2 * Copyright (C) 2018-2019 Alyssa Rosenzweig <alyssa@rosenzweig.io>
4 * Permission is hereby granted, free of charge, to any person obtaining a
5 * copy of this software and associated documentation files (the "Software"),
6 * to deal in the Software without restriction, including without limitation
7 * the rights to use, copy, modify, merge, publish, distribute, sublicense,
8 * and/or sell copies of the Software, and to permit persons to whom the
9 * Software is furnished to do so, subject to the following conditions:
11 * The above copyright notice and this permission notice (including the next
12 * paragraph) shall be included in all copies or substantial portions of the
15 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
16 * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
17 * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL
18 * THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
19 * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
20 * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
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/u_math.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"
48 #include "midgard_ops.h"
52 #include "disassemble.h"
54 static const struct debug_named_value debug_options
[] = {
55 {"msgs", MIDGARD_DBG_MSGS
, "Print debug messages"},
56 {"shaders", MIDGARD_DBG_SHADERS
, "Dump shaders in NIR and MIR"},
57 {"shaderdb", MIDGARD_DBG_SHADERDB
, "Prints shader-db statistics"},
61 DEBUG_GET_ONCE_FLAGS_OPTION(midgard_debug
, "MIDGARD_MESA_DEBUG", debug_options
, 0)
63 unsigned SHADER_DB_COUNT
= 0;
65 int midgard_debug
= 0;
67 #define DBG(fmt, ...) \
68 do { if (midgard_debug & MIDGARD_DBG_MSGS) \
69 fprintf(stderr, "%s:%d: "fmt, \
70 __FUNCTION__, __LINE__, ##__VA_ARGS__); } while (0)
73 midgard_is_branch_unit(unsigned unit
)
75 return (unit
== ALU_ENAB_BRANCH
) || (unit
== ALU_ENAB_BR_COMPACT
);
79 midgard_block_add_successor(midgard_block
*block
, midgard_block
*successor
)
81 block
->successors
[block
->nr_successors
++] = successor
;
82 assert(block
->nr_successors
<= ARRAY_SIZE(block
->successors
));
85 /* Helpers to generate midgard_instruction's using macro magic, since every
86 * driver seems to do it that way */
88 #define EMIT(op, ...) emit_mir_instruction(ctx, v_##op(__VA_ARGS__));
90 #define M_LOAD_STORE(name, rname, uname) \
91 static midgard_instruction m_##name(unsigned ssa, unsigned address) { \
92 midgard_instruction i = { \
93 .type = TAG_LOAD_STORE_4, \
101 .op = midgard_op_##name, \
102 .swizzle = SWIZZLE_XYZW, \
110 #define M_LOAD(name) M_LOAD_STORE(name, dest, src0)
111 #define M_STORE(name) M_LOAD_STORE(name, src0, dest)
113 /* Inputs a NIR ALU source, with modifiers attached if necessary, and outputs
114 * the corresponding Midgard source */
116 static midgard_vector_alu_src
117 vector_alu_modifiers(nir_alu_src
*src
, bool is_int
, unsigned broadcast_count
,
118 bool half
, bool sext
)
120 if (!src
) return blank_alu_src
;
122 /* Figure out how many components there are so we can adjust the
123 * swizzle. Specifically we want to broadcast the last channel so
124 * things like ball2/3 work
127 if (broadcast_count
) {
128 uint8_t last_component
= src
->swizzle
[broadcast_count
- 1];
130 for (unsigned c
= broadcast_count
; c
< NIR_MAX_VEC_COMPONENTS
; ++c
) {
131 src
->swizzle
[c
] = last_component
;
135 midgard_vector_alu_src alu_src
= {
139 .swizzle
= SWIZZLE_FROM_ARRAY(src
->swizzle
)
143 alu_src
.mod
= midgard_int_normal
;
145 /* Sign/zero-extend if needed */
149 midgard_int_sign_extend
150 : midgard_int_zero_extend
;
153 /* These should have been lowered away */
154 assert(!(src
->abs
|| src
->negate
));
156 alu_src
.mod
= (src
->abs
<< 0) | (src
->negate
<< 1);
162 /* load/store instructions have both 32-bit and 16-bit variants, depending on
163 * whether we are using vectors composed of highp or mediump. At the moment, we
164 * don't support half-floats -- this requires changes in other parts of the
165 * compiler -- therefore the 16-bit versions are commented out. */
167 //M_LOAD(ld_attr_16);
169 //M_LOAD(ld_vary_16);
171 //M_LOAD(ld_uniform_16);
172 M_LOAD(ld_uniform_32
);
173 M_LOAD(ld_color_buffer_8
);
174 //M_STORE(st_vary_16);
176 M_STORE(st_cubemap_coords
);
178 static midgard_instruction
179 v_alu_br_compact_cond(midgard_jmp_writeout_op op
, unsigned tag
, signed offset
, unsigned cond
)
181 midgard_branch_cond branch
= {
189 memcpy(&compact
, &branch
, sizeof(branch
));
191 midgard_instruction ins
= {
193 .unit
= ALU_ENAB_BR_COMPACT
,
194 .prepacked_branch
= true,
195 .compact_branch
= true,
196 .br_compact
= compact
199 if (op
== midgard_jmp_writeout_op_writeout
)
205 static midgard_instruction
206 v_branch(bool conditional
, bool invert
)
208 midgard_instruction ins
= {
210 .unit
= ALU_ENAB_BRANCH
,
211 .compact_branch
= true,
213 .conditional
= conditional
,
214 .invert_conditional
= invert
221 static midgard_branch_extended
222 midgard_create_branch_extended( midgard_condition cond
,
223 midgard_jmp_writeout_op op
,
225 signed quadword_offset
)
227 /* For unclear reasons, the condition code is repeated 8 times */
228 uint16_t duplicated_cond
=
238 midgard_branch_extended branch
= {
240 .dest_tag
= dest_tag
,
241 .offset
= quadword_offset
,
242 .cond
= duplicated_cond
249 attach_constants(compiler_context
*ctx
, midgard_instruction
*ins
, void *constants
, int name
)
251 ins
->has_constants
= true;
252 memcpy(&ins
->constants
, constants
, 16);
256 glsl_type_size(const struct glsl_type
*type
, bool bindless
)
258 return glsl_count_attribute_slots(type
, false);
261 /* Lower fdot2 to a vector multiplication followed by channel addition */
263 midgard_nir_lower_fdot2_body(nir_builder
*b
, nir_alu_instr
*alu
)
265 if (alu
->op
!= nir_op_fdot2
)
268 b
->cursor
= nir_before_instr(&alu
->instr
);
270 nir_ssa_def
*src0
= nir_ssa_for_alu_src(b
, alu
, 0);
271 nir_ssa_def
*src1
= nir_ssa_for_alu_src(b
, alu
, 1);
273 nir_ssa_def
*product
= nir_fmul(b
, src0
, src1
);
275 nir_ssa_def
*sum
= nir_fadd(b
,
276 nir_channel(b
, product
, 0),
277 nir_channel(b
, product
, 1));
279 /* Replace the fdot2 with this sum */
280 nir_ssa_def_rewrite_uses(&alu
->dest
.dest
.ssa
, nir_src_for_ssa(sum
));
284 midgard_nir_sysval_for_intrinsic(nir_intrinsic_instr
*instr
)
286 switch (instr
->intrinsic
) {
287 case nir_intrinsic_load_viewport_scale
:
288 return PAN_SYSVAL_VIEWPORT_SCALE
;
289 case nir_intrinsic_load_viewport_offset
:
290 return PAN_SYSVAL_VIEWPORT_OFFSET
;
297 nir_dest_index(compiler_context
*ctx
, nir_dest
*dst
)
300 return dst
->ssa
.index
;
302 assert(!dst
->reg
.indirect
);
303 return ctx
->func
->impl
->ssa_alloc
+ dst
->reg
.reg
->index
;
307 static int sysval_for_instr(compiler_context
*ctx
, nir_instr
*instr
,
310 nir_intrinsic_instr
*intr
;
311 nir_dest
*dst
= NULL
;
315 switch (instr
->type
) {
316 case nir_instr_type_intrinsic
:
317 intr
= nir_instr_as_intrinsic(instr
);
318 sysval
= midgard_nir_sysval_for_intrinsic(intr
);
321 case nir_instr_type_tex
:
322 tex
= nir_instr_as_tex(instr
);
323 if (tex
->op
!= nir_texop_txs
)
326 sysval
= PAN_SYSVAL(TEXTURE_SIZE
,
327 PAN_TXS_SYSVAL_ID(tex
->texture_index
,
328 nir_tex_instr_dest_size(tex
) -
329 (tex
->is_array
? 1 : 0),
338 *dest
= nir_dest_index(ctx
, dst
);
344 midgard_nir_assign_sysval_body(compiler_context
*ctx
, nir_instr
*instr
)
348 sysval
= sysval_for_instr(ctx
, instr
, NULL
);
352 /* We have a sysval load; check if it's already been assigned */
354 if (_mesa_hash_table_u64_search(ctx
->sysval_to_id
, sysval
))
357 /* It hasn't -- so assign it now! */
359 unsigned id
= ctx
->sysval_count
++;
360 _mesa_hash_table_u64_insert(ctx
->sysval_to_id
, sysval
, (void *) ((uintptr_t) id
+ 1));
361 ctx
->sysvals
[id
] = sysval
;
365 midgard_nir_assign_sysvals(compiler_context
*ctx
, nir_shader
*shader
)
367 ctx
->sysval_count
= 0;
369 nir_foreach_function(function
, shader
) {
370 if (!function
->impl
) continue;
372 nir_foreach_block(block
, function
->impl
) {
373 nir_foreach_instr_safe(instr
, block
) {
374 midgard_nir_assign_sysval_body(ctx
, instr
);
381 midgard_nir_lower_fdot2(nir_shader
*shader
)
383 bool progress
= false;
385 nir_foreach_function(function
, shader
) {
386 if (!function
->impl
) continue;
389 nir_builder
*b
= &_b
;
390 nir_builder_init(b
, function
->impl
);
392 nir_foreach_block(block
, function
->impl
) {
393 nir_foreach_instr_safe(instr
, block
) {
394 if (instr
->type
!= nir_instr_type_alu
) continue;
396 nir_alu_instr
*alu
= nir_instr_as_alu(instr
);
397 midgard_nir_lower_fdot2_body(b
, alu
);
403 nir_metadata_preserve(function
->impl
, nir_metadata_block_index
| nir_metadata_dominance
);
410 /* Flushes undefined values to zero */
413 optimise_nir(nir_shader
*nir
)
416 unsigned lower_flrp
=
417 (nir
->options
->lower_flrp16
? 16 : 0) |
418 (nir
->options
->lower_flrp32
? 32 : 0) |
419 (nir
->options
->lower_flrp64
? 64 : 0);
421 NIR_PASS(progress
, nir
, nir_lower_regs_to_ssa
);
422 NIR_PASS(progress
, nir
, midgard_nir_lower_fdot2
);
423 NIR_PASS(progress
, nir
, nir_lower_idiv
);
425 nir_lower_tex_options lower_tex_1st_pass_options
= {
430 nir_lower_tex_options lower_tex_2nd_pass_options
= {
431 .lower_txs_lod
= true,
434 NIR_PASS(progress
, nir
, nir_lower_tex
, &lower_tex_1st_pass_options
);
435 NIR_PASS(progress
, nir
, nir_lower_tex
, &lower_tex_2nd_pass_options
);
440 NIR_PASS(progress
, nir
, nir_lower_var_copies
);
441 NIR_PASS(progress
, nir
, nir_lower_vars_to_ssa
);
443 NIR_PASS(progress
, nir
, nir_copy_prop
);
444 NIR_PASS(progress
, nir
, nir_opt_dce
);
445 NIR_PASS(progress
, nir
, nir_opt_dead_cf
);
446 NIR_PASS(progress
, nir
, nir_opt_cse
);
447 NIR_PASS(progress
, nir
, nir_opt_peephole_select
, 64, false, true);
448 NIR_PASS(progress
, nir
, nir_opt_algebraic
);
449 NIR_PASS(progress
, nir
, nir_opt_constant_folding
);
451 if (lower_flrp
!= 0) {
452 bool lower_flrp_progress
= false;
453 NIR_PASS(lower_flrp_progress
,
457 false /* always_precise */,
458 nir
->options
->lower_ffma
);
459 if (lower_flrp_progress
) {
460 NIR_PASS(progress
, nir
,
461 nir_opt_constant_folding
);
465 /* Nothing should rematerialize any flrps, so we only
466 * need to do this lowering once.
