary_32(reg
, offset
);
1266 midgard_varying_parameter p
= {
1268 .interpolation
= midgard_interp_default
,
1269 .flat
= /*var->data.interpolation == INTERP_MODE_FLAT*/ 0
1273 memcpy(&u
, &p
, sizeof(p
));
1274 ins
.load_store
.varying_parameters
= u
;
1276 ins
.load_store
.unknown
= 0x1e9e; /* xxx: what is this? */
1277 emit_mir_instruction(ctx
, ins
);
1278 } else if (ctx
->is_blend
&& instr
->intrinsic
== nir_intrinsic_load_uniform
) {
1279 /* Constant encoded as a pinned constant */
1281 midgard_instruction ins
= v_fmov(SSA_FIXED_REGISTER(REGISTER_CONSTANT
), blank_alu_src
, reg
);
1282 ins
.has_constants
= true;
1283 ins
.has_blend_constant
= true;
1284 emit_mir_instruction(ctx
, ins
);
1285 } else if (ctx
->is_blend
) {
1286 /* For blend shaders, a load might be
1287 * translated various ways depending on what
1288 * we're loading. Figure out how this is used */
1290 nir_variable
*out
= NULL
;
1292 nir_foreach_variable(var
, &ctx
->nir
->inputs
) {
1293 int drvloc
= var
->data
.driver_location
;
1295 if (nir_intrinsic_base(instr
) == drvloc
) {
1303 if (out
->data
.location
== VARYING_SLOT_COL0
) {
1304 /* Source color preloaded to r0 */
1306 midgard_pin_output(ctx
, reg
, 0);
1307 } else if (out
->data
.location
== VARYING_SLOT_COL1
) {
1308 /* Destination color must be read from framebuffer */
1310 midgard_instruction ins
= m_load_color_buffer_8(reg
, 0);
1311 ins
.load_store
.swizzle
= 0; /* xxxx */
1313 /* Read each component sequentially */
1315 for (int c
= 0; c
< 4; ++c
) {
1316 ins
.load_store
.mask
= (1 << c
);
1317 ins
.load_store
.unknown
= c
;
1318 emit_mir_instruction(ctx
, ins
);
1321 /* vadd.u2f hr2, abs(hr2), #0 */
1323 midgard_vector_alu_src alu_src
= blank_alu_src
;
1325 alu_src
.half
= true;
1327 midgard_instruction u2f
= {
1331 .src1
= SSA_UNUSED_0
,
1333 .inline_constant
= true
1336 .op
= midgard_alu_op_u2f
,
1337 .reg_mode
= midgard_reg_mode_half
,
1338 .dest_override
= midgard_dest_override_none
,
1340 .src1
= vector_alu_srco_unsigned(alu_src
),
1341 .src2
= vector_alu_srco_unsigned(blank_alu_src
),
1345 emit_mir_instruction(ctx
, u2f
);
1347 /* vmul.fmul.sat r1, hr2, #0.00392151 */
1349 alu_src
.abs
= false;
1351 midgard_instruction fmul
= {
1353 .inline_constant
= _mesa_float_to_half(1.0 / 255.0),
1357 .src1
= SSA_UNUSED_0
,
1358 .inline_constant
= true
1361 .op
= midgard_alu_op_fmul
,
1362 .reg_mode
= midgard_reg_mode_full
,
1363 .dest_override
= midgard_dest_override_none
,
1364 .outmod
= midgard_outmod_sat
,
1366 .src1
= vector_alu_srco_unsigned(alu_src
),
1367 .src2
= vector_alu_srco_unsigned(blank_alu_src
),
1371 emit_mir_instruction(ctx
, fmul
);
1373 DBG("Unknown input in blend shader\n");
1376 } else if (ctx
->stage
== MESA_SHADER_VERTEX
) {
1377 midgard_instruction ins
= m_load_attr_32(reg
, offset
);
1378 ins
.load_store
.unknown
= 0x1E1E; /* XXX: What is this? */
1379 ins
.load_store
.mask
= (1 << instr
->num_components
) - 1;
1380 emit_mir_instruction(ctx
, ins
);
1382 DBG("Unknown load\n");
1388 case nir_intrinsic_store_output
:
1389 const_offset
= nir_src_as_const_value(instr
->src
[1]);
1390 assert(const_offset
&& "no indirect outputs");
1392 offset
= nir_intrinsic_base(instr
) + const_offset
->u32
[0];
1394 reg
= nir_src_index(ctx
, &instr
->src
[0]);
1396 if (ctx
->stage
== MESA_SHADER_FRAGMENT
) {
1397 /* gl_FragColor is not emitted with load/store
1398 * instructions. Instead, it gets plonked into
1399 * r0 at the end of the shader and we do the
1400 * framebuffer writeout dance. TODO: Defer
1403 midgard_pin_output(ctx
, reg
, 0);
1405 /* Save the index we're writing to for later reference
1406 * in the epilogue */
1408 ctx
->fragment_output
= reg
;
1409 } else if (ctx
->stage
== MESA_SHADER_VERTEX
) {
1410 /* Varyings are written into one of two special
1411 * varying register, r26 or r27. The register itself is selected as the register
1412 * in the st_vary instruction, minus the base of 26. E.g. write into r27 and then call st_vary(1)
1414 * Normally emitting fmov's is frowned upon,
1415 * but due to unique constraints of
1416 * REGISTER_VARYING, fmov emission + a
1417 * dedicated cleanup pass is the only way to
1418 * guarantee correctness when considering some
1419 * (common) edge cases XXX: FIXME */
1421 /* If this varying corresponds to a constant (why?!),
1422 * emit that now since it won't get picked up by
1423 * hoisting (since there is no corresponding move
1424 * emitted otherwise) */
1426 void *constant_value
= _mesa_hash_table_u64_search(ctx
->ssa_constants
, reg
+ 1);
1428 if (constant_value
) {
1429 /* Special case: emit the varying write
1430 * directly to r26 (looks funny in asm but it's
1431 * fine) and emit the store _now_. Possibly
1432 * slightly slower, but this is a really stupid
1433 * special case anyway (why on earth would you
1434 * have a constant varying? Your own fault for
1435 * slightly worse perf :P) */
1437 midgard_instruction ins
= v_fmov(SSA_FIXED_REGISTER(REGISTER_CONSTANT
), blank_alu_src
, SSA_FIXED_REGISTER(26));
1438 attach_constants(ctx
, &ins
, constant_value
, reg
+ 1);
1439 emit_mir_instruction(ctx
, ins
);
1441 midgard_instruction st
= m_store_vary_32(SSA_FIXED_REGISTER(0), offset
);
1442 st
.load_store
.unknown
= 0x1E9E; /* XXX: What is this? */
1443 emit_mir_instruction(ctx
, st
);
1445 /* Do not emit the varying yet -- instead, just mark down that we need to later */
1447 _mesa_hash_table_u64_insert(ctx
->ssa_varyings
, reg
+ 1, (void *) ((uintptr_t) (offset
+ 1)));
1450 DBG("Unknown store\n");
1456 case nir_intrinsic_load_alpha_ref_float
:
1457 assert(instr
->dest
.is_ssa
);
1459 float ref_value
= ctx
->alpha_ref
;
1461 float *v
= ralloc_array(NULL
, float, 4);
1462 memcpy(v
, &ref_value
, sizeof(float));
1463 _mesa_hash_table_u64_insert(ctx
->ssa_constants
, instr
->dest
.ssa
.index
+ 1, v
);
1468 printf ("Unhandled intrinsic\n");
1475 midgard_tex_format(enum glsl_sampler_dim dim
)
1478 case GLSL_SAMPLER_DIM_2D
:
1479 case GLSL_SAMPLER_DIM_EXTERNAL
:
1482 case GLSL_SAMPLER_DIM_3D
:
1485 case GLSL_SAMPLER_DIM_CUBE
:
1486 return TEXTURE_CUBE
;
1489 DBG("Unknown sampler dim type\n");
1496 emit_tex(compiler_context
*ctx
, nir_tex_instr
*instr
)
1499 //assert (!instr->sampler);
1500 //assert (!instr->texture_array_size);
1501 assert (instr
->op
== nir_texop_tex
);
1503 /* Allocate registers via a round robin scheme to alternate between the two registers */
1504 int reg
= ctx
->texture_op_count
& 1;
1505 int in_reg
= reg
, out_reg
= reg
;
1507 /* Make room for the reg */
1509 if (ctx
->texture_index
[reg
] > -1)
1510 unalias_ssa(ctx
, ctx
->texture_index
[reg
]);
1512 int texture_index
= instr
->texture_index
;
1513 int sampler_index
= texture_index
;
1515 for (unsigned i
= 0; i
< instr
->num_srcs
; ++i
) {
1516 switch (instr
->src
[i
].src_type
) {
1517 case nir_tex_src_coord
: {
1518 int index
= nir_src_index(ctx
, &instr
->src
[i
].src
);
1520 midgard_vector_alu_src alu_src
= blank_alu_src
;
1521 alu_src
.swizzle
= (COMPONENT_Y
<< 2);
1523 midgard_instruction ins
= v_fmov(index
, alu_src
, SSA_FIXED_REGISTER(REGISTER_TEXTURE_BASE
+ in_reg
));
1524 emit_mir_instruction(ctx
, ins
);
1526 //midgard_pin_output(ctx, index, REGISTER_TEXTURE_BASE + in_reg);
1532 DBG("Unknown source type\n");
1539 /* No helper to build texture words -- we do it all here */
1540 midgard_instruction ins
= {
1541 .type
= TAG_TEXTURE_4
,
1543 .op
= TEXTURE_OP_NORMAL
,
1544 .format
= midgard_tex_format(instr
->sampler_dim
),
1545 .texture_handle
= texture_index
,
1546 .sampler_handle
= sampler_index
,
1548 /* TODO: Don't force xyzw */
1549 .swizzle
= SWIZZLE(COMPONENT_X
, COMPONENT_Y
, COMPONENT_Z
, COMPONENT_W
),
1561 /* Assume we can continue; hint it out later */
1566 /* Set registers to read and write from the same place */
1567 ins
.texture
.in_reg_select
= in_reg
;
1568 ins
.texture
.out_reg_select
= out_reg
;
1570 /* TODO: Dynamic swizzle input selection, half-swizzles? */
1571 if (instr
->sampler_dim
== GLSL_SAMPLER_DIM_3D
) {
1572 ins
.texture
.in_reg_swizzle_right
= COMPONENT_X
;
1573 ins
.texture
.in_reg_swizzle_left
= COMPONENT_Y
;
1574 //ins.texture.in_reg_swizzle_third = COMPONENT_Z;
1576 ins
.texture
.in_reg_swizzle_left
= COMPONENT_X
;
1577 ins
.texture
.in_reg_swizzle_right
= COMPONENT_Y
;