471 NIR_PASS(progress
, nir
, nir_opt_undef
);
472 NIR_PASS(progress
, nir
, nir_undef_to_zero
);
474 NIR_PASS(progress
, nir
, nir_opt_loop_unroll
,
477 nir_var_function_temp
);
479 NIR_PASS(progress
, nir
, nir_opt_vectorize
);
482 /* Must be run at the end to prevent creation of fsin/fcos ops */
483 NIR_PASS(progress
, nir
, midgard_nir_scale_trig
);
488 NIR_PASS(progress
, nir
, nir_opt_dce
);
489 NIR_PASS(progress
, nir
, nir_opt_algebraic
);
490 NIR_PASS(progress
, nir
, nir_opt_constant_folding
);
491 NIR_PASS(progress
, nir
, nir_copy_prop
);
494 NIR_PASS(progress
, nir
, nir_opt_algebraic_late
);
496 /* We implement booleans as 32-bit 0/~0 */
497 NIR_PASS(progress
, nir
, nir_lower_bool_to_int32
);
499 /* Now that booleans are lowered, we can run out late opts */
500 NIR_PASS(progress
, nir
, midgard_nir_lower_algebraic_late
);
502 /* Lower mods for float ops only. Integer ops don't support modifiers
503 * (saturate doesn't make sense on integers, neg/abs require dedicated
506 NIR_PASS(progress
, nir
, nir_lower_to_source_mods
, nir_lower_float_source_mods
);
507 NIR_PASS(progress
, nir
, nir_copy_prop
);
508 NIR_PASS(progress
, nir
, nir_opt_dce
);
510 /* Take us out of SSA */
511 NIR_PASS(progress
, nir
, nir_lower_locals_to_regs
);
512 NIR_PASS(progress
, nir
, nir_convert_from_ssa
, true);
514 /* We are a vector architecture; write combine where possible */
515 NIR_PASS(progress
, nir
, nir_move_vec_src_uses_to_dest
);
516 NIR_PASS(progress
, nir
, nir_lower_vec_to_movs
);
518 NIR_PASS(progress
, nir
, nir_opt_dce
);
521 /* Front-half of aliasing the SSA slots, merely by inserting the flag in the
522 * appropriate hash table. Intentional off-by-one to avoid confusing NULL with
523 * r0. See the comments in compiler_context */
526 alias_ssa(compiler_context
*ctx
, int dest
, int src
)
528 _mesa_hash_table_u64_insert(ctx
->ssa_to_alias
, dest
+ 1, (void *) ((uintptr_t) src
+ 1));
529 _mesa_set_add(ctx
->leftover_ssa_to_alias
, (void *) (uintptr_t) (dest
+ 1));
532 /* ...or undo it, after which the original index will be used (dummy move should be emitted alongside this) */
535 unalias_ssa(compiler_context
*ctx
, int dest
)
537 _mesa_hash_table_u64_remove(ctx
->ssa_to_alias
, dest
+ 1);
538 /* TODO: Remove from leftover or no? */
541 /* Do not actually emit a load; instead, cache the constant for inlining */
544 emit_load_const(compiler_context
*ctx
, nir_load_const_instr
*instr
)
546 nir_ssa_def def
= instr
->def
;
548 float *v
= rzalloc_array(NULL
, float, 4);
549 nir_const_load_to_arr(v
, instr
, f32
);
550 _mesa_hash_table_u64_insert(ctx
->ssa_constants
, def
.index
+ 1, v
);
554 nir_src_index(compiler_context
*ctx
, nir_src
*src
)
557 return src
->ssa
->index
;
559 assert(!src
->reg
.indirect
);
560 return ctx
->func
->impl
->ssa_alloc
+ src
->reg
.reg
->index
;
565 nir_alu_src_index(compiler_context
*ctx
, nir_alu_src
*src
)
567 return nir_src_index(ctx
, &src
->src
);
571 nir_is_non_scalar_swizzle(nir_alu_src
*src
, unsigned nr_components
)
573 unsigned comp
= src
->swizzle
[0];
575 for (unsigned c
= 1; c
< nr_components
; ++c
) {
576 if (src
->swizzle
[c
] != comp
)
583 /* Midgard puts scalar conditionals in r31.w; move an arbitrary source (the
584 * output of a conditional test) into that register */
587 emit_condition(compiler_context
*ctx
, nir_src
*src
, bool for_branch
, unsigned component
)
589 int condition
= nir_src_index(ctx
, src
);
591 /* Source to swizzle the desired component into w */
593 const midgard_vector_alu_src alu_src
= {
594 .swizzle
= SWIZZLE(component
, component
, component
, component
),
597 /* There is no boolean move instruction. Instead, we simulate a move by
598 * ANDing the condition with itself to get it into r31.w */
600 midgard_instruction ins
= {
603 /* We need to set the conditional as close as possible */
604 .precede_break
= true,
605 .unit
= for_branch
? UNIT_SMUL
: UNIT_SADD
,
606 .mask
= 1 << COMPONENT_W
,
611 .dest
= SSA_FIXED_REGISTER(31),
615 .op
= midgard_alu_op_iand
,
616 .outmod
= midgard_outmod_int_wrap
,
617 .reg_mode
= midgard_reg_mode_32
,
618 .dest_override
= midgard_dest_override_none
,
619 .src1
= vector_alu_srco_unsigned(alu_src
),
620 .src2
= vector_alu_srco_unsigned(alu_src
)
624 emit_mir_instruction(ctx
, ins
);
627 /* Or, for mixed conditions (with csel_v), here's a vector version using all of
631 emit_condition_mixed(compiler_context
*ctx
, nir_alu_src
*src
, unsigned nr_comp
)
633 int condition
= nir_src_index(ctx
, &src
->src
);
635 /* Source to swizzle the desired component into w */
637 const midgard_vector_alu_src alu_src
= {
638 .swizzle
= SWIZZLE_FROM_ARRAY(src
->swizzle
),
641 /* There is no boolean move instruction. Instead, we simulate a move by
642 * ANDing the condition with itself to get it into r31.w */
644 midgard_instruction ins
= {
646 .precede_break
= true,
647 .mask
= mask_of(nr_comp
),
651 .dest
= SSA_FIXED_REGISTER(31),
654 .op
= midgard_alu_op_iand
,
655 .outmod
= midgard_outmod_int_wrap
,
656 .reg_mode
= midgard_reg_mode_32
,
657 .dest_override
= midgard_dest_override_none
,
658 .src1
= vector_alu_srco_unsigned(alu_src
),
659 .src2
= vector_alu_srco_unsigned(alu_src
)
663 emit_mir_instruction(ctx
, ins
);
668 /* Likewise, indirect offsets are put in r27.w. TODO: Allow componentwise
669 * pinning to eliminate this move in all known cases */
672 emit_indirect_offset(compiler_context
*ctx
, nir_src
*src
)
674 int offset
= nir_src_index(ctx
, src
);
676 midgard_instruction ins
= {
678 .mask
= 1 << COMPONENT_W
,
680 .src0
= SSA_UNUSED_1
,
682 .dest
= SSA_FIXED_REGISTER(REGISTER_OFFSET
),
685 .op
= midgard_alu_op_imov
,
686 .outmod
= midgard_outmod_int_wrap
,
687 .reg_mode
= midgard_reg_mode_32
,
688 .dest_override
= midgard_dest_override_none
,
689 .src1
= vector_alu_srco_unsigned(zero_alu_src
),
690 .src2
= vector_alu_srco_unsigned(blank_alu_src_xxxx
)
694 emit_mir_instruction(ctx
, ins
);
697 #define ALU_CASE(nir, _op) \
699 op = midgard_alu_op_##_op; \
700 assert(src_bitsize == dst_bitsize); \
703 #define ALU_CASE_BCAST(nir, _op, count) \
705 op = midgard_alu_op_##_op; \
706 broadcast_swizzle = count; \
707 assert(src_bitsize == dst_bitsize); \
710 nir_is_fzero_constant(nir_src src
)
712 if (!nir_src_is_const(src
))
715 for (unsigned c
= 0; c
< nir_src_num_components(src
); ++c
) {
716 if (nir_src_comp_as_float(src
, c
) != 0.0)
723 /* Analyze the sizes of the inputs to determine which reg mode. Ops needed
724 * special treatment override this anyway. */
726 static midgard_reg_mode
727 reg_mode_for_nir(nir_alu_instr
*instr
)
729 unsigned src_bitsize
= nir_src_bit_size(instr
->src
[0].src
);
731 switch (src_bitsize
) {
733 return midgard_reg_mode_8
;
735 return midgard_reg_mode_16
;
737 return midgard_reg_mode_32
;
739 return midgard_reg_mode_64
;
741 unreachable("Invalid bit size");
746 emit_alu(compiler_context
*ctx
, nir_alu_instr
*instr
)
748 bool is_ssa
= instr
->dest
.dest
.is_ssa
;
750 unsigned dest
= nir_dest_index(ctx
, &instr
->dest
.dest
);
751 unsigned nr_components
= nir_dest_num_components(instr
->dest
.dest
);
752 unsigned nr_inputs
= nir_op_infos
[instr
->op
].num_inputs
;
754 /* Most Midgard ALU ops have a 1:1 correspondance to NIR ops; these are
755 * supported. A few do not and are commented for now. Also, there are a
756 * number of NIR ops which Midgard does not support and need to be
757 * lowered, also TODO. This switch block emits the opcode and calling
758 * convention of the Midgard instruction; actual packing is done in
763 /* Number of components valid to check for the instruction (the rest
764 * will be forced to the last), or 0 to use as-is. Relevant as
765 * ball-type instructions have a channel count in NIR but are all vec4
768 unsigned broadcast_swizzle
= 0;
770 /* What register mode should we operate in? */
771 midgard_reg_mode reg_mode
=
772 reg_mode_for_nir(instr
);
774 /* Do we need a destination override? Used for inline
777 midgard_dest_override dest_override
=
778 midgard_dest_override_none
;
780 /* Should we use a smaller respective source and sign-extend? */
782 bool half_1
= false, sext_1
= false;
783 bool half_2
= false, sext_2
= false;
785 unsigned src_bitsize
= nir_src_bit_size(instr
->src
[0].src
);
786 unsigned dst_bitsize
= nir_dest_bit_size(instr
->dest
.dest
);
789 ALU_CASE(fadd
, fadd
);
790 ALU_CASE(fmul
, fmul
);
791 ALU_CASE(fmin
, fmin
);
792 ALU_CASE(fmax
, fmax
);
793 ALU_CASE(imin
, imin
);
794 ALU_CASE(imax
, imax
);
795 ALU_CASE(umin
, umin
);
796 ALU_CASE(umax
, umax
);
797 ALU_CASE(ffloor
, ffloor
);
798 ALU_CASE(fround_even
, froundeven
);
799 ALU_CASE(ftrunc
, ftrunc
);
800 ALU_CASE(fceil
, fceil
);
801 ALU_CASE(fdot3
, fdot3
);
802 ALU_CASE(fdot4
, fdot4
);
803 ALU_CASE(iadd
, iadd
);
804 ALU_CASE(isub
, isub
);
805 ALU_CASE(imul
, imul
);
807 /* Zero shoved as second-arg */
808 ALU_CASE(iabs
, iabsdiff
);
812 ALU_CASE(feq32
, feq
);
813 ALU_CASE(fne32
, fne
);
814 ALU_CASE(flt32
, flt
);
815 ALU_CASE(ieq32
, ieq
);
816 ALU_CASE(ine32
, ine
);
817 ALU_CASE(ilt32
, ilt
);
818 ALU_CASE(ult32
, ult
);
820 /* We don't have a native b2f32 instruction. Instead, like many
821 * GPUs, we exploit booleans as 0/~0 for false/true, and
822 * correspondingly AND
823 * by 1.0 to do the type conversion. For the moment, prime us
826 * iand [whatever], #0
828 * At the end of emit_alu (as MIR), we'll fix-up the constant
831 ALU_CASE(b2f32
, iand
);
832 ALU_CASE(b2i32
, iand
);
834 /* Likewise, we don't have a dedicated f2b32 instruction, but
835 * we can do a "not equal to 0.0" test. */
837 ALU_CASE(f2b32
, fne
);
838 ALU_CASE(i2b32
, ine
);
840 ALU_CASE(frcp
, frcp
);
841 ALU_CASE(frsq
, frsqrt
);
842 ALU_CASE(fsqrt
, fsqrt
);
843 ALU_CASE(fexp2
, fexp2
);
844 ALU_CASE(flog2
, flog2
);
846 ALU_CASE(f2i32
, f2i_rtz
);
847 ALU_CASE(f2u32
, f2u_rtz
);
848 ALU_CASE(i2f32
, i2f_rtz
);
849 ALU_CASE(u2f32
, u2f_rtz
);
851 ALU_CASE(f2i16
, f2i_rtz
);
852 ALU_CASE(f2u16
, f2u_rtz
);
853 ALU_CASE(i2f16
, i2f_rtz
);
854 ALU_CASE(u2f16
, u2f_rtz
);
856 ALU_CASE(fsin
, fsin
);
857 ALU_CASE(fcos
, fcos
);
859 /* Second op implicit #0 */
860 ALU_CASE(inot
, inor
);
861 ALU_CASE(iand
, iand
);
863 ALU_CASE(ixor
, ixor
);
864 ALU_CASE(ishl
, ishl
);
865 ALU_CASE(ishr
, iasr
);
866 ALU_CASE(ushr
, ilsr
);
868 ALU_CASE_BCAST(b32all_fequal2
, fball_eq
, 2);
869 ALU_CASE_BCAST(b32all_fequal3
, fball_eq
, 3);
870 ALU_CASE(b32all_fequal4
, fball_eq
);
872 ALU_CASE_BCAST(b32any_fnequal2
, fbany_neq
, 2);
873 ALU_CASE_BCAST(b32any_fnequal3
, fbany_neq
, 3);
874 ALU_CASE(b32any_fnequal4
, fbany_neq
);
876 ALU_CASE_BCAST(b32all_iequal2
, iball_eq
, 2);
877 ALU_CASE_BCAST(b32all_iequal3
, iball_eq
, 3);
878 ALU_CASE(b32all_iequal4
, iball_eq
);
880 ALU_CASE_BCAST(b32any_inequal2
, ibany_neq
, 2);
881 ALU_CASE_BCAST(b32any_inequal3
, ibany_neq
, 3);
882 ALU_CASE(b32any_inequal4
, ibany_neq
);
884 /* Source mods will be shoved in later */
885 ALU_CASE(fabs
, fmov
);
886 ALU_CASE(fneg
, fmov
);
887 ALU_CASE(fsat
, fmov
);
889 /* For size conversion, we use a move. Ideally though we would squash
890 * these ops together; maybe that has to happen after in NIR as part of
891 * propagation...? An earlier algebraic pass ensured we step down by
892 * only / exactly one size. If stepping down, we use a dest override to
893 * reduce the size; if stepping up, we use a larger-sized move with a
894 * half source and a sign/zero-extension modifier */
899 /* If we end up upscale, we'll need a sign-extend on the
900 * operand (the second argument) */
906 op
= midgard_alu_op_imov
;
908 if (dst_bitsize
== (src_bitsize
* 2)) {
912 /* Use a greater register mode */
914 } else if (src_bitsize
== (dst_bitsize
* 2)) {
915 /* Converting down */
916 dest_override
= midgard_dest_override_lower
;
923 assert(src_bitsize
== 32);
925 op
= midgard_alu_op_fmov
;
926 dest_override
= midgard_dest_override_lower
;
931 assert(src_bitsize
== 16);
933 op
= midgard_alu_op_fmov
;
940 /* For greater-or-equal, we lower to less-or-equal and flip the
948 instr
->op
== nir_op_fge
? midgard_alu_op_fle
:
949 instr
->op
== nir_op_fge32
? midgard_alu_op_fle
:
950 instr
->op
== nir_op_ige32
? midgard_alu_op_ile
:
951 instr
->op
== nir_op_uge32
? midgard_alu_op_ule
:
954 /* Swap via temporary */
955 nir_alu_src temp
= instr
->src
[1];
956 instr
->src
[1] = instr
->src
[0];
957 instr
->src
[0] = temp
;
962 case nir_op_b32csel
: {
963 /* Midgard features both fcsel and icsel, depending on
964 * the type of the arguments/output. However, as long
965 * as we're careful we can _always_ use icsel and
966 * _never_ need fcsel, since the latter does additional
967 * floating-point-specific processing whereas the
968 * former just moves bits on the wire. It's not obvious
969 * why these are separate opcodes, save for the ability
970 * to do things like sat/pos/abs/neg for free */
972 bool mixed
= nir_is_non_scalar_swizzle(&instr
->src
[0], nr_components
);
973 op
= mixed
? midgard_alu_op_icsel_v
: midgard_alu_op_icsel
;
975 /* csel works as a two-arg in Midgard, since the condition is hardcoded in r31.w */
978 /* Emit the condition into r31 */
981 emit_condition_mixed(ctx
, &instr
->src
[0], nr_components
);
983 emit_condition(ctx
, &instr
->src
[0].src
, false, instr
->src
[0].swizzle
[0]);
985 /* The condition is the first argument; move the other
986 * arguments up one to be a binary instruction for
989 memmove(instr
->src
, instr
->src
+ 1, 2 * sizeof(nir_alu_src
));
994 DBG("Unhandled ALU op %s\n", nir_op_infos
[instr
->op
].name
);
999 /* Midgard can perform certain modifiers on output of an ALU op */
1002 if (midgard_is_integer_out_op(op
)) {
1003 outmod
= midgard_outmod_int_wrap
;
1005 bool sat
= instr
->dest
.saturate
|| instr
->op
== nir_op_fsat
;
1006 outmod
= sat
? midgard_outmod_sat
: midgard_outmod_none
;
1009 /* fmax(a, 0.0) can turn into a .pos modifier as an optimization */
1011 if (instr
->op
== nir_op_fmax
) {
1012 if (nir_is_fzero_constant(instr
->src
[0].src
)) {
1013 op
= midgard_alu_op_fmov
;
1015 outmod
= midgard_outmod_pos
;
1016 instr
->src
[0] = instr
->src
[1];
1017 } else if (nir_is_fzero_constant(instr
->src
[1].src
)) {
1018 op
= midgard_alu_op_fmov
;
1020 outmod
= midgard_outmod_pos
;
1024 /* Fetch unit, quirks, etc information */
1025 unsigned opcode_props
= alu_opcode_props
[op
].props
;
1026 bool quirk_flipped_r24
= opcode_props
& QUIRK_FLIPPED_R24
;
1028 /* src0 will always exist afaik, but src1 will not for 1-argument
1029 * instructions. The latter can only be fetched if the instruction
1030 * needs it, or else we may segfault. */
1032 unsigned src0
= nir_alu_src_index(ctx
, &instr
->src
[0]);
1033 unsigned src1
= nr_inputs
== 2 ? nir_alu_src_index(ctx
, &instr
->src
[1]) : SSA_UNUSED_0
;
1035 /* Rather than use the instruction generation helpers, we do it
1036 * ourselves here to avoid the mess */
1038 midgard_instruction ins
= {
1041 .src0
= quirk_flipped_r24
? SSA_UNUSED_1
: src0
,
1042 .src1
= quirk_flipped_r24
? src0
: src1
,
1047 nir_alu_src
*nirmods
[2] = { NULL
};
1049 if (nr_inputs
== 2) {
1050 nirmods
[0] = &instr
->src
[0];
1051 nirmods
[1] = &instr
->src
[1];
1052 } else if (nr_inputs
== 1) {
1053 nirmods
[quirk_flipped_r24
] = &instr
->src
[0];
1058 /* These were lowered to a move, so apply the corresponding mod */
1060 if (instr
->op
== nir_op_fneg
|| instr
->op
== nir_op_fabs
) {
1061 nir_alu_src
*s
= nirmods
[quirk_flipped_r24
];
1063 if (instr
->op
== nir_op_fneg
)
1064 s
->negate
= !s
->negate
;
1066 if (instr
->op
== nir_op_fabs
)
1070 bool is_int
= midgard_is_integer_op(op
);
1072 ins
.mask
= mask_of(nr_components
);
1074 midgard_vector_alu alu
= {
1076 .reg_mode
= reg_mode
,
1077 .dest_override
= dest_override
,
1080 .src1
= vector_alu_srco_unsigned(vector_alu_modifiers(nirmods
[0], is_int
, broadcast_swizzle
, half_1
, sext_1
)),
1081 .src2
= vector_alu_srco_unsigned(vector_alu_modifiers(nirmods
[1], is_int
, broadcast_swizzle
, half_2
, sext_2
)),
1084 /* Apply writemask if non-SSA, keeping in mind that we can't write to components that don't exist */
1087 ins
.mask
&= instr
->dest
.write_mask
;
1091 /* Late fixup for emulated instructions */
1093 if (instr
->op
== nir_op_b2f32
|| instr
->op
== nir_op_b2i32
) {
1094 /* Presently, our second argument is an inline #0 constant.