1578 //ins.texture.in_reg_swizzle_third = COMPONENT_X;
1581 emit_mir_instruction(ctx
, ins
);
1583 /* Simultaneously alias the destination and emit a move for it. The move will be eliminated if possible */
1585 int o_reg
= REGISTER_TEXTURE_BASE
+ out_reg
, o_index
= nir_dest_index(ctx
, &instr
->dest
);
1586 alias_ssa(ctx
, o_index
, SSA_FIXED_REGISTER(o_reg
));
1587 ctx
->texture_index
[reg
] = o_index
;
1589 midgard_instruction ins2
= v_fmov(SSA_FIXED_REGISTER(o_reg
), blank_alu_src
, o_index
);
1590 emit_mir_instruction(ctx
, ins2
);
1592 /* Used for .cont and .last hinting */
1593 ctx
->texture_op_count
++;
1597 emit_jump(compiler_context
*ctx
, nir_jump_instr
*instr
)
1599 switch (instr
->type
) {
1600 case nir_jump_break
: {
1601 /* Emit a branch out of the loop */
1602 struct midgard_instruction br
= v_branch(false, false);
1603 br
.branch
.target_type
= TARGET_BREAK
;
1604 br
.branch
.target_break
= ctx
->current_loop
;
1605 emit_mir_instruction(ctx
, br
);
1612 DBG("Unknown jump type %d\n", instr
->type
);
1618 emit_instr(compiler_context
*ctx
, struct nir_instr
*instr
)
1620 switch (instr
->type
) {
1621 case nir_instr_type_load_const
:
1622 emit_load_const(ctx
, nir_instr_as_load_const(instr
));
1625 case nir_instr_type_intrinsic
:
1626 emit_intrinsic(ctx
, nir_instr_as_intrinsic(instr
));
1629 case nir_instr_type_alu
:
1630 emit_alu(ctx
, nir_instr_as_alu(instr
));
1633 case nir_instr_type_tex
:
1634 emit_tex(ctx
, nir_instr_as_tex(instr
));
1637 case nir_instr_type_jump
:
1638 emit_jump(ctx
, nir_instr_as_jump(instr
));
1641 case nir_instr_type_ssa_undef
:
1646 DBG("Unhandled instruction type\n");
1651 /* Determine the actual hardware from the index based on the RA results or special values */
1654 dealias_register(compiler_context
*ctx
, struct ra_graph
*g
, int reg
, int maxreg
)
1656 if (reg
>= SSA_FIXED_MINIMUM
)
1657 return SSA_REG_FROM_FIXED(reg
);
1660 assert(reg
< maxreg
);
1661 int r
= ra_get_node_reg(g
, reg
);
1662 ctx
->work_registers
= MAX2(ctx
->work_registers
, r
);
1667 /* fmov style unused */
1669 return REGISTER_UNUSED
;
1671 /* lut style unused */
1673 return REGISTER_UNUSED
;
1676 DBG("Unknown SSA register alias %d\n", reg
);
1683 midgard_ra_select_callback(struct ra_graph
*g
, BITSET_WORD
*regs
, void *data
)
1685 /* Choose the first available register to minimise reported register pressure */
1687 for (int i
= 0; i
< 16; ++i
) {
1688 if (BITSET_TEST(regs
, i
)) {
1698 midgard_is_live_in_instr(midgard_instruction
*ins
, int src
)
1700 if (ins
->ssa_args
.src0
== src
) return true;
1701 if (ins
->ssa_args
.src1
== src
) return true;
1707 is_live_after(compiler_context
*ctx
, midgard_block
*block
, midgard_instruction
*start
, int src
)
1709 /* Check the rest of the block for liveness */
1710 mir_foreach_instr_in_block_from(block
, ins
, mir_next_op(start
)) {
1711 if (midgard_is_live_in_instr(ins
, src
))
1715 /* Check the rest of the blocks for liveness */
1716 mir_foreach_block_from(ctx
, mir_next_block(block
), b
) {
1717 mir_foreach_instr_in_block(b
, ins
) {
1718 if (midgard_is_live_in_instr(ins
, src
))
1723 /* TODO: How does control flow interact in complex shaders? */
1729 allocate_registers(compiler_context
*ctx
)
1731 /* First, initialize the RA */
1732 struct ra_regs
*regs
= ra_alloc_reg_set(NULL
, 32, true);
1734 /* Create a primary (general purpose) class, as well as special purpose
1735 * pipeline register classes */
1737 int primary_class
= ra_alloc_reg_class(regs
);
1738 int varying_class
= ra_alloc_reg_class(regs
);
1740 /* Add the full set of work registers */
1741 int work_count
= 16 - MAX2((ctx
->uniform_cutoff
- 8), 0);
1742 for (int i
= 0; i
< work_count
; ++i
)
1743 ra_class_add_reg(regs
, primary_class
, i
);
1745 /* Add special registers */
1746 ra_class_add_reg(regs
, varying_class
, REGISTER_VARYING_BASE
);
1747 ra_class_add_reg(regs
, varying_class
, REGISTER_VARYING_BASE
+ 1);
1749 /* We're done setting up */
1750 ra_set_finalize(regs
, NULL
);
1752 /* Transform the MIR into squeezed index form */
1753 mir_foreach_block(ctx
, block
) {
1754 mir_foreach_instr_in_block(block
, ins
) {
1755 if (ins
->compact_branch
) continue;
1757 ins
->ssa_args
.src0
= find_or_allocate_temp(ctx
, ins
->ssa_args
.src0
);
1758 ins
->ssa_args
.src1
= find_or_allocate_temp(ctx
, ins
->ssa_args
.src1
);
1759 ins
->ssa_args
.dest
= find_or_allocate_temp(ctx
, ins
->ssa_args
.dest
);
1761 if (midgard_debug
& MIDGARD_DBG_SHADERS
)
1762 print_mir_block(block
);
1765 /* Let's actually do register allocation */
1766 int nodes
= ctx
->temp_count
;
1767 struct ra_graph
*g
= ra_alloc_interference_graph(regs
, nodes
);
1769 /* Set everything to the work register class, unless it has somewhere
1772 mir_foreach_block(ctx
, block
) {
1773 mir_foreach_instr_in_block(block
, ins
) {
1774 if (ins
->compact_branch
) continue;
1776 if (ins
->ssa_args
.dest
< 0) continue;
1778 if (ins
->ssa_args
.dest
>= SSA_FIXED_MINIMUM
) continue;
1780 int class = primary_class
;
1782 ra_set_node_class(g
, ins
->ssa_args
.dest
, class);
1786 for (int index
= 0; index
<= ctx
->max_hash
; ++index
) {
1787 unsigned temp
= (uintptr_t) _mesa_hash_table_u64_search(ctx
->ssa_to_register
, index
+ 1);
1790 unsigned reg
= temp
- 1;
1791 int t
= find_or_allocate_temp(ctx
, index
);
1792 ra_set_node_reg(g
, t
, reg
);
1796 /* Determine liveness */
1798 int *live_start
= malloc(nodes
* sizeof(int));
1799 int *live_end
= malloc(nodes
* sizeof(int));
1801 /* Initialize as non-existent */
1803 for (int i
= 0; i
< nodes
; ++i
) {
1804 live_start
[i
] = live_end
[i
] = -1;
1809 mir_foreach_block(ctx
, block
) {
1810 mir_foreach_instr_in_block(block
, ins
) {
1811 if (ins
->compact_branch
) continue;
1813 if (ins
->ssa_args
.dest
< SSA_FIXED_MINIMUM
) {
1814 /* If this destination is not yet live, it is now since we just wrote it */
1816 int dest
= ins
->ssa_args
.dest
;
1818 if (live_start
[dest
] == -1)
1819 live_start
[dest
] = d
;
1822 /* Since we just used a source, the source might be
1823 * dead now. Scan the rest of the block for
1824 * invocations, and if there are none, the source dies
1827 int sources
[2] = { ins
->ssa_args
.src0
, ins
->ssa_args
.src1
};
1829 for (int src
= 0; src
< 2; ++src
) {
1830 int s
= sources
[src
];
1832 if (s
< 0) continue;
1834 if (s
>= SSA_FIXED_MINIMUM
) continue;
1836 if (!is_live_after(ctx
, block
, ins
, s
)) {
1845 /* If a node still hasn't been killed, kill it now */
1847 for (int i
= 0; i
< nodes
; ++i
) {
1848 /* live_start == -1 most likely indicates a pinned output */
1850 if (live_end
[i
] == -1)
1854 /* Setup interference between nodes that are live at the same time */
1856 for (int i
= 0; i
< nodes
; ++i
) {
1857 for (int j
= i
+ 1; j
< nodes
; ++j
) {
1858 if (!(live_start
[i
] >= live_end
[j
] || live_start
[j
] >= live_end
[i
]))
1859 ra_add_node_interference(g
, i
, j
);
1863 ra_set_select_reg_callback(g
, midgard_ra_select_callback
, NULL
);
1865 if (!ra_allocate(g
)) {
1866 DBG("Error allocating registers\n");
1874 mir_foreach_block(ctx
, block
) {
1875 mir_foreach_instr_in_block(block
, ins
) {
1876 if (ins
->compact_branch
) continue;
1878 ssa_args args
= ins
->ssa_args
;
1880 switch (ins
->type
) {
1882 ins
->registers
.src1_reg
= dealias_register(ctx
, g
, args
.src0
, nodes
);
1884 ins
->registers
.src2_imm
= args
.inline_constant
;
1886 if (args
.inline_constant
) {
1887 /* Encode inline 16-bit constant as a vector by default */
1889 ins
->registers
.src2_reg
= ins
->inline_constant
>> 11;
1891 int lower_11
= ins
->inline_constant
& ((1 << 12) - 1);
1893 uint16_t imm
= ((lower_11
>> 8) & 0x7) | ((lower_11
& 0xFF) << 3);
1894 ins
->alu
.src2
= imm
<< 2;
1896 ins
->registers
.src2_reg
= dealias_register(ctx
, g
, args
.src1
, nodes
);
1899 ins
->registers
.out_reg
= dealias_register(ctx
, g
, args
.dest
, nodes
);
1903 case TAG_LOAD_STORE_4
: {
1904 if (OP_IS_STORE(ins
->load_store
.op
)) {
1905 /* TODO: use ssa_args for store_vary */
1906 ins
->load_store
.reg
= 0;
1908 bool has_dest
= args
.dest
>= 0;
1909 int ssa_arg
= has_dest
? args
.dest
: args
.src0
;
1911 ins
->load_store
.reg
= dealias_register(ctx
, g
, ssa_arg
, nodes
);
1924 /* Midgard IR only knows vector ALU types, but we sometimes need to actually
1925 * use scalar ALU instructions, for functional or performance reasons. To do