1095 * Switch over to an embedded 1.0 constant (that can't fit
1096 * inline, since we're 32-bit, not 16-bit like the inline
1099 ins
.ssa_args
.inline_constant
= false;
1100 ins
.ssa_args
.src1
= SSA_FIXED_REGISTER(REGISTER_CONSTANT
);
1101 ins
.has_constants
= true;
1103 if (instr
->op
== nir_op_b2f32
) {
1104 ins
.constants
[0] = 1.0f
;
1106 /* Type pun it into place */
1108 memcpy(&ins
.constants
[0], &one
, sizeof(uint32_t));
1111 ins
.alu
.src2
= vector_alu_srco_unsigned(blank_alu_src_xxxx
);
1112 } else if (nr_inputs
== 1 && !quirk_flipped_r24
) {
1113 /* Lots of instructions need a 0 plonked in */
1114 ins
.ssa_args
.inline_constant
= false;
1115 ins
.ssa_args
.src1
= SSA_FIXED_REGISTER(REGISTER_CONSTANT
);
1116 ins
.has_constants
= true;
1117 ins
.constants
[0] = 0.0f
;
1118 ins
.alu
.src2
= vector_alu_srco_unsigned(blank_alu_src_xxxx
);
1119 } else if (instr
->op
== nir_op_inot
) {
1120 /* ~b = ~(b & b), so duplicate the source */
1121 ins
.ssa_args
.src1
= ins
.ssa_args
.src0
;
1122 ins
.alu
.src2
= ins
.alu
.src1
;
1125 if ((opcode_props
& UNITS_ALL
) == UNIT_VLUT
) {
1126 /* To avoid duplicating the lookup tables (probably), true LUT
1127 * instructions can only operate as if they were scalars. Lower
1128 * them here by changing the component. */
1130 uint8_t original_swizzle
[4];
1131 memcpy(original_swizzle
, nirmods
[0]->swizzle
, sizeof(nirmods
[0]->swizzle
));
1132 unsigned orig_mask
= ins
.mask
;
1134 for (int i
= 0; i
< nr_components
; ++i
) {
1135 /* Mask the associated component, dropping the
1136 * instruction if needed */
1139 ins
.mask
&= orig_mask
;
1144 for (int j
= 0; j
< 4; ++j
)
1145 nirmods
[0]->swizzle
[j
] = original_swizzle
[i
]; /* Pull from the correct component */
1147 ins
.alu
.src1
= vector_alu_srco_unsigned(vector_alu_modifiers(nirmods
[0], is_int
, broadcast_swizzle
, half_1
, false));
1148 emit_mir_instruction(ctx
, ins
);
1151 emit_mir_instruction(ctx
, ins
);
1157 /* Uniforms and UBOs use a shared code path, as uniforms are just (slightly
1158 * optimized) versions of UBO #0 */
1162 compiler_context
*ctx
,
1165 nir_src
*indirect_offset
,
1168 /* TODO: half-floats */
1170 if (!indirect_offset
&& offset
< ctx
->uniform_cutoff
&& index
== 0) {
1171 /* Fast path: For the first 16 uniforms, direct accesses are
1172 * 0-cycle, since they're just a register fetch in the usual
1173 * case. So, we alias the registers while we're still in
1176 int reg_slot
= 23 - offset
;
1177 alias_ssa(ctx
, dest
, SSA_FIXED_REGISTER(reg_slot
));
1179 /* Otherwise, read from the 'special' UBO to access
1180 * higher-indexed uniforms, at a performance cost. More
1181 * generally, we're emitting a UBO read instruction. */
1183 midgard_instruction ins
= m_ld_uniform_32(dest
, offset
);
1185 /* TODO: Don't split */
1186 ins
.load_store
.varying_parameters
= (offset
& 7) << 7;
1187 ins
.load_store
.address
= offset
>> 3;
1189 if (indirect_offset
) {
1190 emit_indirect_offset(ctx
, indirect_offset
);
1191 ins
.load_store
.unknown
= 0x8700 | index
; /* xxx: what is this? */
1193 ins
.load_store
.unknown
= 0x1E00 | index
; /* xxx: what is this? */
1196 /* TODO respect index */
1198 emit_mir_instruction(ctx
, ins
);
1204 compiler_context
*ctx
,
1205 unsigned dest
, unsigned offset
,
1206 unsigned nr_comp
, unsigned component
,
1207 nir_src
*indirect_offset
, nir_alu_type type
)
1209 /* XXX: Half-floats? */
1210 /* TODO: swizzle, mask */
1212 midgard_instruction ins
= m_ld_vary_32(dest
, offset
);
1213 ins
.mask
= mask_of(nr_comp
);
1214 ins
.load_store
.swizzle
= SWIZZLE_XYZW
>> (2 * component
);
1216 midgard_varying_parameter p
= {
1218 .interpolation
= midgard_interp_default
,
1219 .flat
= /*var->data.interpolation == INTERP_MODE_FLAT*/ 0
1223 memcpy(&u
, &p
, sizeof(p
));
1224 ins
.load_store
.varying_parameters
= u
;
1226 if (indirect_offset
) {
1227 /* We need to add in the dynamic index, moved to r27.w */
1228 emit_indirect_offset(ctx
, indirect_offset
);
1229 ins
.load_store
.unknown
= 0x79e; /* xxx: what is this? */
1231 /* Just a direct load */
1232 ins
.load_store
.unknown
= 0x1e9e; /* xxx: what is this? */
1235 /* Use the type appropriate load */
1239 ins
.load_store
.op
= midgard_op_ld_vary_32u
;
1242 ins
.load_store
.op
= midgard_op_ld_vary_32i
;
1244 case nir_type_float
:
1245 ins
.load_store
.op
= midgard_op_ld_vary_32
;
1248 unreachable("Attempted to load unknown type");
1252 emit_mir_instruction(ctx
, ins
);
1256 emit_sysval_read(compiler_context
*ctx
, nir_instr
*instr
)
1259 /* Figure out which uniform this is */
1260 int sysval
= sysval_for_instr(ctx
, instr
, &dest
);
1261 void *val
= _mesa_hash_table_u64_search(ctx
->sysval_to_id
, sysval
);
1263 /* Sysvals are prefix uniforms */
1264 unsigned uniform
= ((uintptr_t) val
) - 1;
1266 /* Emit the read itself -- this is never indirect */
1267 emit_ubo_read(ctx
, dest
, uniform
, NULL
, 0);
1271 emit_intrinsic(compiler_context
*ctx
, nir_intrinsic_instr
*instr
)
1273 unsigned offset
= 0, reg
;
1275 switch (instr
->intrinsic
) {
1276 case nir_intrinsic_discard_if
:
1277 emit_condition(ctx
, &instr
->src
[0], true, COMPONENT_X
);
1281 case nir_intrinsic_discard
: {
1282 bool conditional
= instr
->intrinsic
== nir_intrinsic_discard_if
;
1283 struct midgard_instruction discard
= v_branch(conditional
, false);
1284 discard
.branch
.target_type
= TARGET_DISCARD
;
1285 emit_mir_instruction(ctx
, discard
);
1287 ctx
->can_discard
= true;
1291 case nir_intrinsic_load_uniform
:
1292 case nir_intrinsic_load_ubo
:
1293 case nir_intrinsic_load_input
: {
1294 bool is_uniform
= instr
->intrinsic
== nir_intrinsic_load_uniform
;
1295 bool is_ubo
= instr
->intrinsic
== nir_intrinsic_load_ubo
;
1297 /* Get the base type of the intrinsic */
1298 /* TODO: Infer type? Does it matter? */
1300 is_ubo
? nir_type_uint
: nir_intrinsic_type(instr
);
1301 t
= nir_alu_type_get_base_type(t
);
1304 offset
= nir_intrinsic_base(instr
);
1307 unsigned nr_comp
= nir_intrinsic_dest_components(instr
);
1309 nir_src
*src_offset
= nir_get_io_offset_src(instr
);
1311 bool direct
= nir_src_is_const(*src_offset
);
1314 offset
+= nir_src_as_uint(*src_offset
);
1316 /* We may need to apply a fractional offset */
1317 int component
= instr
->intrinsic
== nir_intrinsic_load_input
?
1318 nir_intrinsic_component(instr
) : 0;
1319 reg
= nir_dest_index(ctx
, &instr
->dest
);
1321 if (is_uniform
&& !ctx
->is_blend
) {
1322 emit_ubo_read(ctx
, reg
, ctx
->sysval_count
+ offset
, !direct
? &instr
->src
[0] : NULL
, 0);
1323 } else if (is_ubo
) {
1324 nir_src index
= instr
->src
[0];
1326 /* We don't yet support indirect UBOs. For indirect
1327 * block numbers (if that's possible), we don't know
1328 * enough about the hardware yet. For indirect sources,
1329 * we know what we need but we need to add some NIR
1330 * support for lowering correctly with respect to
1333 assert(nir_src_is_const(index
));
1334 assert(nir_src_is_const(*src_offset
));
1336 /* TODO: Alignment */
1337 assert((offset
& 0xF) == 0);
1339 uint32_t uindex
= nir_src_as_uint(index
) + 1;
1340 emit_ubo_read(ctx
, reg
, offset
/ 16, NULL
, uindex
);
1341 } else if (ctx
->stage
== MESA_SHADER_FRAGMENT
&& !ctx
->is_blend
) {
1342 emit_varying_read(ctx
, reg
, offset
, nr_comp
, component
, !direct
? &instr
->src
[0] : NULL
, t
);
1343 } else if (ctx
->is_blend
) {
1344 /* For blend shaders, load the input color, which is
1345 * preloaded to r0 */
1347 midgard_instruction move
= v_mov(reg
, blank_alu_src
, SSA_FIXED_REGISTER(0));
1348 emit_mir_instruction(ctx
, move
);
1349 } else if (ctx
->stage
== MESA_SHADER_VERTEX
) {
1350 midgard_instruction ins
= m_ld_attr_32(reg
, offset
);
1351 ins
.load_store
.unknown
= 0x1E1E; /* XXX: What is this? */
1352 ins
.mask
= mask_of(nr_comp
);
1354 /* Use the type appropriate load */
1358 ins
.load_store
.op
= midgard_op_ld_attr_32u
;
1361 ins
.load_store
.op
= midgard_op_ld_attr_32i
;
1363 case nir_type_float
:
1364 ins
.load_store
.op
= midgard_op_ld_attr_32
;
1367 unreachable("Attempted to load unknown type");
1371 emit_mir_instruction(ctx
, ins
);
1373 DBG("Unknown load\n");
1380 /* Reads 128-bit value raw off the tilebuffer during blending, tasty */
1382 case nir_intrinsic_load_raw_output_pan
:
1383 reg
= nir_dest_index(ctx
, &instr
->dest
);
1384 assert(ctx
->is_blend
);
1386 midgard_instruction ins
= m_ld_color_buffer_8(reg
, 0);
1387 emit_mir_instruction(ctx
, ins
);
1390 case nir_intrinsic_load_blend_const_color_rgba
: {
1391 assert(ctx
->is_blend
);
1392 reg
= nir_dest_index(ctx
, &instr
->dest
);
1394 /* Blend constants are embedded directly in the shader and
1395 * patched in, so we use some magic routing */
1397 midgard_instruction ins
= v_mov(SSA_FIXED_REGISTER(REGISTER_CONSTANT
), blank_alu_src
, reg
);
1398 ins
.has_constants
= true;
1399 ins
.has_blend_constant
= true;
1400 emit_mir_instruction(ctx
, ins
);
1404 case nir_intrinsic_store_output
:
1405 assert(nir_src_is_const(instr
->src
[1]) && "no indirect outputs");
1407 offset
= nir_intrinsic_base(instr
) + nir_src_as_uint(instr
->src
[1]);
1409 reg
= nir_src_index(ctx
, &instr
->src
[0]);
1411 if (ctx
->stage
== MESA_SHADER_FRAGMENT
) {
1412 /* gl_FragColor is not emitted with load/store
1413 * instructions. Instead, it gets plonked into
1414 * r0 at the end of the shader and we do the
1415 * framebuffer writeout dance. TODO: Defer
1418 midgard_instruction move
= v_mov(reg
, blank_alu_src
, SSA_FIXED_REGISTER(0));
1419 emit_mir_instruction(ctx
, move
);
1421 /* Save the index we're writing to for later reference
1422 * in the epilogue */
1424 ctx
->fragment_output
= reg
;
1425 } else if (ctx
->stage
== MESA_SHADER_VERTEX
) {
1426 /* Varyings are written into one of two special
1427 * varying register, r26 or r27. The register itself is
1428 * selected as the register in the st_vary instruction,
1429 * minus the base of 26. E.g. write into r27 and then
1430 * call st_vary(1) */
1432 midgard_instruction ins
= v_mov(reg
, blank_alu_src
, SSA_FIXED_REGISTER(26));
1433 emit_mir_instruction(ctx
, ins
);
1435 /* We should have been vectorized, though we don't
1436 * currently check that st_vary is emitted only once
1437 * per slot (this is relevant, since there's not a mask