1926 * this, we just demote vector ALU payloads to scalar. */
1929 component_from_mask(unsigned mask
)
1931 for (int c
= 0; c
< 4; ++c
) {
1932 if (mask
& (3 << (2 * c
)))
1941 is_single_component_mask(unsigned mask
)
1945 for (int c
= 0; c
< 4; ++c
)
1946 if (mask
& (3 << (2 * c
)))
1949 return components
== 1;
1952 /* Create a mask of accessed components from a swizzle to figure out vector
1956 swizzle_to_access_mask(unsigned swizzle
)
1958 unsigned component_mask
= 0;
1960 for (int i
= 0; i
< 4; ++i
) {
1961 unsigned c
= (swizzle
>> (2 * i
)) & 3;
1962 component_mask
|= (1 << c
);
1965 return component_mask
;
1969 vector_to_scalar_source(unsigned u
)
1971 midgard_vector_alu_src v
;
1972 memcpy(&v
, &u
, sizeof(v
));
1974 midgard_scalar_alu_src s
= {
1978 .component
= (v
.swizzle
& 3) << 1
1982 memcpy(&o
, &s
, sizeof(s
));
1984 return o
& ((1 << 6) - 1);
1987 static midgard_scalar_alu
1988 vector_to_scalar_alu(midgard_vector_alu v
, midgard_instruction
*ins
)
1990 /* The output component is from the mask */
1991 midgard_scalar_alu s
= {
1993 .src1
= vector_to_scalar_source(v
.src1
),
1994 .src2
= vector_to_scalar_source(v
.src2
),
1997 .output_full
= 1, /* TODO: Half */
1998 .output_component
= component_from_mask(v
.mask
) << 1,
2001 /* Inline constant is passed along rather than trying to extract it
2004 if (ins
->ssa_args
.inline_constant
) {
2006 int lower_11
= ins
->inline_constant
& ((1 << 12) - 1);
2007 imm
|= (lower_11
>> 9) & 3;
2008 imm
|= (lower_11
>> 6) & 4;
2009 imm
|= (lower_11
>> 2) & 0x38;
2010 imm
|= (lower_11
& 63) << 6;
2018 /* Midgard prefetches instruction types, so during emission we need to
2019 * lookahead too. Unless this is the last instruction, in which we return 1. Or
2020 * if this is the second to last and the last is an ALU, then it's also 1... */
2022 #define IS_ALU(tag) (tag == TAG_ALU_4 || tag == TAG_ALU_8 || \
2023 tag == TAG_ALU_12 || tag == TAG_ALU_16)
2025 #define EMIT_AND_COUNT(type, val) util_dynarray_append(emission, type, val); \
2026 bytes_emitted += sizeof(type)
2029 emit_binary_vector_instruction(midgard_instruction
*ains
,
2030 uint16_t *register_words
, int *register_words_count
,
2031 uint64_t *body_words
, size_t *body_size
, int *body_words_count
,
2032 size_t *bytes_emitted
)
2034 memcpy(®ister_words
[(*register_words_count
)++], &ains
->registers
, sizeof(ains
->registers
));
2035 *bytes_emitted
+= sizeof(midgard_reg_info
);
2037 body_size
[*body_words_count
] = sizeof(midgard_vector_alu
);
2038 memcpy(&body_words
[(*body_words_count
)++], &ains
->alu
, sizeof(ains
->alu
));
2039 *bytes_emitted
+= sizeof(midgard_vector_alu
);
2042 /* Checks for an SSA data hazard between two adjacent instructions, keeping in
2043 * mind that we are a vector architecture and we can write to different
2044 * components simultaneously */
2047 can_run_concurrent_ssa(midgard_instruction
*first
, midgard_instruction
*second
)
2049 /* Each instruction reads some registers and writes to a register. See
2050 * where the first writes */
2052 /* Figure out where exactly we wrote to */
2053 int source
= first
->ssa_args
.dest
;
2054 int source_mask
= first
->type
== TAG_ALU_4
? squeeze_writemask(first
->alu
.mask
) : 0xF;
2056 /* As long as the second doesn't read from the first, we're okay */
2057 if (second
->ssa_args
.src0
== source
) {
2058 if (first
->type
== TAG_ALU_4
) {
2059 /* Figure out which components we just read from */
2061 int q
= second
->alu
.src1
;
2062 midgard_vector_alu_src
*m
= (midgard_vector_alu_src
*) &q
;
2064 /* Check if there are components in common, and fail if so */
2065 if (swizzle_to_access_mask(m
->swizzle
) & source_mask
)
2072 if (second
->ssa_args
.src1
== source
)
2075 /* Otherwise, it's safe in that regard. Another data hazard is both
2076 * writing to the same place, of course */
2078 if (second
->ssa_args
.dest
== source
) {
2079 /* ...but only if the components overlap */
2080 int dest_mask
= second
->type
== TAG_ALU_4
? squeeze_writemask(second
->alu
.mask
) : 0xF;
2082 if (dest_mask
& source_mask
)
2092 midgard_instruction
**segment
, unsigned segment_size
,
2093 midgard_instruction
*ains
)
2095 for (int s
= 0; s
< segment_size
; ++s
)
2096 if (!can_run_concurrent_ssa(segment
[s
], ains
))
2104 /* Schedules, but does not emit, a single basic block. After scheduling, the
2105 * final tag and size of the block are known, which are necessary for branching
2108 static midgard_bundle
2109 schedule_bundle(compiler_context
*ctx
, midgard_block
*block
, midgard_instruction
*ins
, int *skip
)
2111 int instructions_emitted
= 0, instructions_consumed
= -1;
2112 midgard_bundle bundle
= { 0 };
2114 uint8_t tag
= ins
->type
;
2116 /* Default to the instruction's tag */
2119 switch (ins
->type
) {
2121 uint32_t control
= 0;
2122 size_t bytes_emitted
= sizeof(control
);
2124 /* TODO: Constant combining */
2125 int index
= 0, last_unit
= 0;
2127 /* Previous instructions, for the purpose of parallelism */
2128 midgard_instruction
*segment
[4] = {0};
2129 int segment_size
= 0;
2131 instructions_emitted
= -1;
2132 midgard_instruction
*pins
= ins
;
2135 midgard_instruction
*ains
= pins
;
2137 /* Advance instruction pointer */
2139 ains
= mir_next_op(pins
);
2143 /* Out-of-work condition */
2144 if ((struct list_head
*) ains
== &block
->instructions
)
2147 /* Ensure that the chain can continue */
2148 if (ains
->type
!= TAG_ALU_4
) break;
2150 /* According to the presentation "The ARM
2151 * Mali-T880 Mobile GPU" from HotChips 27,
2152 * there are two pipeline stages. Branching
2153 * position determined experimentally. Lines
2154 * are executed in parallel:
2157 * [ VADD ] [ SMUL ] [ LUT ] [ BRANCH ]
2159 * Verify that there are no ordering dependencies here.
2161 * TODO: Allow for parallelism!!!
2164 /* Pick a unit for it if it doesn't force a particular unit */
2166 int unit
= ains
->unit
;
2169 int op
= ains
->alu
.op
;
2170 int units
= alu_opcode_props
[op
];
2172 /* TODO: Promotion of scalars to vectors */
2173 int vector
= ((!is_single_component_mask(ains
->alu
.mask
)) || ((units
& UNITS_SCALAR
) == 0)) && (units
& UNITS_ANY_VECTOR
);
2176 assert(units
& UNITS_SCALAR
);
2179 if (last_unit
>= UNIT_VADD
) {
2180 if (units
& UNIT_VLUT
)
2185 if ((units
& UNIT_VMUL
) && !(control
& UNIT_VMUL
))
2187 else if ((units
& UNIT_VADD
) && !(control
& UNIT_VADD
))
2189 else if (units
& UNIT_VLUT
)
2195 if (last_unit
>= UNIT_VADD
) {
2196 if ((units
& UNIT_SMUL
) && !(control
& UNIT_SMUL
))
2198 else if (units
& UNIT_VLUT
)
2203 if ((units
& UNIT_SADD
) && !(control
& UNIT_SADD
) && !midgard_has_hazard(segment
, segment_size
, ains
))
2205 else if (units
& UNIT_SMUL
)
2206 unit
= ((units
& UNIT_VMUL
) && !(control
& UNIT_VMUL
)) ? UNIT_VMUL
: UNIT_SMUL
;
2207 else if ((units
& UNIT_VADD
) && !(control
& UNIT_VADD
))
2214 assert(unit
& units
);
2217 /* Late unit check, this time for encoding (not parallelism) */
2218 if (unit
<= last_unit
) break;
2220 /* Clear the segment */
2221 if (last_unit
< UNIT_VADD
&& unit
>= UNIT_VADD
)
2224 if (midgard_has_hazard(segment
, segment_size
, ains
))
2227 /* We're good to go -- emit the instruction */
2230 segment
[segment_size
++] = ains
;
2232 /* Only one set of embedded constants per
2233 * bundle possible; if we have more, we must
2234 * break the chain early, unfortunately */
2236 if (ains
->has_constants
) {
2237 if (bundle
.has_embedded_constants
) {
2238 /* ...but if there are already
2239 * constants but these are the
2240 * *same* constants, we let it
2243 if (memcmp(bundle
.constants
, ains
->constants
, sizeof(bundle
.constants
)))
2246 bundle
.has_embedded_constants
= true;
2247 memcpy(bundle
.constants
, ains
->constants
, sizeof(bundle
.constants
));
2249 /* If this is a blend shader special constant, track it for patching */
2250 if (ains
->has_blend_constant
)
2251 bundle
.has_blend_constant
= true;
2255 if (ains
->unit
& UNITS_ANY_VECTOR
) {
2256 emit_binary_vector_instruction(ains
, bundle
.register_words
,
2257 &bundle
.register_words_count
, bundle
.body_words
,
2258 bundle
.body_size
, &bundle
.body_words_count
, &bytes_emitted
);
2259 } else if (ains
->compact_branch
) {
2260 /* All of r0 has to be written out
2261 * along with the branch writeout.