1438 * parameter available on the store [set to 0 by the
1439 * blob]). We do respect the component by adjusting the
1442 unsigned component
= nir_intrinsic_component(instr
);
1444 midgard_instruction st
= m_st_vary_32(SSA_FIXED_REGISTER(0), offset
);
1445 st
.load_store
.unknown
= 0x1E9E; /* XXX: What is this? */
1446 st
.load_store
.swizzle
= SWIZZLE_XYZW
<< (2*component
);
1447 emit_mir_instruction(ctx
, st
);
1449 DBG("Unknown store\n");
1455 /* Special case of store_output for lowered blend shaders */
1456 case nir_intrinsic_store_raw_output_pan
:
1457 assert (ctx
->stage
== MESA_SHADER_FRAGMENT
);
1458 reg
= nir_src_index(ctx
, &instr
->src
[0]);
1460 midgard_instruction move
= v_mov(reg
, blank_alu_src
, SSA_FIXED_REGISTER(0));
1461 emit_mir_instruction(ctx
, move
);
1462 ctx
->fragment_output
= reg
;
1466 case nir_intrinsic_load_alpha_ref_float
:
1467 assert(instr
->dest
.is_ssa
);
1469 float ref_value
= ctx
->alpha_ref
;
1471 float *v
= ralloc_array(NULL
, float, 4);
1472 memcpy(v
, &ref_value
, sizeof(float));
1473 _mesa_hash_table_u64_insert(ctx
->ssa_constants
, instr
->dest
.ssa
.index
+ 1, v
);
1476 case nir_intrinsic_load_viewport_scale
:
1477 case nir_intrinsic_load_viewport_offset
:
1478 emit_sysval_read(ctx
, &instr
->instr
);
1482 printf ("Unhandled intrinsic\n");
1489 midgard_tex_format(enum glsl_sampler_dim dim
)
1492 case GLSL_SAMPLER_DIM_1D
:
1493 case GLSL_SAMPLER_DIM_BUF
:
1496 case GLSL_SAMPLER_DIM_2D
:
1497 case GLSL_SAMPLER_DIM_EXTERNAL
:
1500 case GLSL_SAMPLER_DIM_3D
:
1503 case GLSL_SAMPLER_DIM_CUBE
:
1504 return MALI_TEX_CUBE
;
1507 DBG("Unknown sampler dim type\n");
1513 /* Tries to attach an explicit LOD / bias as a constant. Returns whether this
1517 pan_attach_constant_bias(
1518 compiler_context
*ctx
,
1520 midgard_texture_word
*word
)
1522 /* To attach as constant, it has to *be* constant */
1524 if (!nir_src_is_const(lod
))
1527 float f
= nir_src_as_float(lod
);
1529 /* Break into fixed-point */
1531 float lod_frac
= f
- lod_int
;
1533 /* Carry over negative fractions */
1534 if (lod_frac
< 0.0) {
1540 word
->bias
= float_to_ubyte(lod_frac
);
1541 word
->bias_int
= lod_int
;
1546 static enum mali_sampler_type
1547 midgard_sampler_type(nir_alu_type t
) {
1548 switch (nir_alu_type_get_base_type(t
))
1550 case nir_type_float
:
1551 return MALI_SAMPLER_FLOAT
;
1553 return MALI_SAMPLER_SIGNED
;
1555 return MALI_SAMPLER_UNSIGNED
;
1557 unreachable("Unknown sampler type");
1562 emit_texop_native(compiler_context
*ctx
, nir_tex_instr
*instr
,
1563 unsigned midgard_texop
)
1566 //assert (!instr->sampler);
1567 //assert (!instr->texture_array_size);
1569 /* Allocate registers via a round robin scheme to alternate between the two registers */
1570 int reg
= ctx
->texture_op_count
& 1;
1571 int in_reg
= reg
, out_reg
= reg
;
1573 /* Make room for the reg */
1575 if (ctx
->texture_index
[reg
] > -1)
1576 unalias_ssa(ctx
, ctx
->texture_index
[reg
]);
1578 int texture_index
= instr
->texture_index
;
1579 int sampler_index
= texture_index
;
1581 /* No helper to build texture words -- we do it all here */
1582 midgard_instruction ins
= {
1583 .type
= TAG_TEXTURE_4
,
1586 .op
= midgard_texop
,
1587 .format
= midgard_tex_format(instr
->sampler_dim
),
1588 .texture_handle
= texture_index
,
1589 .sampler_handle
= sampler_index
,
1591 /* TODO: Regalloc it in */
1592 .swizzle
= SWIZZLE_XYZW
,
1598 .sampler_type
= midgard_sampler_type(instr
->dest_type
),
1602 for (unsigned i
= 0; i
< instr
->num_srcs
; ++i
) {
1603 int reg
= SSA_FIXED_REGISTER(REGISTER_TEXTURE_BASE
+ in_reg
);
1604 int index
= nir_src_index(ctx
, &instr
->src
[i
].src
);
1605 int nr_comp
= nir_src_num_components(instr
->src
[i
].src
);
1606 midgard_vector_alu_src alu_src
= blank_alu_src
;
1608 switch (instr
->src
[i
].src_type
) {
1609 case nir_tex_src_coord
: {
1610 if (instr
->sampler_dim
== GLSL_SAMPLER_DIM_CUBE
) {
1611 /* texelFetch is undefined on samplerCube */
1612 assert(midgard_texop
!= TEXTURE_OP_TEXEL_FETCH
);
1614 /* For cubemaps, we need to load coords into
1615 * special r27, and then use a special ld/st op
1616 * to select the face and copy the xy into the
1617 * texture register */
1619 alu_src
.swizzle
= SWIZZLE(COMPONENT_X
, COMPONENT_Y
, COMPONENT_Z
, COMPONENT_X
);
1621 midgard_instruction move
= v_mov(index
, alu_src
, SSA_FIXED_REGISTER(27));
1622 emit_mir_instruction(ctx
, move
);
1624 midgard_instruction st
= m_st_cubemap_coords(reg
, 0);
1625 st
.load_store
.unknown
= 0x24; /* XXX: What is this? */
1626 st
.mask
= 0x3; /* xy */
1627 st
.load_store
.swizzle
= alu_src
.swizzle
;
1628 emit_mir_instruction(ctx
, st
);
1630 ins
.texture
.in_reg_swizzle
= swizzle_of(2);
1632 ins
.texture
.in_reg_swizzle
= alu_src
.swizzle
= swizzle_of(nr_comp
);
1634 midgard_instruction mov
= v_mov(index
, alu_src
, reg
);
1635 mov
.mask
= mask_of(nr_comp
);
1636 emit_mir_instruction(ctx
, mov
);
1638 if (midgard_texop
== TEXTURE_OP_TEXEL_FETCH
) {
1639 /* Texel fetch opcodes care about the
1640 * values of z and w, so we actually
1641 * need to spill into a second register
1642 * for a texel fetch with register bias
1643 * (for non-2D). TODO: Implement that
1646 assert(instr
->sampler_dim
== GLSL_SAMPLER_DIM_2D
);
1648 midgard_instruction zero
= v_mov(index
, alu_src
, reg
);
1649 zero
.ssa_args
.inline_constant
= true;
1650 zero
.ssa_args
.src1
= SSA_FIXED_REGISTER(REGISTER_CONSTANT
);
1651 zero
.has_constants
= true;
1652 zero
.mask
= ~mov
.mask
;
1653 emit_mir_instruction(ctx
, zero
);
1655 ins
.texture
.in_reg_swizzle
= SWIZZLE_XYZZ
;
1657 /* Non-texel fetch doesn't need that
1658 * nonsense. However we do use the Z
1659 * for array indexing */
1660 bool is_3d
= instr
->sampler_dim
== GLSL_SAMPLER_DIM_3D
;
1661 ins
.texture
.in_reg_swizzle
= is_3d
? SWIZZLE_XYZZ
: SWIZZLE_XYXZ
;
1668 case nir_tex_src_bias
:
1669 case nir_tex_src_lod
: {
1670 /* Try as a constant if we can */
1672 bool is_txf
= midgard_texop
== TEXTURE_OP_TEXEL_FETCH
;
1673 if (!is_txf
&& pan_attach_constant_bias(ctx
, instr
->src
[i
].src
, &ins
.texture
))
1676 /* Otherwise we use a register. To keep RA simple, we
1677 * put the bias/LOD into the w component of the input
1678 * source, which is otherwise in xy */
1680 alu_src
.swizzle
= SWIZZLE_XXXX
;
1682 midgard_instruction mov
= v_mov(index
, alu_src
, reg
);
1683 mov
.mask
= 1 << COMPONENT_W
;
1684 emit_mir_instruction(ctx
, mov
);
1686 ins
.texture
.lod_register
= true;
1688 midgard_tex_register_select sel
= {
1698 memcpy(&packed
, &sel
, sizeof(packed
));
1699 ins
.texture
.bias
= packed
;
1705 unreachable("Unknown texture source type\n");
1709 /* Set registers to read and write from the same place */
1710 ins
.texture
.in_reg_select
= in_reg
;
1711 ins
.texture
.out_reg_select
= out_reg
;
1713 emit_mir_instruction(ctx
, ins
);
1715 int o_reg
= REGISTER_TEXTURE_BASE
+ out_reg
, o_index
= nir_dest_index(ctx
, &instr
->dest
);
1716 midgard_instruction ins2
= v_mov(SSA_FIXED_REGISTER(o_reg
), blank_alu_src
, o_index
);
1717 emit_mir_instruction(ctx
, ins2
);
1719 /* Used for .cont and .last hinting */
1720 ctx
->texture_op_count
++;
1724 emit_tex(compiler_context
*ctx
, nir_tex_instr
*instr
)
1726 /* Fixup op, since only textureLod is permitted in VS but NIR can give
1727 * generic tex in some cases (which confuses the hardware) */
1729 bool is_vertex
= ctx
->stage
== MESA_SHADER_VERTEX
;
1731 if (is_vertex
&& instr
->op
== nir_texop_tex
)
1732 instr
->op
= nir_texop_txl
;
1734 switch (instr
->op
) {
1737 emit_texop_native(ctx
, instr
, TEXTURE_OP_NORMAL
);
1740 emit_texop_native(ctx
, instr
, TEXTURE_OP_LOD
);
1743 emit_texop_native(ctx
, instr
, TEXTURE_OP_TEXEL_FETCH
);
1746 emit_sysval_read(ctx
, &instr
->instr
);
1749 unreachable("Unhanlded texture op");
1754 emit_jump(compiler_context
*ctx
, nir_jump_instr
*instr
)
1756 switch (instr
->type
) {
1757 case nir_jump_break
: {
1758 /* Emit a branch out of the loop */
1759 struct midgard_instruction br
= v_branch(false, false);
1760 br
.branch
.target_type
= TARGET_BREAK
;
1761 br
.branch
.target_break
= ctx
->current_loop_depth
;
1762 emit_mir_instruction(ctx
, br
);
1769 DBG("Unknown jump type %d\n", instr
->type
);
1775 emit_instr(compiler_context
*ctx
, struct nir_instr
*instr
)
1777 switch (instr
->type
) {
1778 case nir_instr_type_load_const
:
1779 emit_load_const(ctx
, nir_instr_as_load_const(instr
));
1782 case nir_instr_type_intrinsic
:
1783 emit_intrinsic(ctx
, nir_instr_as_intrinsic(instr
));
1786 case nir_instr_type_alu
:
1787 emit_alu(ctx
, nir_instr_as_alu(instr
));
1790 case nir_instr_type_tex
:
1791 emit_tex(ctx
, nir_instr_as_tex(instr
));
1794 case nir_instr_type_jump
:
1795 emit_jump(ctx
, nir_instr_as_jump(instr
));
1798 case nir_instr_type_ssa_undef
:
1803 DBG("Unhandled instruction type\n");
1809 /* ALU instructions can inline or embed constants, which decreases register
1810 * pressure and saves space. */
1812 #define CONDITIONAL_ATTACH(src) { \
1813 void *entry = _mesa_hash_table_u64_search(ctx->ssa_constants, alu->ssa_args.src + 1); \
1816 attach_constants(ctx, alu, entry, alu->ssa_args.src + 1); \
1817 alu->ssa_args.src = SSA_FIXED_REGISTER(REGISTER_CONSTANT); \
1822 inline_alu_constants(compiler_context
*ctx
)
1824 mir_foreach_instr(ctx
, alu
) {
1825 /* Other instructions cannot inline constants */
1826 if (alu
->type
!= TAG_ALU_4
) continue;
1828 /* If there is already a constant here, we can do nothing */
1829 if (alu
->has_constants
) continue;
1831 /* It makes no sense to inline constants on a branch */
1832 if (alu
->compact_branch
|| alu
->prepacked_branch
) continue;
1834 CONDITIONAL_ATTACH(src0
);
1836 if (!alu
->has_constants
) {
1837 CONDITIONAL_ATTACH(src1
)
1838 } else if (!alu
->inline_constant
) {
1839 /* Corner case: _two_ vec4 constants, for instance with a
1840 * csel. For this case, we can only use a constant
1841 * register for one, we'll have to emit a move for the
1842 * other. Note, if both arguments are constants, then
1843 * necessarily neither argument depends on the value of
1844 * any particular register. As the destination register
1845 * will be wiped, that means we can spill the constant
1846 * to the destination register.
1849 void *entry
= _mesa_hash_table_u64_search(ctx
->ssa_constants
, alu
->ssa_args
.src1
+ 1);
1850 unsigned scratch
= alu
->ssa_args
.dest
;
1853 midgard_instruction ins
= v_mov(SSA_FIXED_REGISTER(REGISTER_CONSTANT
), blank_alu_src
, scratch
);
1854 attach_constants(ctx
, &ins
, entry
, alu
->ssa_args
.src1
+ 1);
1856 /* Force a break XXX Defer r31 writes */
1857 ins
.unit
= UNIT_VLUT
;
1859 /* Set the source */
1860 alu
->ssa_args
.src1
= scratch
;
1862 /* Inject us -before- the last instruction which set r31 */
1863 mir_insert_instruction_before(mir_prev_op(alu
), ins
);
1869 /* Midgard supports two types of constants, embedded constants (128-bit) and
1870 * inline constants (16-bit). Sometimes, especially with scalar ops, embedded
1871 * constants can be demoted to inline constants, for space savings and
1872 * sometimes a performance boost */
1875 embedded_to_inline_constant(compiler_context
*ctx
)
1877 mir_foreach_instr(ctx
, ins
) {
1878 if (!ins
->has_constants
) continue;
1880 if (ins
->ssa_args
.inline_constant
) continue;
1882 /* Blend constants must not be inlined by definition */
1883 if (ins
->has_blend_constant
) continue;
1885 /* We can inline 32-bit (sometimes) or 16-bit (usually) */
1886 bool is_16
= ins
->alu
.reg_mode
== midgard_reg_mode_16
;
1887 bool is_32
= ins
->alu
.reg_mode
== midgard_reg_mode_32
;
1889 if (!(is_16
|| is_32
))
1892 /* src1 cannot be an inline constant due to encoding
1893 * restrictions. So, if possible we try to flip the arguments
1896 int op
= ins
->alu
.op
;
1898 if (ins
->ssa_args
.src0
== SSA_FIXED_REGISTER(REGISTER_CONSTANT
)) {
1900 /* These ops require an operational change to flip
1901 * their arguments TODO */
1902 case midgard_alu_op_flt
:
1903 case midgard_alu_op_fle
:
1904 case midgard_alu_op_ilt
:
1905 case midgard_alu_op_ile
:
1906 case midgard_alu_op_fcsel
:
1907 case midgard_alu_op_icsel
:
1908 DBG("Missed non-commutative flip (%s)\n", alu_opcode_props
[op
].name
);
1913 if (alu_opcode_props
[op
].props
& OP_COMMUTES
) {
1914 /* Flip the SSA numbers */
1915 ins
->ssa_args
.src0
= ins
->ssa_args
.src1
;
1916 ins
->ssa_args
.src1
= SSA_FIXED_REGISTER(REGISTER_CONSTANT
);
1918 /* And flip the modifiers */
1922 src_temp
= ins
->alu
.src2
;
1923 ins
->alu
.src2
= ins
->alu
.src1
;
1924 ins
->alu
.src1
= src_temp
;
1928 if (ins
->ssa_args
.src1
== SSA_FIXED_REGISTER(REGISTER_CONSTANT
)) {
1929 /* Extract the source information */
1931 midgard_vector_alu_src
*src
;
1932 int q
= ins
->alu
.src2
;
1933 midgard_vector_alu_src
*m
= (midgard_vector_alu_src
*) &q
;
1936 /* Component is from the swizzle, e.g. r26.w -> w component. TODO: What if x is masked out? */
1937 int component
= src
->swizzle
& 3;
1939 /* Scale constant appropriately, if we can legally */
1940 uint16_t scaled_constant
= 0;
1942 if (midgard_is_integer_op(op
) || is_16
) {
1943 unsigned int *iconstants
= (unsigned int *) ins
->constants
;
1944 scaled_constant
= (uint16_t) iconstants
[component
];
1946 /* Constant overflow after resize */
1947 if (scaled_constant
!= iconstants
[component
])
1950 float original
= (float) ins
->constants
[component
];
1951 scaled_constant
= _mesa_float_to_half(original
);
1953 /* Check for loss of precision. If this is
1954 * mediump, we don't care, but for a highp
1955 * shader, we need to pay attention. NIR
1956 * doesn't yet tell us which mode we're in!
1957 * Practically this prevents most constants
1958 * from being inlined, sadly. */
1960 float fp32
= _mesa_half_to_float(scaled_constant
);
1962 if (fp32
!= original
)
1966 /* We don't know how to handle these with a constant */
1968 if (src
->mod
|| src
->half
|| src
->rep_low
|| src
->rep_high
) {
1969 DBG("Bailing inline constant...\n");
1973 /* Make sure that the constant is not itself a
1974 * vector by checking if all accessed values
1975 * (by the swizzle) are the same. */
1977 uint32_t *cons
= (uint32_t *) ins
->constants
;
1978 uint32_t value
= cons
[component
];
1980 bool is_vector
= false;
1981 unsigned mask
= effective_writemask(&ins
->alu
, ins
->mask
);
1983 for (int c
= 1; c
< 4; ++c
) {
1984 /* We only care if this component is actually used */
1985 if (!(mask
& (1 << c
)))
1988 uint32_t test
= cons
[(src
->swizzle
>> (2 * c
)) & 3];
1990 if (test
!= value
) {
1999 /* Get rid of the embedded constant */
2000 ins
->has_constants
= false;
2001 ins
->ssa_args
.src1
= SSA_UNUSED_0
;
2002 ins
->ssa_args
.inline_constant
= true;
2003 ins
->inline_constant
= scaled_constant
;
2008 /* Map normal SSA sources to other SSA sources / fixed registers (like
2012 map_ssa_to_alias(compiler_context
*ctx
, int *ref
)
2014 /* Sign is used quite deliberately for unused */
2018 unsigned int alias
= (uintptr_t) _mesa_hash_table_u64_search(ctx
->ssa_to_alias
, *ref
+ 1);
2021 /* Remove entry in leftovers to avoid a redunant fmov */
2023 struct set_entry
*leftover
= _mesa_set_search(ctx
->leftover_ssa_to_alias
, ((void *) (uintptr_t) (*ref
+ 1)));
2026 _mesa_set_remove(ctx
->leftover_ssa_to_alias
, leftover
);
2028 /* Assign the alias map */
2034 /* Basic dead code elimination on the MIR itself, which cleans up e.g. the
2035 * texture pipeline */
2038 midgard_opt_dead_code_eliminate(compiler_context
*ctx
, midgard_block
*block
)
2040 bool progress
= false;
2042 mir_foreach_instr_in_block_safe(block
, ins
) {
2043 if (ins
->type
!= TAG_ALU_4
) continue;
2044 if (ins
->compact_branch
) continue;
2046 if (ins
->ssa_args
.dest
>= SSA_FIXED_MINIMUM
) continue;
2047 if (mir_is_live_after(ctx
, block
, ins
, ins
->ssa_args
.dest
)) continue;
2049 mir_remove_instruction(ins
);
2056 /* Dead code elimination for branches at the end of a block - only one branch
2057 * per block is legal semantically */
2060 midgard_opt_cull_dead_branch(compiler_context
*ctx
, midgard_block
*block
)
2062 bool branched
= false;
2064 mir_foreach_instr_in_block_safe(block
, ins
) {
2065 if (!midgard_is_branch_unit(ins
->unit
)) continue;
2067 /* We ignore prepacked branches since the fragment epilogue is
2068 * just generally special */
2069 if (ins
->prepacked_branch
) continue;
2071 /* Discards are similarly special and may not correspond to the
2074 if (ins
->branch
.target_type
== TARGET_DISCARD
) continue;
2077 /* We already branched, so this is dead */
2078 mir_remove_instruction(ins
);
2086 mir_nontrivial_mod(midgard_vector_alu_src src
, bool is_int
, unsigned mask
)
2089 if (!is_int
&& src
.mod
) return true;
2091 /* Other int mods don't matter in isolation */
2092 if (is_int
&& src
.mod
== midgard_int_shift
) return true;
2094 /* size-conversion */
2095 if (src
.half
) return true;
2098 for (unsigned c
= 0; c
< 4; ++c
) {
2099 if (!(mask
& (1 << c
))) continue;
2100 if (((src
.swizzle
>> (2*c
)) & 3) != c
) return true;
2107 mir_nontrivial_source2_mod(midgard_instruction
*ins
)
2109 bool is_int
= midgard_is_integer_op(ins
->alu
.op
);
2111 midgard_vector_alu_src src2
=
2112 vector_alu_from_unsigned(ins
->alu
.src2
);
2114 return mir_nontrivial_mod(src2
, is_int
, ins
->mask
);
2118 mir_nontrivial_outmod(midgard_instruction
*ins
)
2120 bool is_int
= midgard_is_integer_op(ins
->alu
.op
);
2121 unsigned mod
= ins
->alu
.outmod
;
2123 /* Type conversion is a sort of outmod */
2124 if (ins
->alu
.dest_override
!= midgard_dest_override_none
)
2128 return mod
!= midgard_outmod_int_wrap
;
2130 return mod
!= midgard_outmod_none
;
2134 midgard_opt_copy_prop(compiler_context
*ctx
, midgard_block
*block
)
2136 bool progress
= false;
2138 mir_foreach_instr_in_block_safe(block
, ins
) {
2139 if (ins
->type
!= TAG_ALU_4
) continue;
2140 if (!OP_IS_MOVE(ins
->alu
.op
)) continue;
2142 unsigned from
= ins
->ssa_args
.src1
;
2143 unsigned to
= ins
->ssa_args
.dest
;
2145 /* We only work on pure SSA */
2147 if (to
>= SSA_FIXED_MINIMUM
) continue;
2148 if (from
>= SSA_FIXED_MINIMUM
) continue;
2149 if (to
>= ctx
->func
->impl
->ssa_alloc
) continue;
2150 if (from
>= ctx
->func
->impl
->ssa_alloc
) continue;
2152 /* Constant propagation is not handled here, either */
2153 if (ins
->ssa_args
.inline_constant
) continue;
2154 if (ins
->has_constants
) continue;
2156 if (mir_nontrivial_source2_mod(ins
)) continue;
2157 if (mir_nontrivial_outmod(ins
)) continue;
2159 /* We're clear -- rewrite */
2160 mir_rewrite_index_src(ctx
, to
, from
);
2161 mir_remove_instruction(ins
);
2168 /* fmov.pos is an idiom for fpos. Propoagate the .pos up to the source, so then
2169 * the move can be propagated away entirely */
2172 mir_compose_float_outmod(midgard_outmod_float
*outmod
, midgard_outmod_float comp
)
2175 if (comp
== midgard_outmod_none
)
2178 if (*outmod
== midgard_outmod_none
) {
2183 /* TODO: Compose rules */
2188 midgard_opt_pos_propagate(compiler_context
*ctx
, midgard_block
*block
)
2190 bool progress
= false;
2192 mir_foreach_instr_in_block_safe(block
, ins
) {
2193 if (ins
->type
!= TAG_ALU_4
) continue;
2194 if (ins
->alu
.op
!= midgard_alu_op_fmov
) continue;
2195 if (ins
->alu
.outmod
!= midgard_outmod_pos
) continue;
2197 /* TODO: Registers? */
2198 unsigned src
= ins
->ssa_args
.src1
;
2199 if (src
>= ctx
->func
->impl
->ssa_alloc
) continue;
2200 assert(!mir_has_multiple_writes(ctx
, src
));
2202 /* There might be a source modifier, too */
2203 if (mir_nontrivial_source2_mod(ins
)) continue;
2205 /* Backpropagate the modifier */
2206 mir_foreach_instr_in_block_from_rev(block
, v
, mir_prev_op(ins
)) {
2207 if (v
->type
!= TAG_ALU_4
) continue;
2208 if (v
->ssa_args
.dest
!= src
) continue;
2210 /* Can we even take a float outmod? */
2211 if (midgard_is_integer_out_op(v
->alu
.op
)) continue;
2213 midgard_outmod_float temp
= v
->alu
.outmod
;
2214 progress
|= mir_compose_float_outmod(&temp
, ins
->alu
.outmod
);
2216 /* Throw in the towel.. */
2217 if (!progress
) break;
2219 /* Otherwise, transfer the modifier */
2220 v
->alu
.outmod
= temp
;
2221 ins
->alu
.outmod
= midgard_outmod_none
;
2230 /* The following passes reorder MIR instructions to enable better scheduling */
2233 midgard_pair_load_store(compiler_context
*ctx
, midgard_block
*block
)
2235 mir_foreach_instr_in_block_safe(block
, ins
) {
2236 if (ins
->type
!= TAG_LOAD_STORE_4
) continue;
2238 /* We've found a load/store op. Check if next is also load/store. */
2239 midgard_instruction
*next_op
= mir_next_op(ins
);
2240 if (&next_op
->link
!= &block
->instructions
) {
2241 if (next_op
->type
== TAG_LOAD_STORE_4
) {
2242 /* If so, we're done since we're a pair */
2243 ins
= mir_next_op(ins
);
2247 /* Maximum search distance to pair, to avoid register pressure disasters */
2248 int search_distance
= 8;
2250 /* Otherwise, we have an orphaned load/store -- search for another load */
2251 mir_foreach_instr_in_block_from(block
, c
, mir_next_op(ins
)) {
2252 /* Terminate search if necessary */
2253 if (!(search_distance
--)) break;
2255 if (c
->type
!= TAG_LOAD_STORE_4
) continue;
2257 /* Stores cannot be reordered, since they have
2258 * dependencies. For the same reason, indirect
2259 * loads cannot be reordered as their index is
2260 * loaded in r27.w */
2262 if (OP_IS_STORE(c
->load_store
.op
)) continue;
2264 /* It appears the 0x800 bit is set whenever a
2265 * load is direct, unset when it is indirect.