2264 if (ains
->writeout
) {
2266 midgard_instruction ins
= v_fmov(0, blank_alu_src
, SSA_FIXED_REGISTER(0));
2267 ins
.unit
= UNIT_VMUL
;
2269 control
|= ins
.unit
;
2271 emit_binary_vector_instruction(&ins
, bundle
.register_words
,
2272 &bundle
.register_words_count
, bundle
.body_words
,
2273 bundle
.body_size
, &bundle
.body_words_count
, &bytes_emitted
);
2275 /* Analyse the group to see if r0 is written in full, on-time, without hanging dependencies*/
2276 bool written_late
= false;
2277 bool components
[4] = { 0 };
2278 uint16_t register_dep_mask
= 0;
2279 uint16_t written_mask
= 0;
2281 midgard_instruction
*qins
= ins
;
2282 for (int t
= 0; t
< index
; ++t
) {
2283 if (qins
->registers
.out_reg
!= 0) {
2284 /* Mark down writes */
2286 written_mask
|= (1 << qins
->registers
.out_reg
);
2288 /* Mark down the register dependencies for errata check */
2290 if (qins
->registers
.src1_reg
< 16)
2291 register_dep_mask
|= (1 << qins
->registers
.src1_reg
);
2293 if (qins
->registers
.src2_reg
< 16)
2294 register_dep_mask
|= (1 << qins
->registers
.src2_reg
);
2296 int mask
= qins
->alu
.mask
;
2298 for (int c
= 0; c
< 4; ++c
)
2299 if (mask
& (0x3 << (2 * c
)))
2300 components
[c
] = true;
2302 /* ..but if the writeout is too late, we have to break up anyway... for some reason */
2304 if (qins
->unit
== UNIT_VLUT
)
2305 written_late
= true;
2308 /* Advance instruction pointer */
2309 qins
= mir_next_op(qins
);
2313 /* 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) */
2314 if (register_dep_mask
& written_mask
) {
2315 DBG("ERRATA WORKAROUND: Breakup for writeout dependency masks %X vs %X (common %X)\n", register_dep_mask
, written_mask
, register_dep_mask
& written_mask
);
2322 /* If even a single component is not written, break it up (conservative check). */
2323 bool breakup
= false;
2325 for (int c
= 0; c
< 4; ++c
)
2332 /* Otherwise, we're free to proceed */
2336 if (ains
->unit
== ALU_ENAB_BRANCH
) {
2337 bundle
.body_size
[bundle
.body_words_count
] = sizeof(midgard_branch_extended
);
2338 memcpy(&bundle
.body_words
[bundle
.body_words_count
++], &ains
->branch_extended
, sizeof(midgard_branch_extended
));
2339 bytes_emitted
+= sizeof(midgard_branch_extended
);
2341 bundle
.body_size
[bundle
.body_words_count
] = sizeof(ains
->br_compact
);
2342 memcpy(&bundle
.body_words
[bundle
.body_words_count
++], &ains
->br_compact
, sizeof(ains
->br_compact
));
2343 bytes_emitted
+= sizeof(ains
->br_compact
);
2346 memcpy(&bundle
.register_words
[bundle
.register_words_count
++], &ains
->registers
, sizeof(ains
->registers
));
2347 bytes_emitted
+= sizeof(midgard_reg_info
);
2349 bundle
.body_size
[bundle
.body_words_count
] = sizeof(midgard_scalar_alu
);
2350 bundle
.body_words_count
++;
2351 bytes_emitted
+= sizeof(midgard_scalar_alu
);
2354 /* Defer marking until after writing to allow for break */
2355 control
|= ains
->unit
;
2356 last_unit
= ains
->unit
;
2357 ++instructions_emitted
;
2361 /* Bubble up the number of instructions for skipping */
2362 instructions_consumed
= index
- 1;
2366 /* Pad ALU op to nearest word */
2368 if (bytes_emitted
& 15) {
2369 padding
= 16 - (bytes_emitted
& 15);
2370 bytes_emitted
+= padding
;
2373 /* Constants must always be quadwords */
2374 if (bundle
.has_embedded_constants
)
2375 bytes_emitted
+= 16;
2377 /* Size ALU instruction for tag */
2378 bundle
.tag
= (TAG_ALU_4
) + (bytes_emitted
/ 16) - 1;
2379 bundle
.padding
= padding
;
2380 bundle
.control
= bundle
.tag
| control
;
2385 case TAG_LOAD_STORE_4
: {
2386 /* Load store instructions have two words at once. If
2387 * we only have one queued up, we need to NOP pad.
2388 * Otherwise, we store both in succession to save space
2389 * and cycles -- letting them go in parallel -- skip
2390 * the next. The usefulness of this optimisation is
2391 * greatly dependent on the quality of the instruction
2395 midgard_instruction
*next_op
= mir_next_op(ins
);
2397 if ((struct list_head
*) next_op
!= &block
->instructions
&& next_op
->type
== TAG_LOAD_STORE_4
) {
2398 /* As the two operate concurrently, make sure
2399 * they are not dependent */
2401 if (can_run_concurrent_ssa(ins
, next_op
) || true) {
2402 /* Skip ahead, since it's redundant with the pair */
2403 instructions_consumed
= 1 + (instructions_emitted
++);
2411 /* Texture ops default to single-op-per-bundle scheduling */
2415 /* Copy the instructions into the bundle */
2416 bundle
.instruction_count
= instructions_emitted
+ 1;
2420 midgard_instruction
*uins
= ins
;
2421 for (int i
= 0; used_idx
< bundle
.instruction_count
; ++i
) {
2422 bundle
.instructions
[used_idx
++] = *uins
;
2423 uins
= mir_next_op(uins
);
2426 *skip
= (instructions_consumed
== -1) ? instructions_emitted
: instructions_consumed
;
2432 quadword_size(int tag
)
2447 case TAG_LOAD_STORE_4
:
2459 /* Schedule a single block by iterating its instruction to create bundles.
2460 * While we go, tally about the bundle sizes to compute the block size. */
2463 schedule_block(compiler_context
*ctx
, midgard_block
*block
)
2465 util_dynarray_init(&block
->bundles
, NULL
);
2467 block
->quadword_count
= 0;
2469 mir_foreach_instr_in_block(block
, ins
) {
2471 midgard_bundle bundle
= schedule_bundle(ctx
, block
, ins
, &skip
);
2472 util_dynarray_append(&block
->bundles
, midgard_bundle
, bundle
);
2474 if (bundle
.has_blend_constant
) {
2475 /* TODO: Multiblock? */
2476 int quadwords_within_block
= block
->quadword_count
+ quadword_size(bundle
.tag
) - 1;
2477 ctx
->blend_constant_offset
= quadwords_within_block
* 0x10;
2481 ins
= mir_next_op(ins
);
2483 block
->quadword_count
+= quadword_size(bundle
.tag
);
2486 block
->is_scheduled
= true;
2490 schedule_program(compiler_context
*ctx
)
2492 allocate_registers(ctx
);
2494 mir_foreach_block(ctx
, block
) {
2495 schedule_block(ctx
, block
);
2499 /* After everything is scheduled, emit whole bundles at a time */
2502 emit_binary_bundle(compiler_context
*ctx
, midgard_bundle
*bundle
, struct util_dynarray
*emission
, int next_tag
)
2504 int lookahead
= next_tag
<< 4;
2506 switch (bundle
->tag
) {
2511 /* Actually emit each component */
2512 util_dynarray_append(emission
, uint32_t, bundle
->control
| lookahead
);
2514 for (int i
= 0; i
< bundle
->register_words_count
; ++i
)
2515 util_dynarray_append(emission
, uint16_t, bundle
->register_words
[i
]);
2517 /* Emit body words based on the instructions bundled */
2518 for (int i
= 0; i
< bundle
->instruction_count
; ++i
) {
2519 midgard_instruction
*ins
= &bundle
->instructions
[i
];
2521 if (ins
->unit
& UNITS_ANY_VECTOR
) {
2522 memcpy(util_dynarray_grow(emission
, sizeof(midgard_vector_alu
)), &ins
->alu
, sizeof(midgard_vector_alu
));
2523 } else if (ins
->compact_branch
) {
2524 /* Dummy move, XXX DRY */
2525 if ((i
== 0) && ins
->writeout
) {
2526 midgard_instruction ins
= v_fmov(0, blank_alu_src
, SSA_FIXED_REGISTER(0));
2527 memcpy(util_dynarray_grow(emission
, sizeof(midgard_vector_alu
)), &ins
.alu
, sizeof(midgard_vector_alu
));
2530 if (ins
->unit
== ALU_ENAB_BR_COMPACT
) {
2531 memcpy(util_dynarray_grow(emission
, sizeof(ins
->br_compact
)), &ins
->br_compact
, sizeof(ins
->br_compact
));
2533 memcpy(util_dynarray_grow(emission
, sizeof(ins
->branch_extended
)), &ins
->branch_extended
, sizeof(ins
->branch_extended
));
2537 midgard_scalar_alu scalarised
= vector_to_scalar_alu(ins
->alu
, ins
);
2538 memcpy(util_dynarray_grow(emission
, sizeof(scalarised
)), &scalarised
, sizeof(scalarised
));
2542 /* Emit padding (all zero) */
2543 memset(util_dynarray_grow(emission
, bundle
->padding
), 0, bundle
->padding
);
2545 /* Tack on constants */
2547 if (bundle
->has_embedded_constants
) {
2548 util_dynarray_append(emission
, float, bundle
->constants
[0]);
2549 util_dynarray_append(emission
, float, bundle
->constants
[1]);
2550 util_dynarray_append(emission
, float, bundle
->constants
[2]);
2551 util_dynarray_append(emission
, float, bundle
->constants
[3]);
2557 case TAG_LOAD_STORE_4
: {
2558 /* One or two composing instructions */
2560 uint64_t current64
, next64
= LDST_NOP
;
2562 memcpy(¤t64
, &bundle
->instructions
[0].load_store
, sizeof(current64
));
2564 if (bundle
->instruction_count
== 2)
2565 memcpy(&next64
, &bundle
->instructions
[1].load_store
, sizeof(next64
));
2567 midgard_load_store instruction
= {
2568 .type
= bundle
->tag
,
2569 .next_type
= next_tag
,
2574 util_dynarray_append(emission
, midgard_load_store
, instruction
);
2579 case TAG_TEXTURE_4
: {
2580 /* Texture instructions are easy, since there is no
2581 * pipelining nor VLIW to worry about. We may need to set the .last flag */
2583 midgard_instruction
*ins
= &bundle
->instructions
[0];
2585 ins
->texture
.type
= TAG_TEXTURE_4
;
2586 ins
->texture
.next_type
= next_tag
;
2588 ctx
->texture_op_count
--;
2590 if (!ctx
->texture_op_count
) {
2591 ins
->texture
.cont
= 0;
2592 ins
->texture
.last
= 1;
2595 util_dynarray_append(emission
, midgard_texture_word
, ins
->texture
);
2600 DBG("Unknown midgard instruction type\n");
2607 /* ALU instructions can inline or embed constants, which decreases register
2608 * pressure and saves space. */
2610 #define CONDITIONAL_ATTACH(src) { \
2611 void *entry = _mesa_hash_table_u64_search(ctx->ssa_constants, alu->ssa_args.src + 1); \
2614 attach_constants(ctx, alu, entry, alu->ssa_args.src + 1); \
2615 alu->ssa_args.src = SSA_FIXED_REGISTER(REGISTER_CONSTANT); \
2620 inline_alu_constants(compiler_context
*ctx
)
2622 mir_foreach_instr(ctx
, alu
) {
2623 /* Other instructions cannot inline constants */
2624 if (alu
->type
!= TAG_ALU_4
) continue;
2626 /* If there is already a constant here, we can do nothing */
2627 if (alu
->has_constants
) continue;
2629 CONDITIONAL_ATTACH(src0
);
2631 if (!alu
->has_constants
) {
2632 CONDITIONAL_ATTACH(src1
)
2633 } else if (!alu
->inline_constant
) {
2634 /* Corner case: _two_ vec4 constants, for instance with a
2635 * csel. For this case, we can only use a constant
2636 * register for one, we'll have to emit a move for the
2637 * other. Note, if both arguments are constants, then
2638 * necessarily neither argument depends on the value of
2639 * any particular register. As the destination register
2640 * will be wiped, that means we can spill the constant
2641 * to the destination register.