2266 * Skip indirect loads. */
2268 if (!(c
->load_store
.unknown
& 0x800)) continue;
2270 /* We found one! Move it up to pair and remove it from the old location */
2272 mir_insert_instruction_before(ins
, *c
);
2273 mir_remove_instruction(c
);
2281 /* If there are leftovers after the below pass, emit actual fmov
2282 * instructions for the slow-but-correct path */
2285 emit_leftover_move(compiler_context
*ctx
)
2287 set_foreach(ctx
->leftover_ssa_to_alias
, leftover
) {
2288 int base
= ((uintptr_t) leftover
->key
) - 1;
2291 map_ssa_to_alias(ctx
, &mapped
);
2292 EMIT(mov
, mapped
, blank_alu_src
, base
);
2297 actualise_ssa_to_alias(compiler_context
*ctx
)
2299 mir_foreach_instr(ctx
, ins
) {
2300 map_ssa_to_alias(ctx
, &ins
->ssa_args
.src0
);
2301 map_ssa_to_alias(ctx
, &ins
->ssa_args
.src1
);
2304 emit_leftover_move(ctx
);
2308 emit_fragment_epilogue(compiler_context
*ctx
)
2310 /* Special case: writing out constants requires us to include the move
2311 * explicitly now, so shove it into r0 */
2313 void *constant_value
= _mesa_hash_table_u64_search(ctx
->ssa_constants
, ctx
->fragment_output
+ 1);
2315 if (constant_value
) {
2316 midgard_instruction ins
= v_mov(SSA_FIXED_REGISTER(REGISTER_CONSTANT
), blank_alu_src
, SSA_FIXED_REGISTER(0));
2317 attach_constants(ctx
, &ins
, constant_value
, ctx
->fragment_output
+ 1);
2318 emit_mir_instruction(ctx
, ins
);
2321 /* Perform the actual fragment writeout. We have two writeout/branch
2322 * instructions, forming a loop until writeout is successful as per the
2323 * docs. TODO: gl_FragDepth */
2325 EMIT(alu_br_compact_cond
, midgard_jmp_writeout_op_writeout
, TAG_ALU_4
, 0, midgard_condition_always
);
2326 EMIT(alu_br_compact_cond
, midgard_jmp_writeout_op_writeout
, TAG_ALU_4
, -1, midgard_condition_always
);
2329 static midgard_block
*
2330 emit_block(compiler_context
*ctx
, nir_block
*block
)
2332 midgard_block
*this_block
= calloc(sizeof(midgard_block
), 1);
2333 list_addtail(&this_block
->link
, &ctx
->blocks
);
2335 this_block
->is_scheduled
= false;
2338 ctx
->texture_index
[0] = -1;
2339 ctx
->texture_index
[1] = -1;
2341 /* Add us as a successor to the block we are following */
2342 if (ctx
->current_block
)
2343 midgard_block_add_successor(ctx
->current_block
, this_block
);
2345 /* Set up current block */
2346 list_inithead(&this_block
->instructions
);
2347 ctx
->current_block
= this_block
;
2349 nir_foreach_instr(instr
, block
) {
2350 emit_instr(ctx
, instr
);
2351 ++ctx
->instruction_count
;
2354 inline_alu_constants(ctx
);
2355 embedded_to_inline_constant(ctx
);
2357 /* Perform heavylifting for aliasing */
2358 actualise_ssa_to_alias(ctx
);
2360 midgard_pair_load_store(ctx
, this_block
);
2362 /* Append fragment shader epilogue (value writeout) */
2363 if (ctx
->stage
== MESA_SHADER_FRAGMENT
) {
2364 if (block
== nir_impl_last_block(ctx
->func
->impl
)) {
2365 emit_fragment_epilogue(ctx
);
2369 if (block
== nir_start_block(ctx
->func
->impl
))
2370 ctx
->initial_block
= this_block
;
2372 if (block
== nir_impl_last_block(ctx
->func
->impl
))
2373 ctx
->final_block
= this_block
;
2375 /* Allow the next control flow to access us retroactively, for
2377 ctx
->current_block
= this_block
;
2379 /* Document the fallthrough chain */
2380 ctx
->previous_source_block
= this_block
;
2385 static midgard_block
*emit_cf_list(struct compiler_context
*ctx
, struct exec_list
*list
);
2388 emit_if(struct compiler_context
*ctx
, nir_if
*nif
)
2390 /* Conditional branches expect the condition in r31.w; emit a move for
2391 * that in the _previous_ block (which is the current block). */
2392 emit_condition(ctx
, &nif
->condition
, true, COMPONENT_X
);
2394 /* Speculatively emit the branch, but we can't fill it in until later */
2395 EMIT(branch
, true, true);
2396 midgard_instruction
*then_branch
= mir_last_in_block(ctx
->current_block
);
2398 /* Emit the two subblocks */
2399 midgard_block
*then_block
= emit_cf_list(ctx
, &nif
->then_list
);
2401 /* Emit a jump from the end of the then block to the end of the else */
2402 EMIT(branch
, false, false);
2403 midgard_instruction
*then_exit
= mir_last_in_block(ctx
->current_block
);
2405 /* Emit second block, and check if it's empty */
2407 int else_idx
= ctx
->block_count
;
2408 int count_in
= ctx
->instruction_count
;
2409 midgard_block
*else_block
= emit_cf_list(ctx
, &nif
->else_list
);
2410 int after_else_idx
= ctx
->block_count
;
2412 /* Now that we have the subblocks emitted, fix up the branches */
2417 if (ctx
->instruction_count
== count_in
) {
2418 /* The else block is empty, so don't emit an exit jump */
2419 mir_remove_instruction(then_exit
);
2420 then_branch
->branch
.target_block
= after_else_idx
;
2422 then_branch
->branch
.target_block
= else_idx
;
2423 then_exit
->branch
.target_block
= after_else_idx
;
2428 emit_loop(struct compiler_context
*ctx
, nir_loop
*nloop
)
2430 /* Remember where we are */
2431 midgard_block
*start_block
= ctx
->current_block
;
2433 /* Allocate a loop number, growing the current inner loop depth */
2434 int loop_idx
= ++ctx
->current_loop_depth
;
2436 /* Get index from before the body so we can loop back later */
2437 int start_idx
= ctx
->block_count
;
2439 /* Emit the body itself */
2440 emit_cf_list(ctx
, &nloop
->body
);
2442 /* Branch back to loop back */
2443 struct midgard_instruction br_back
= v_branch(false, false);
2444 br_back
.branch
.target_block
= start_idx
;
2445 emit_mir_instruction(ctx
, br_back
);
2447 /* Mark down that branch in the graph. Note that we're really branching
2448 * to the block *after* we started in. TODO: Why doesn't the branch
2449 * itself have an off-by-one then...? */
2450 midgard_block_add_successor(ctx
->current_block
, start_block
->successors
[0]);
2452 /* Find the index of the block about to follow us (note: we don't add
2453 * one; blocks are 0-indexed so we get a fencepost problem) */
2454 int break_block_idx
= ctx
->block_count
;
2456 /* Fix up the break statements we emitted to point to the right place,
2457 * now that we can allocate a block number for them */
2459 list_for_each_entry_from(struct midgard_block
, block
, start_block
, &ctx
->blocks
, link
) {
2460 mir_foreach_instr_in_block(block
, ins
) {
2461 if (ins
->type
!= TAG_ALU_4
) continue;
2462 if (!ins
->compact_branch
) continue;
2463 if (ins
->prepacked_branch
) continue;
2465 /* We found a branch -- check the type to see if we need to do anything */
2466 if (ins
->branch
.target_type
!= TARGET_BREAK
) continue;
2468 /* It's a break! Check if it's our break */
2469 if (ins
->branch
.target_break
!= loop_idx
) continue;
2471 /* Okay, cool, we're breaking out of this loop.