2644 void *entry
= _mesa_hash_table_u64_search(ctx
->ssa_constants
, alu
->ssa_args
.src1
+ 1);
2645 unsigned scratch
= alu
->ssa_args
.dest
;
2648 midgard_instruction ins
= v_fmov(SSA_FIXED_REGISTER(REGISTER_CONSTANT
), blank_alu_src
, scratch
);
2649 attach_constants(ctx
, &ins
, entry
, alu
->ssa_args
.src1
+ 1);
2651 /* Force a break XXX Defer r31 writes */
2652 ins
.unit
= UNIT_VLUT
;
2654 /* Set the source */
2655 alu
->ssa_args
.src1
= scratch
;
2657 /* Inject us -before- the last instruction which set r31 */
2658 mir_insert_instruction_before(mir_prev_op(alu
), ins
);
2664 /* Midgard supports two types of constants, embedded constants (128-bit) and
2665 * inline constants (16-bit). Sometimes, especially with scalar ops, embedded
2666 * constants can be demoted to inline constants, for space savings and
2667 * sometimes a performance boost */
2670 embedded_to_inline_constant(compiler_context
*ctx
)
2672 mir_foreach_instr(ctx
, ins
) {
2673 if (!ins
->has_constants
) continue;
2675 if (ins
->ssa_args
.inline_constant
) continue;
2677 /* Blend constants must not be inlined by definition */
2678 if (ins
->has_blend_constant
) continue;
2680 /* src1 cannot be an inline constant due to encoding
2681 * restrictions. So, if possible we try to flip the arguments
2684 int op
= ins
->alu
.op
;
2686 if (ins
->ssa_args
.src0
== SSA_FIXED_REGISTER(REGISTER_CONSTANT
)) {
2687 /* Flip based on op. Fallthrough intentional */
2690 /* These ops require an operational change to flip their arguments TODO */
2691 case midgard_alu_op_flt
:
2692 case midgard_alu_op_fle
:
2693 case midgard_alu_op_ilt
:
2694 case midgard_alu_op_ile
:
2695 case midgard_alu_op_fcsel
:
2696 case midgard_alu_op_icsel
:
2697 case midgard_alu_op_isub
:
2698 DBG("Missed non-commutative flip (%s)\n", alu_opcode_names
[op
]);
2701 /* These ops are commutative and Just Flip */
2702 case midgard_alu_op_fne
:
2703 case midgard_alu_op_fadd
:
2704 case midgard_alu_op_fmul
:
2705 case midgard_alu_op_fmin
:
2706 case midgard_alu_op_fmax
:
2707 case midgard_alu_op_iadd
:
2708 case midgard_alu_op_imul
:
2709 case midgard_alu_op_feq
:
2710 case midgard_alu_op_ieq
:
2711 case midgard_alu_op_ine
:
2712 case midgard_alu_op_iand
:
2713 case midgard_alu_op_ior
:
2714 case midgard_alu_op_ixor
:
2715 /* Flip the SSA numbers */
2716 ins
->ssa_args
.src0
= ins
->ssa_args
.src1
;
2717 ins
->ssa_args
.src1
= SSA_FIXED_REGISTER(REGISTER_CONSTANT
);
2719 /* And flip the modifiers */
2723 src_temp
= ins
->alu
.src2
;
2724 ins
->alu
.src2
= ins
->alu
.src1
;
2725 ins
->alu
.src1
= src_temp
;
2732 if (ins
->ssa_args
.src1
== SSA_FIXED_REGISTER(REGISTER_CONSTANT
)) {
2733 /* Extract the source information */
2735 midgard_vector_alu_src
*src
;
2736 int q
= ins
->alu
.src2
;
2737 midgard_vector_alu_src
*m
= (midgard_vector_alu_src
*) &q
;
2740 /* Component is from the swizzle, e.g. r26.w -> w component. TODO: What if x is masked out? */
2741 int component
= src
->swizzle
& 3;
2743 /* Scale constant appropriately, if we can legally */
2744 uint16_t scaled_constant
= 0;
2746 /* XXX: Check legality */
2747 if (midgard_is_integer_op(op
)) {
2748 /* TODO: Inline integer */
2751 unsigned int *iconstants
= (unsigned int *) ins
->constants
;
2752 scaled_constant
= (uint16_t) iconstants
[component
];
2754 /* Constant overflow after resize */
2755 if (scaled_constant
!= iconstants
[component
])
2758 scaled_constant
= _mesa_float_to_half((float) ins
->constants
[component
]);
2761 /* We don't know how to handle these with a constant */
2763 if (src
->abs
|| src
->negate
|| src
->half
|| src
->rep_low
|| src
->rep_high
) {
2764 DBG("Bailing inline constant...\n");
2768 /* Make sure that the constant is not itself a
2769 * vector by checking if all accessed values
2770 * (by the swizzle) are the same. */
2772 uint32_t *cons
= (uint32_t *) ins
->constants
;
2773 uint32_t value
= cons
[component
];
2775 bool is_vector
= false;
2776 unsigned mask
= effective_writemask(&ins
->alu
);
2778 for (int c
= 1; c
< 4; ++c
) {
2779 /* We only care if this component is actually used */
2780 if (!(mask
& (1 << c
)))
2783 uint32_t test
= cons
[(src
->swizzle
>> (2 * c
)) & 3];
2785 if (test
!= value
) {
2794 /* Get rid of the embedded constant */
2795 ins
->has_constants
= false;
2796 ins
->ssa_args
.src1
= SSA_UNUSED_0
;
2797 ins
->ssa_args
.inline_constant
= true;
2798 ins
->inline_constant
= scaled_constant
;
2803 /* Map normal SSA sources to other SSA sources / fixed registers (like
2807 map_ssa_to_alias(compiler_context
*ctx
, int *ref
)
2809 unsigned int alias
= (uintptr_t) _mesa_hash_table_u64_search(ctx
->ssa_to_alias
, *ref
+ 1);
2812 /* Remove entry in leftovers to avoid a redunant fmov */
2814 struct set_entry
*leftover
= _mesa_set_search(ctx
->leftover_ssa_to_alias
, ((void *) (uintptr_t) (*ref
+ 1)));
2817 _mesa_set_remove(ctx
->leftover_ssa_to_alias
, leftover
);
2819 /* Assign the alias map */
2825 #define AS_SRC(to, u) \
2826 int q##to = ins->alu.src2; \
2827 midgard_vector_alu_src *to = (midgard_vector_alu_src *) &q##to;
2829 /* Removing unused moves is necessary to clean up the texture pipeline results.
2831 * To do so, we find moves in the MIR. We check if their destination is live later. If it's not, the move is redundant. */
2834 midgard_eliminate_orphan_moves(compiler_context
*ctx
, midgard_block
*block
)
2836 mir_foreach_instr_in_block_safe(block
, ins
) {
2837 if (ins
->type
!= TAG_ALU_4
) continue;
2839 if (ins
->alu
.op
!= midgard_alu_op_fmov
) continue;
2841 if (ins
->ssa_args
.dest
>= SSA_FIXED_MINIMUM
) continue;
2843 if (midgard_is_pinned(ctx
, ins
->ssa_args
.dest
)) continue;
2845 if (is_live_after(ctx
, block
, ins
, ins
->ssa_args
.dest
)) continue;
2847 mir_remove_instruction(ins
);
2851 /* The following passes reorder MIR instructions to enable better scheduling */
2854 midgard_pair_load_store(compiler_context
*ctx
, midgard_block
*block
)
2856 mir_foreach_instr_in_block_safe(block
, ins
) {
2857 if (ins
->type
!= TAG_LOAD_STORE_4
) continue;
2859 /* We've found a load/store op. Check if next is also load/store. */
2860 midgard_instruction
*next_op
= mir_next_op(ins
);
2861 if (&next_op
->link
!= &block
->instructions
) {
2862 if (next_op
->type
== TAG_LOAD_STORE_4
) {
2863 /* If so, we're done since we're a pair */
2864 ins
= mir_next_op(ins
);
2868 /* Maximum search distance to pair, to avoid register pressure disasters */
2869 int search_distance
= 8;
2871 /* Otherwise, we have an orphaned load/store -- search for another load */
2872 mir_foreach_instr_in_block_from(block
, c
, mir_next_op(ins
)) {
2873 /* Terminate search if necessary */
2874 if (!(search_distance
--)) break;
2876 if (c
->type
!= TAG_LOAD_STORE_4
) continue;
2878 if (OP_IS_STORE(c
->load_store
.op
)) continue;
2880 /* We found one! Move it up to pair and remove it from the old location */
2882 mir_insert_instruction_before(ins
, *c
);
2883 mir_remove_instruction(c
);
2891 /* Emit varying stores late */
2894 midgard_emit_store(compiler_context
*ctx
, midgard_block
*block
) {
2895 /* Iterate in reverse to get the final write, rather than the first */
2897 mir_foreach_instr_in_block_safe_rev(block
, ins
) {
2898 /* Check if what we just wrote needs a store */
2899 int idx
= ins
->ssa_args
.dest
;
2900 uintptr_t varying
= ((uintptr_t) _mesa_hash_table_u64_search(ctx
->ssa_varyings
, idx
+ 1));
2902 if (!varying
) continue;
2906 /* We need to store to the appropriate varying, so emit the
2909 /* TODO: Integrate with special purpose RA (and scheduler?) */
2910 bool high_varying_register
= false;
2912 midgard_instruction mov
= v_fmov(idx
, blank_alu_src
, SSA_FIXED_REGISTER(REGISTER_VARYING_BASE
+ high_varying_register
));
2914 midgard_instruction st
= m_store_vary_32(SSA_FIXED_REGISTER(high_varying_register
), varying
);
2915 st
.load_store
.unknown
= 0x1E9E; /* XXX: What is this? */
2917 mir_insert_instruction_before(mir_next_op(ins
), st
);
2918 mir_insert_instruction_before(mir_next_op(ins
), mov
);
2920 /* We no longer need to store this varying */
2921 _mesa_hash_table_u64_remove(ctx
->ssa_varyings
, idx
+ 1);
2925 /* If there are leftovers after the below pass, emit actual fmov
2926 * instructions for the slow-but-correct path */
2929 emit_leftover_move(compiler_context
*ctx
)
2931 set_foreach(ctx
->leftover_ssa_to_alias
, leftover
) {
2932 int base
= ((uintptr_t) leftover
->key
) - 1;
2935 map_ssa_to_alias(ctx
, &mapped
);
2936 EMIT(fmov
, mapped
, blank_alu_src
, base
);
2941 actualise_ssa_to_alias(compiler_context
*ctx
)
2943 mir_foreach_instr(ctx
, ins
) {
2944 map_ssa_to_alias(ctx
, &ins
->ssa_args
.src0
);
2945 map_ssa_to_alias(ctx
, &ins
->ssa_args
.src1
);
2948 emit_leftover_move(ctx
);
2951 /* Vertex shaders do not write gl_Position as is; instead, they write a
2952 * transformed screen space position as a varying. See section 12.5 "Coordinate
2953 * Transformation" of the ES 3.2 full specification for details.
2955 * This transformation occurs early on, as NIR and prior to optimisation, in
2956 * order to take advantage of NIR optimisation passes of the transform itself.
2960 write_transformed_position(nir_builder
*b
, nir_src input_point_src
, int uniform_no
)
2962 nir_ssa_def
*input_point
= nir_ssa_for_src(b
, input_point_src
, 4);
2964 /* Get viewport from the uniforms */
2965 nir_intrinsic_instr
*load
;
2966 load
= nir_intrinsic_instr_create(b
->shader
, nir_intrinsic_load_uniform
);
2967 load
->num_components
= 4;
2968 load
->src
[0] = nir_src_for_ssa(nir_imm_int(b
, uniform_no
));
2969 nir_ssa_dest_init(&load
->instr
, &load
->dest
, 4, 32, NULL
);
2970 nir_builder_instr_insert(b
, &load
->instr
);
2972 /* Formatted as <width, height, centerx, centery> */
2973 nir_ssa_def
*viewport_vec4
= &load
->dest
.ssa
;
2974 nir_ssa_def
*viewport_width_2
= nir_channel(b
, viewport_vec4
, 0);
2975 nir_ssa_def
*viewport_height_2
= nir_channel(b
, viewport_vec4
, 1);
2976 nir_ssa_def
*viewport_offset
= nir_channels(b
, viewport_vec4
, 0x8 | 0x4);
2978 /* XXX: From uniforms? */
2979 nir_ssa_def
*depth_near
= nir_imm_float(b
, 0.0);
2980 nir_ssa_def
*depth_far
= nir_imm_float(b
, 1.0);
2982 /* World space to normalised device coordinates */
2984 nir_ssa_def
*w_recip
= nir_frcp(b
, nir_channel(b
, input_point
, 3));
2985 nir_ssa_def
*ndc_point
= nir_fmul(b
, nir_channels(b
, input_point
, 0x7), w_recip
);
2987 /* Normalised device coordinates to screen space */
2989 nir_ssa_def
*viewport_multiplier
= nir_vec2(b
, viewport_width_2
, viewport_height_2
);
2990 nir_ssa_def
*viewport_xy
= nir_fadd(b
, nir_fmul(b
, nir_channels(b
, ndc_point
, 0x3), viewport_multiplier
), viewport_offset
);
2992 nir_ssa_def
*depth_multiplier
= nir_fmul(b
, nir_fsub(b
, depth_far
, depth_near
), nir_imm_float(b
, 0.5f
));
2993 nir_ssa_def
*depth_offset
= nir_fmul(b
, nir_fadd(b
, depth_far
, depth_near
), nir_imm_float(b
, 0.5f
));
2994 nir_ssa_def
*screen_depth
= nir_fadd(b
, nir_fmul(b
, nir_channel(b
, ndc_point
, 2), depth_multiplier
), depth_offset
);
2996 /* gl_Position will be written out in screenspace xyz, with w set to
2997 * the reciprocal we computed earlier. The transformed w component is
2998 * then used for perspective-correct varying interpolation. The
2999 * transformed w component must preserve its original sign; this is
3000 * used in depth clipping computations */
3002 nir_ssa_def
*screen_space
= nir_vec4(b
,
3003 nir_channel(b
, viewport_xy
, 0),
3004 nir_channel(b
, viewport_xy
, 1),
3008 /* Finally, write out the transformed values to the varying */
3010 nir_intrinsic_instr
*store
;
3011 store
= nir_intrinsic_instr_create(b
->shader
, nir_intrinsic_store_output
);
3012 store
->num_components
= 4;
3013 nir_intrinsic_set_base(store
, 0);
3014 nir_intrinsic_set_write_mask(store
, 0xf);
3015 store
->src
[0].ssa
= screen_space
;
3016 store
->src
[0].is_ssa
= true;
3017 store
->src
[1] = nir_src_for_ssa(nir_imm_int(b
, 0));
3018 nir_builder_instr_insert(b
, &store
->instr
);
3022 transform_position_writes(nir_shader
*shader
)
3024 nir_foreach_function(func
, shader
) {
3025 nir_foreach_block(block
, func
->impl
) {
3026 nir_foreach_instr_safe(instr
, block
) {
3027 if (instr
->type
!= nir_instr_type_intrinsic
) continue;
3029 nir_intrinsic_instr
*intr
= nir_instr_as_intrinsic(instr
);
3030 nir_variable
*out
= NULL
;
3032 switch (intr
->intrinsic
) {
3033 case nir_intrinsic_store_output
:
3034 /* already had i/o lowered.. lookup the matching output var: */
3035 nir_foreach_variable(var
, &shader
->outputs
) {
3036 int drvloc
= var
->data
.driver_location
;
3038 if (nir_intrinsic_base(intr
) == drvloc
) {
3052 if (out
->data
.mode
!= nir_var_shader_out
)
3055 if (out
->data
.location
!= VARYING_SLOT_POS
)
3059 nir_builder_init(&b
, func
->impl
);
3060 b
.cursor
= nir_before_instr(instr
);
3062 write_transformed_position(&b
, intr
->src
[0], UNIFORM_VIEWPORT
);
3063 nir_instr_remove(instr
);
3070 emit_fragment_epilogue(compiler_context
*ctx
)
3072 /* Special case: writing out constants requires us to include the move