2472 * Rewrite from a break to a goto */
2474 ins
->branch
.target_type
= TARGET_GOTO
;
2475 ins
->branch
.target_block
= break_block_idx
;
2479 /* Now that we've finished emitting the loop, free up the depth again
2480 * so we play nice with recursion amid nested loops */
2481 --ctx
->current_loop_depth
;
2483 /* Dump loop stats */
2487 static midgard_block
*
2488 emit_cf_list(struct compiler_context
*ctx
, struct exec_list
*list
)
2490 midgard_block
*start_block
= NULL
;
2492 foreach_list_typed(nir_cf_node
, node
, node
, list
) {
2493 switch (node
->type
) {
2494 case nir_cf_node_block
: {
2495 midgard_block
*block
= emit_block(ctx
, nir_cf_node_as_block(node
));
2498 start_block
= block
;
2503 case nir_cf_node_if
:
2504 emit_if(ctx
, nir_cf_node_as_if(node
));
2507 case nir_cf_node_loop
:
2508 emit_loop(ctx
, nir_cf_node_as_loop(node
));
2511 case nir_cf_node_function
:
2520 /* Due to lookahead, we need to report the first tag executed in the command
2521 * stream and in branch targets. An initial block might be empty, so iterate
2522 * until we find one that 'works' */
2525 midgard_get_first_tag_from_block(compiler_context
*ctx
, unsigned block_idx
)
2527 midgard_block
*initial_block
= mir_get_block(ctx
, block_idx
);
2529 unsigned first_tag
= 0;
2532 midgard_bundle
*initial_bundle
= util_dynarray_element(&initial_block
->bundles
, midgard_bundle
, 0);
2534 if (initial_bundle
) {
2535 first_tag
= initial_bundle
->tag
;
2539 /* Initial block is empty, try the next block */
2540 initial_block
= list_first_entry(&(initial_block
->link
), midgard_block
, link
);
2541 } while(initial_block
!= NULL
);
2548 midgard_compile_shader_nir(nir_shader
*nir
, midgard_program
*program
, bool is_blend
)
2550 struct util_dynarray
*compiled
= &program
->compiled
;
2552 midgard_debug
= debug_get_option_midgard_debug();
2554 compiler_context ictx
= {
2556 .stage
= nir
->info
.stage
,
2558 .is_blend
= is_blend
,
2559 .blend_constant_offset
= 0,
2561 .alpha_ref
= program
->alpha_ref
2564 compiler_context
*ctx
= &ictx
;
2566 /* TODO: Decide this at runtime */
2567 ctx
->uniform_cutoff
= 8;
2569 /* Initialize at a global (not block) level hash tables */
2571 ctx
->ssa_constants
= _mesa_hash_table_u64_create(NULL
);
2572 ctx
->ssa_to_alias
= _mesa_hash_table_u64_create(NULL
);
2573 ctx
->hash_to_temp
= _mesa_hash_table_u64_create(NULL
);
2574 ctx
->sysval_to_id
= _mesa_hash_table_u64_create(NULL
);
2575 ctx
->leftover_ssa_to_alias
= _mesa_set_create(NULL
, _mesa_hash_pointer
, _mesa_key_pointer_equal
);
2577 /* Record the varying mapping for the command stream's bookkeeping */
2579 struct exec_list
*varyings
=
2580 ctx
->stage
== MESA_SHADER_VERTEX
? &nir
->outputs
: &nir
->inputs
;
2582 unsigned max_varying
= 0;
2583 nir_foreach_variable(var
, varyings
) {
2584 unsigned loc
= var
->data
.driver_location
;
2585 unsigned sz
= glsl_type_size(var
->type
, FALSE
);
2587 for (int c
= 0; c
< sz
; ++c
) {
2588 program
->varyings
[loc
+ c
] = var
->data
.location
+ c
;
2589 max_varying
= MAX2(max_varying
, loc
+ c
);
2593 /* Lower gl_Position pre-optimisation, but after lowering vars to ssa
2594 * (so we don't accidentally duplicate the epilogue since mesa/st has
2595 * messed with our I/O quite a bit already) */
2597 NIR_PASS_V(nir
, nir_lower_vars_to_ssa
);
2599 if (ctx
->stage
== MESA_SHADER_VERTEX
) {
2600 NIR_PASS_V(nir
, nir_lower_viewport_transform
);
2601 NIR_PASS_V(nir
, nir_clamp_psiz
, 1.0, 1024.0);
2604 NIR_PASS_V(nir
, nir_lower_var_copies
);
2605 NIR_PASS_V(nir
, nir_lower_vars_to_ssa
);
2606 NIR_PASS_V(nir
, nir_split_var_copies
);
2607 NIR_PASS_V(nir
, nir_lower_var_copies
);
2608 NIR_PASS_V(nir
, nir_lower_global_vars_to_local
);
2609 NIR_PASS_V(nir
, nir_lower_var_copies
);
2610 NIR_PASS_V(nir
, nir_lower_vars_to_ssa
);
2612 NIR_PASS_V(nir
, nir_lower_io
, nir_var_all
, glsl_type_size
, 0);
2614 /* Optimisation passes */
2618 if (midgard_debug
& MIDGARD_DBG_SHADERS
) {
2619 nir_print_shader(nir
, stdout
);
2622 /* Assign sysvals and counts, now that we're sure
2623 * (post-optimisation) */
2625 midgard_nir_assign_sysvals(ctx
, nir
);
2627 program
->uniform_count
= nir
->num_uniforms
;
2628 program
->sysval_count
= ctx
->sysval_count
;
2629 memcpy(program
->sysvals
, ctx
->sysvals
, sizeof(ctx
->sysvals
[0]) * ctx
->sysval_count
);
2631 program
->attribute_count
= (ctx
->stage
== MESA_SHADER_VERTEX
) ? nir
->num_inputs
: 0;
2632 program
->varying_count
= max_varying
+ 1; /* Fencepost off-by-one */
2634 nir_foreach_function(func
, nir
) {
2638 list_inithead(&ctx
->blocks
);
2639 ctx
->block_count
= 0;
2642 emit_cf_list(ctx
, &func
->impl
->body
);
2643 emit_block(ctx
, func
->impl
->end_block
);
2645 break; /* TODO: Multi-function shaders */
2648 util_dynarray_init(compiled
, NULL
);
2650 /* MIR-level optimizations */
2652 bool progress
= false;
2657 mir_foreach_block(ctx
, block
) {
2658 progress
|= midgard_opt_pos_propagate(ctx
, block
);
2659 progress
|= midgard_opt_copy_prop(ctx
, block
);
2660 progress
|= midgard_opt_dead_code_eliminate(ctx
, block
);
2664 /* Nested control-flow can result in dead branches at the end of the
2665 * block. This messes with our analysis and is just dead code, so cull
2667 mir_foreach_block(ctx
, block
) {
2668 midgard_opt_cull_dead_branch(ctx
, block
);
2672 schedule_program(ctx
);
2674 /* Now that all the bundles are scheduled and we can calculate block
2675 * sizes, emit actual branch instructions rather than placeholders */
2677 int br_block_idx
= 0;
2679 mir_foreach_block(ctx
, block
) {
2680 util_dynarray_foreach(&block
->bundles
, midgard_bundle
, bundle
) {
2681 for (int c
= 0; c
< bundle
->instruction_count
; ++c
) {
2682 midgard_instruction
*ins
= bundle
->instructions
[c
];
2684 if (!midgard_is_branch_unit(ins
->unit
)) continue;
2686 if (ins
->prepacked_branch
) continue;
2688 /* Parse some basic branch info */
2689 bool is_compact
= ins
->unit
== ALU_ENAB_BR_COMPACT
;
2690 bool is_conditional
= ins
->branch
.conditional
;
2691 bool is_inverted
= ins
->branch
.invert_conditional
;
2692 bool is_discard
= ins
->branch
.target_type
== TARGET_DISCARD
;
2694 /* Determine the block we're jumping to */
2695 int target_number
= ins
->branch
.target_block
;
2697 /* Report the destination tag */
2698 int dest_tag
= is_discard
? 0 : midgard_get_first_tag_from_block(ctx
, target_number
);
2700 /* Count up the number of quadwords we're
2701 * jumping over = number of quadwords until
2702 * (br_block_idx, target_number) */
2704 int quadword_offset
= 0;
2707 /* Jump to the end of the shader. We
2708 * need to include not only the
2709 * following blocks, but also the
2710 * contents of our current block (since
2711 * discard can come in the middle of
2714 midgard_block
*blk
= mir_get_block(ctx
, br_block_idx
+ 1);
2716 for (midgard_bundle
*bun
= bundle
+ 1; bun
< (midgard_bundle
*)((char*) block
->bundles
.data
+ block
->bundles
.size
); ++bun
) {
2717 quadword_offset
+= quadword_size(bun
->tag
);
2720 mir_foreach_block_from(ctx
, blk
, b
) {
2721 quadword_offset
+= b
->quadword_count
;
2724 } else if (target_number
> br_block_idx
) {
2727 for (int idx
= br_block_idx
+ 1; idx
< target_number
; ++idx
) {
2728 midgard_block
*blk
= mir_get_block(ctx
, idx
);
2731 quadword_offset
+= blk
->quadword_count
;
2734 /* Jump backwards */
2736 for (int idx
= br_block_idx
; idx
>= target_number
; --idx
) {
2737 midgard_block
*blk
= mir_get_block(ctx
, idx
);
2740 quadword_offset
-= blk
->quadword_count
;
2744 /* Unconditional extended branches (far jumps)
2745 * have issues, so we always use a conditional
2746 * branch, setting the condition to always for
2747 * unconditional. For compact unconditional
2748 * branches, cond isn't used so it doesn't
2749 * matter what we pick. */
2751 midgard_condition cond
=
2752 !is_conditional
? midgard_condition_always
:
2753 is_inverted
? midgard_condition_false
:
2754 midgard_condition_true
;
2756 midgard_jmp_writeout_op op
=
2757 is_discard
? midgard_jmp_writeout_op_discard
:
2758 (is_compact
&& !is_conditional
) ? midgard_jmp_writeout_op_branch_uncond
:
2759 midgard_jmp_writeout_op_branch_cond
;
2762 midgard_branch_extended branch
=
2763 midgard_create_branch_extended(
2768 memcpy(&ins
->branch_extended
, &branch
, sizeof(branch
));
2769 } else if (is_conditional
|| is_discard
) {
2770 midgard_branch_cond branch
= {
2772 .dest_tag
= dest_tag
,
2773 .offset
= quadword_offset
,
2777 assert(branch
.offset
== quadword_offset
);
2779 memcpy(&ins
->br_compact
, &branch
, sizeof(branch
));
2781 assert(op
== midgard_jmp_writeout_op_branch_uncond
);
2783 midgard_branch_uncond branch
= {
2785 .dest_tag
= dest_tag
,
2786 .offset
= quadword_offset
,
2790 assert(branch
.offset
== quadword_offset
);
2792 memcpy(&ins
->br_compact
, &branch
, sizeof(branch
));
2800 /* Emit flat binary from the instruction arrays. Iterate each block in
2801 * sequence. Save instruction boundaries such that lookahead tags can
2802 * be assigned easily */
2804 /* Cache _all_ bundles in source order for lookahead across failed branches */
2806 int bundle_count
= 0;
2807 mir_foreach_block(ctx
, block
) {
2808 bundle_count
+= block
->bundles
.size
/ sizeof(midgard_bundle
);
2810 midgard_bundle
**source_order_bundles
= malloc(sizeof(midgard_bundle
*) * bundle_count
);
2812 mir_foreach_block(ctx
, block
) {
2813 util_dynarray_foreach(&block
->bundles
, midgard_bundle
, bundle
) {
2814 source_order_bundles
[bundle_idx
++] = bundle
;
2818 int current_bundle
= 0;
2820 /* Midgard prefetches instruction types, so during emission we
2821 * need to lookahead. Unless this is the last instruction, in
2822 * which we return 1. Or if this is the second to last and the
2823 * last is an ALU, then it's also 1... */
2825 mir_foreach_block(ctx
, block
) {
2826 mir_foreach_bundle_in_block(block
, bundle
) {
2829 if (current_bundle
+ 1 < bundle_count
) {
2830 uint8_t next
= source_order_bundles
[current_bundle
+ 1]->tag
;
2832 if (!(current_bundle
+ 2 < bundle_count
) && IS_ALU(next
)) {
2839 emit_binary_bundle(ctx
, bundle
, compiled
, lookahead
);
2843 /* TODO: Free deeper */
2844 //util_dynarray_fini(&block->instructions);
2847 free(source_order_bundles
);
2849 /* Report the very first tag executed */
2850 program
->first_tag
= midgard_get_first_tag_from_block(ctx
, 0);
2852 /* Deal with off-by-one related to the fencepost problem */
2853 program
->work_register_count
= ctx
->work_registers
+ 1;
2855 program
->can_discard
= ctx
->can_discard
;
2856 program
->uniform_cutoff
= ctx
->uniform_cutoff
;
2858 program
->blend_patch_offset
= ctx
->blend_constant_offset
;
2860 if (midgard_debug
& MIDGARD_DBG_SHADERS
)
2861 disassemble_midgard(program
->compiled
.data
, program
->compiled
.size
);
2863 if (midgard_debug
& MIDGARD_DBG_SHADERDB
) {
2864 unsigned nr_bundles
= 0, nr_ins
= 0, nr_quadwords
= 0;
2866 /* Count instructions and bundles */
2868 mir_foreach_instr_global(ctx
, ins
) {
2872 mir_foreach_block(ctx
, block
) {
2873 nr_bundles
+= util_dynarray_num_elements(
2874 &block
->bundles
, midgard_bundle
);
2876 nr_quadwords
+= block
->quadword_count
;
2879 /* Calculate thread count. There are certain cutoffs by
2880 * register count for thread count */
2882 unsigned nr_registers
= program
->work_register_count
;
2884 unsigned nr_threads
=
2885 (nr_registers
<= 4) ? 4 :
2886 (nr_registers
<= 8) ? 2 :
2891 fprintf(stderr
, "shader%d - %s shader: "
2892 "%u inst, %u bundles, %u quadwords, "
2893 "%u registers, %u threads, %u loops\n",
2895 gl_shader_stage_name(ctx
->stage
),
2896 nr_ins
, nr_bundles
, nr_quadwords
,
2897 nr_registers
, nr_threads
,