3073 * explicitly now, so shove it into r0 */
3075 void *constant_value
= _mesa_hash_table_u64_search(ctx
->ssa_constants
, ctx
->fragment_output
+ 1);
3077 if (constant_value
) {
3078 midgard_instruction ins
= v_fmov(SSA_FIXED_REGISTER(REGISTER_CONSTANT
), blank_alu_src
, SSA_FIXED_REGISTER(0));
3079 attach_constants(ctx
, &ins
, constant_value
, ctx
->fragment_output
+ 1);
3080 emit_mir_instruction(ctx
, ins
);
3083 /* Perform the actual fragment writeout. We have two writeout/branch
3084 * instructions, forming a loop until writeout is successful as per the
3085 * docs. TODO: gl_FragDepth */
3087 EMIT(alu_br_compact_cond
, midgard_jmp_writeout_op_writeout
, TAG_ALU_4
, 0, midgard_condition_always
);
3088 EMIT(alu_br_compact_cond
, midgard_jmp_writeout_op_writeout
, TAG_ALU_4
, -1, midgard_condition_always
);
3091 /* For the blend epilogue, we need to convert the blended fragment vec4 (stored
3092 * in r0) to a RGBA8888 value by scaling and type converting. We then output it
3093 * with the int8 analogue to the fragment epilogue */
3096 emit_blend_epilogue(compiler_context
*ctx
)
3098 /* vmul.fmul.none.fulllow hr48, r0, #255 */
3100 midgard_instruction scale
= {
3103 .inline_constant
= _mesa_float_to_half(255.0),
3105 .src0
= SSA_FIXED_REGISTER(0),
3106 .src1
= SSA_UNUSED_0
,
3107 .dest
= SSA_FIXED_REGISTER(24),
3108 .inline_constant
= true
3111 .op
= midgard_alu_op_fmul
,
3112 .reg_mode
= midgard_reg_mode_full
,
3113 .dest_override
= midgard_dest_override_lower
,
3115 .src1
= vector_alu_srco_unsigned(blank_alu_src
),
3116 .src2
= vector_alu_srco_unsigned(blank_alu_src
),
3120 emit_mir_instruction(ctx
, scale
);
3122 /* vadd.f2u8.pos.low hr0, hr48, #0 */
3124 midgard_vector_alu_src alu_src
= blank_alu_src
;
3125 alu_src
.half
= true;
3127 midgard_instruction f2u8
= {
3130 .src0
= SSA_FIXED_REGISTER(24),
3131 .src1
= SSA_UNUSED_0
,
3132 .dest
= SSA_FIXED_REGISTER(0),
3133 .inline_constant
= true
3136 .op
= midgard_alu_op_f2u8
,
3137 .reg_mode
= midgard_reg_mode_half
,
3138 .dest_override
= midgard_dest_override_lower
,
3139 .outmod
= midgard_outmod_pos
,
3141 .src1
= vector_alu_srco_unsigned(alu_src
),
3142 .src2
= vector_alu_srco_unsigned(blank_alu_src
),
3146 emit_mir_instruction(ctx
, f2u8
);
3148 /* vmul.imov.quarter r0, r0, r0 */
3150 midgard_instruction imov_8
= {
3153 .src0
= SSA_UNUSED_1
,
3154 .src1
= SSA_FIXED_REGISTER(0),
3155 .dest
= SSA_FIXED_REGISTER(0),
3158 .op
= midgard_alu_op_imov
,
3159 .reg_mode
= midgard_reg_mode_quarter
,
3160 .dest_override
= midgard_dest_override_none
,
3162 .src1
= vector_alu_srco_unsigned(blank_alu_src
),
3163 .src2
= vector_alu_srco_unsigned(blank_alu_src
),
3167 /* Emit branch epilogue with the 8-bit move as the source */
3169 emit_mir_instruction(ctx
, imov_8
);
3170 EMIT(alu_br_compact_cond
, midgard_jmp_writeout_op_writeout
, TAG_ALU_4
, 0, midgard_condition_always
);
3172 emit_mir_instruction(ctx
, imov_8
);
3173 EMIT(alu_br_compact_cond
, midgard_jmp_writeout_op_writeout
, TAG_ALU_4
, -1, midgard_condition_always
);
3176 static midgard_block
*
3177 emit_block(compiler_context
*ctx
, nir_block
*block
)
3179 midgard_block
*this_block
= malloc(sizeof(midgard_block
));
3180 list_addtail(&this_block
->link
, &ctx
->blocks
);
3182 this_block
->is_scheduled
= false;
3185 ctx
->texture_index
[0] = -1;
3186 ctx
->texture_index
[1] = -1;
3188 /* Set up current block */
3189 list_inithead(&this_block
->instructions
);
3190 ctx
->current_block
= this_block
;
3192 nir_foreach_instr(instr
, block
) {
3193 emit_instr(ctx
, instr
);
3194 ++ctx
->instruction_count
;
3197 inline_alu_constants(ctx
);
3198 embedded_to_inline_constant(ctx
);
3200 /* Perform heavylifting for aliasing */
3201 actualise_ssa_to_alias(ctx
);
3203 midgard_emit_store(ctx
, this_block
);
3204 midgard_eliminate_orphan_moves(ctx
, this_block
);
3205 midgard_pair_load_store(ctx
, this_block
);
3207 /* Append fragment shader epilogue (value writeout) */
3208 if (ctx
->stage
== MESA_SHADER_FRAGMENT
) {
3209 if (block
== nir_impl_last_block(ctx
->func
->impl
)) {
3211 emit_blend_epilogue(ctx
);
3213 emit_fragment_epilogue(ctx
);
3217 /* Fallthrough save */
3218 this_block
->next_fallthrough
= ctx
->previous_source_block
;
3220 if (block
== nir_start_block(ctx
->func
->impl
))
3221 ctx
->initial_block
= this_block
;
3223 if (block
== nir_impl_last_block(ctx
->func
->impl
))
3224 ctx
->final_block
= this_block
;
3226 /* Allow the next control flow to access us retroactively, for
3228 ctx
->current_block
= this_block
;
3230 /* Document the fallthrough chain */
3231 ctx
->previous_source_block
= this_block
;
3236 static midgard_block
*emit_cf_list(struct compiler_context
*ctx
, struct exec_list
*list
);
3239 emit_if(struct compiler_context
*ctx
, nir_if
*nif
)
3241 /* Conditional branches expect the condition in r31.w; emit a move for
3242 * that in the _previous_ block (which is the current block). */
3243 emit_condition(ctx
, &nif
->condition
, true);
3245 /* Speculatively emit the branch, but we can't fill it in until later */
3246 EMIT(branch
, true, true);
3247 midgard_instruction
*then_branch
= mir_last_in_block(ctx
->current_block
);
3249 /* Emit the two subblocks */
3250 midgard_block
*then_block
= emit_cf_list(ctx
, &nif
->then_list
);
3252 /* Emit a jump from the end of the then block to the end of the else */
3253 EMIT(branch
, false, false);
3254 midgard_instruction
*then_exit
= mir_last_in_block(ctx
->current_block
);
3256 /* Emit second block, and check if it's empty */
3258 int else_idx
= ctx
->block_count
;
3259 int count_in
= ctx
->instruction_count
;
3260 midgard_block
*else_block
= emit_cf_list(ctx
, &nif
->else_list
);
3261 int after_else_idx
= ctx
->block_count
;
3263 /* Now that we have the subblocks emitted, fix up the branches */
3268 if (ctx
->instruction_count
== count_in
) {
3269 /* The else block is empty, so don't emit an exit jump */
3270 mir_remove_instruction(then_exit
);
3271 then_branch
->branch
.target_block
= after_else_idx
;
3273 then_branch
->branch
.target_block
= else_idx
;
3274 then_exit
->branch
.target_block
= after_else_idx
;
3279 emit_loop(struct compiler_context
*ctx
, nir_loop
*nloop
)
3281 /* Remember where we are */
3282 midgard_block
*start_block
= ctx
->current_block
;
3284 /* Allocate a loop number for this. TODO: Nested loops. Instead of a
3285 * single current_loop variable, maybe we need a stack */
3287 int loop_idx
= ++ctx
->current_loop
;
3289 /* Get index from before the body so we can loop back later */
3290 int start_idx
= ctx
->block_count
;
3292 /* Emit the body itself */
3293 emit_cf_list(ctx
, &nloop
->body
);
3295 /* Branch back to loop back */
3296 struct midgard_instruction br_back
= v_branch(false, false);
3297 br_back
.branch
.target_block
= start_idx
;
3298 emit_mir_instruction(ctx
, br_back
);
3300 /* Find the index of the block about to follow us (note: we don't add
3301 * one; blocks are 0-indexed so we get a fencepost problem) */
3302 int break_block_idx
= ctx
->block_count
;
3304 /* Fix up the break statements we emitted to point to the right place,
3305 * now that we can allocate a block number for them */
3307 list_for_each_entry_from(struct midgard_block
, block
, start_block
, &ctx
->blocks
, link
) {
3308 if (midgard_debug
& MIDGARD_DBG_SHADERS
)
3309 print_mir_block(block
);
3310 mir_foreach_instr_in_block(block
, ins
) {
3311 if (ins
->type
!= TAG_ALU_4
) continue;
3312 if (!ins
->compact_branch
) continue;
3313 if (ins
->prepacked_branch
) continue;
3315 /* We found a branch -- check the type to see if we need to do anything */
3316 if (ins
->branch
.target_type
!= TARGET_BREAK
) continue;
3318 /* It's a break! Check if it's our break */
3319 if (ins
->branch
.target_break
!= loop_idx
) continue;
3321 /* Okay, cool, we're breaking out of this loop.
3322 * Rewrite from a break to a goto */
3324 ins
->branch
.target_type
= TARGET_GOTO
;
3325 ins
->branch
.target_block
= break_block_idx
;
3330 static midgard_block
*
3331 emit_cf_list(struct compiler_context
*ctx
, struct exec_list
*list
)
3333 midgard_block
*start_block
= NULL
;
3335 foreach_list_typed(nir_cf_node
, node
, node
, list
) {
3336 switch (node
->type
) {
3337 case nir_cf_node_block
: {
3338 midgard_block
*block
= emit_block(ctx
, nir_cf_node_as_block(node
));
3341 start_block
= block
;
3346 case nir_cf_node_if
:
3347 emit_if(ctx
, nir_cf_node_as_if(node
));
3350 case nir_cf_node_loop
:
3351 emit_loop(ctx
, nir_cf_node_as_loop(node
));
3354 case nir_cf_node_function
:
3363 /* Due to lookahead, we need to report the first tag executed in the command
3364 * stream and in branch targets. An initial block might be empty, so iterate
3365 * until we find one that 'works' */
3368 midgard_get_first_tag_from_block(compiler_context
*ctx
, unsigned block_idx
)
3370 midgard_block
*initial_block
= mir_get_block(ctx
, block_idx
);
3372 unsigned first_tag
= 0;
3375 midgard_bundle
*initial_bundle
= util_dynarray_element(&initial_block
->bundles
, midgard_bundle
, 0);
3377 if (initial_bundle
) {
3378 first_tag
= initial_bundle
->tag
;
3382 /* Initial block is empty, try the next block */
3383 initial_block
= list_first_entry(&(initial_block
->link
), midgard_block
, link
);
3384 } while(initial_block
!= NULL
);
3391 midgard_compile_shader_nir(nir_shader
*nir
, midgard_program
*program
, bool is_blend
)
3393 struct util_dynarray
*compiled
= &program
->compiled
;
3395 midgard_debug
= debug_get_option_midgard_debug();
3397 compiler_context ictx
= {
3399 .stage
= nir
->info
.stage
,
3401 .is_blend
= is_blend
,
3402 .blend_constant_offset
= -1,
3404 .alpha_ref
= program
->alpha_ref
3407 compiler_context
*ctx
= &ictx
;
3409 /* TODO: Decide this at runtime */
3410 ctx
->uniform_cutoff
= 8;
3412 switch (ctx
->stage
) {
3413 case MESA_SHADER_VERTEX
:
3414 ctx
->special_uniforms
= 1;
3418 ctx
->special_uniforms
= 0;
3422 /* Append epilogue uniforms if necessary. The cmdstream depends on
3423 * these being at the -end-; see assign_var_locations. */
3425 if (ctx
->stage
== MESA_SHADER_VERTEX
) {
3426 nir_variable_create(nir
, nir_var_uniform
, glsl_vec4_type(), "viewport");
3429 /* Assign var locations early, so the epilogue can use them if necessary */
3431 nir_assign_var_locations(&nir
->outputs
, &nir
->num_outputs
, glsl_type_size
);
3432 nir_assign_var_locations(&nir
->inputs
, &nir
->num_inputs
, glsl_type_size
);
3433 nir_assign_var_locations(&nir
->uniforms
, &nir
->num_uniforms
, glsl_type_size
);
3435 /* Initialize at a global (not block) level hash tables */
3437 ctx
->ssa_constants
= _mesa_hash_table_u64_create(NULL
);
3438 ctx
->ssa_varyings
= _mesa_hash_table_u64_create(NULL
);
3439 ctx
->ssa_to_alias
= _mesa_hash_table_u64_create(NULL
);
3440 ctx
->ssa_to_register
= _mesa_hash_table_u64_create(NULL
);
3441 ctx
->hash_to_temp
= _mesa_hash_table_u64_create(NULL
);
3442 ctx
->leftover_ssa_to_alias
= _mesa_set_create(NULL
, _mesa_hash_pointer
, _mesa_key_pointer_equal
);
3444 /* Assign actual uniform location, skipping over samplers */
3446 ctx
->uniform_nir_to_mdg
= _mesa_hash_table_u64_create(NULL
);
3448 nir_foreach_variable(var
, &nir
->uniforms
) {
3449 if (glsl_get_base_type(var
->type
) == GLSL_TYPE_SAMPLER
) continue;
3451 unsigned length
= glsl_get_aoa_size(var
->type
);
3454 length
= glsl_get_length(var
->type
);
3458 length
= glsl_get_matrix_columns(var
->type
);
3461 for (int col
= 0; col
< length
; ++col
) {
3462 int id
= ctx
->uniform_count
++;
3463 _mesa_hash_table_u64_insert(ctx
->uniform_nir_to_mdg
, var
->data
.driver_location
+ col
+ 1, (void *) ((uintptr_t) (id
+ 1)));
3467 /* Record the varying mapping for the command stream's bookkeeping */
3469 struct exec_list
*varyings
=
3470 ctx
->stage
== MESA_SHADER_VERTEX
? &nir
->outputs
: &nir
->inputs
;
3472 nir_foreach_variable(var
, varyings
) {
3473 unsigned loc
= var
->data
.driver_location
;
3474 program
->varyings
[loc
] = var
->data
.location
;
3477 /* Lower vars -- not I/O -- before epilogue */
3479 NIR_PASS_V(nir
, nir_lower_var_copies
);
3480 NIR_PASS_V(nir
, nir_lower_vars_to_ssa
);
3481 NIR_PASS_V(nir
, nir_split_var_copies
);
3482 NIR_PASS_V(nir
, nir_lower_var_copies
);
3483 NIR_PASS_V(nir
, nir_lower_global_vars_to_local
);
3484 NIR_PASS_V(nir
, nir_lower_var_copies
);
3485 NIR_PASS_V(nir
, nir_lower_vars_to_ssa
);
3486 NIR_PASS_V(nir
, nir_lower_io
, nir_var_all
, glsl_type_size
, 0);
3488 /* Append vertex epilogue before optimisation, so the epilogue itself
3491 if (ctx
->stage
== MESA_SHADER_VERTEX
)
3492 transform_position_writes(nir
);
3494 /* Optimisation passes */
3498 if (midgard_debug
& MIDGARD_DBG_SHADERS
) {
3499 nir_print_shader(nir
, stdout
);
3502 /* Assign counts, now that we're sure (post-optimisation) */
3503 program
->uniform_count
= nir
->num_uniforms
;
3505 program
->attribute_count
= (ctx
->stage
== MESA_SHADER_VERTEX
) ? nir
->num_inputs
: 0;
3506 program
->varying_count
= (ctx
->stage
== MESA_SHADER_VERTEX
) ? nir
->num_outputs
: ((ctx
->stage
== MESA_SHADER_FRAGMENT
) ? nir
->num_inputs
: 0);
3509 nir_foreach_function(func
, nir
) {
3513 list_inithead(&ctx
->blocks
);
3514 ctx
->block_count
= 0;
3517 emit_cf_list(ctx
, &func
->impl
->body
);
3518 emit_block(ctx
, func
->impl
->end_block
);
3520 break; /* TODO: Multi-function shaders */
3523 util_dynarray_init(compiled
, NULL
);
3526 schedule_program(ctx
);
3528 /* Now that all the bundles are scheduled and we can calculate block
3529 * sizes, emit actual branch instructions rather than placeholders */
3531 int br_block_idx
= 0;
3533 mir_foreach_block(ctx
, block
) {
3534 util_dynarray_foreach(&block
->bundles
, midgard_bundle
, bundle
) {
3535 for (int c
= 0; c
< bundle
->instruction_count
; ++c
) {
3536 midgard_instruction
*ins
= &bundle
->instructions
[c
];
3538 if (!midgard_is_branch_unit(ins
->unit
)) continue;
3540 if (ins
->prepacked_branch
) continue;
3542 /* Parse some basic branch info */
3543 bool is_compact
= ins
->unit
== ALU_ENAB_BR_COMPACT
;
3544 bool is_conditional
= ins
->branch
.conditional
;
3545 bool is_inverted
= ins
->branch
.invert_conditional
;
3546 bool is_discard
= ins
->branch
.target_type
== TARGET_DISCARD
;
3548 /* Determine the block we're jumping to */
3549 int target_number
= ins
->branch
.target_block
;
3551 /* Report the destination tag. Discards don't need this */
3552 int dest_tag
= is_discard
? 0 : midgard_get_first_tag_from_block(ctx
, target_number
);
3554 /* 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) */
3555 int quadword_offset
= 0;
3558 /* Jump to the end of the shader. We
3559 * need to include not only the
3560 * following blocks, but also the
3561 * contents of our current block (since
3562 * discard can come in the middle of
3565 midgard_block
*blk
= mir_get_block(ctx
, br_block_idx
+ 1);
3567 for (midgard_bundle
*bun
= bundle
+ 1; bun
< (midgard_bundle
*)((char*) block
->bundles
.data
+ block
->bundles
.size
); ++bun
) {
3568 quadword_offset
+= quadword_size(bun
->tag
);
3571 mir_foreach_block_from(ctx
, blk
, b
) {
3572 quadword_offset
+= b
->quadword_count
;
3575 } else if (target_number
> br_block_idx
) {
3578 for (int idx
= br_block_idx
+ 1; idx
< target_number
; ++idx
) {
3579 midgard_block
*blk
= mir_get_block(ctx
, idx
);
3582 quadword_offset
+= blk
->quadword_count
;
3585 /* Jump backwards */
3587 for (int idx
= br_block_idx
; idx
>= target_number
; --idx
) {
3588 midgard_block
*blk
= mir_get_block(ctx
, idx
);
3591 quadword_offset
-= blk
->quadword_count
;
3595 /* Unconditional extended branches (far jumps)
3596 * have issues, so we always use a conditional
3597 * branch, setting the condition to always for
3598 * unconditional. For compact unconditional
3599 * branches, cond isn't used so it doesn't
3600 * matter what we pick. */
3602 midgard_condition cond
=
3603 !is_conditional
? midgard_condition_always
:
3604 is_inverted
? midgard_condition_false
:
3605 midgard_condition_true
;
3607 midgard_jmp_writeout_op op
=
3608 is_discard
? midgard_jmp_writeout_op_discard
:
3609 (is_compact
&& !is_conditional
) ? midgard_jmp_writeout_op_branch_uncond
:
3610 midgard_jmp_writeout_op_branch_cond
;
3613 midgard_branch_extended branch
=
3614 midgard_create_branch_extended(
3619 memcpy(&ins
->branch_extended
, &branch
, sizeof(branch
));
3620 } else if (is_conditional
|| is_discard
) {
3621 midgard_branch_cond branch
= {
3623 .dest_tag
= dest_tag
,
3624 .offset
= quadword_offset
,
3628 assert(branch
.offset
== quadword_offset
);
3630 memcpy(&ins
->br_compact
, &branch
, sizeof(branch
));
3632 assert(op
== midgard_jmp_writeout_op_branch_uncond
);
3634 midgard_branch_uncond branch
= {
3636 .dest_tag
= dest_tag
,
3637 .offset
= quadword_offset
,
3641 assert(branch
.offset
== quadword_offset
);
3643 memcpy(&ins
->br_compact
, &branch
, sizeof(branch
));
3651 /* Emit flat binary from the instruction arrays. Iterate each block in
3652 * sequence. Save instruction boundaries such that lookahead tags can
3653 * be assigned easily */
3655 /* Cache _all_ bundles in source order for lookahead across failed branches */
3657 int bundle_count
= 0;
3658 mir_foreach_block(ctx
, block
) {
3659 bundle_count
+= block
->bundles
.size
/ sizeof(midgard_bundle
);
3661 midgard_bundle
**source_order_bundles
= malloc(sizeof(midgard_bundle
*) * bundle_count
);
3663 mir_foreach_block(ctx
, block
) {
3664 util_dynarray_foreach(&block
->bundles
, midgard_bundle
, bundle
) {
3665 source_order_bundles
[bundle_idx
++] = bundle
;
3669 int current_bundle
= 0;
3671 mir_foreach_block(ctx
, block
) {
3672 util_dynarray_foreach(&block
->bundles
, midgard_bundle
, bundle
) {
3675 if (current_bundle
+ 1 < bundle_count
) {
3676 uint8_t next
= source_order_bundles
[current_bundle
+ 1]->tag
;
3678 if (!(current_bundle
+ 2 < bundle_count
) && IS_ALU(next
)) {
3685 emit_binary_bundle(ctx
, bundle
, compiled
, lookahead
);
3689 /* TODO: Free deeper */
3690 //util_dynarray_fini(&block->instructions);
3693 free(source_order_bundles
);
3695 /* Report the very first tag executed */
3696 program
->first_tag
= midgard_get_first_tag_from_block(ctx
, 0);
3698 /* Deal with off-by-one related to the fencepost problem */
3699 program
->work_register_count
= ctx
->work_registers
+ 1;
3701 program
->can_discard
= ctx
->can_discard
;
3702 program
->uniform_cutoff
= ctx
->uniform_cutoff
;
3704 program
->blend_patch_offset
= ctx
->blend_constant_offset
;
3706 if (midgard_debug
& MIDGARD_DBG_SHADERS
)
3707 disassemble_midgard(program
->compiled
.data
, program
->compiled
.size
);