1 /* Copyright © 2011 Intel Corporation
3 * Permission is hereby granted, free of charge, to any person obtaining a
4 * copy of this software and associated documentation files (the "Software"),
5 * to deal in the Software without restriction, including without limitation
6 * the rights to use, copy, modify, merge, publish, distribute, sublicense,
7 * and/or sell copies of the Software, and to permit persons to whom the
8 * Software is furnished to do so, subject to the following conditions:
10 * The above copyright notice and this permission notice (including the next
11 * paragraph) shall be included in all copies or substantial portions of the
14 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
15 * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
16 * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL
17 * THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
18 * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING
19 * FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS
27 #include "main/macros.h"
28 #include "program/prog_print.h"
29 #include "program/prog_parameter.h"
35 vec4_instruction::get_dst(void)
37 struct brw_reg brw_reg
;
41 brw_reg
= brw_vec8_grf(dst
.reg
+ dst
.reg_offset
, 0);
42 brw_reg
= retype(brw_reg
, dst
.type
);
43 brw_reg
.dw1
.bits
.writemask
= dst
.writemask
;
47 brw_reg
= brw_message_reg(dst
.reg
+ dst
.reg_offset
);
48 brw_reg
= retype(brw_reg
, dst
.type
);
49 brw_reg
.dw1
.bits
.writemask
= dst
.writemask
;
53 brw_reg
= dst
.fixed_hw_reg
;
57 brw_reg
= brw_null_reg();
61 assert(!"not reached");
62 brw_reg
= brw_null_reg();
69 vec4_instruction::get_src(const struct brw_vec4_prog_data
*prog_data
, int i
)
71 struct brw_reg brw_reg
;
73 switch (src
[i
].file
) {
75 brw_reg
= brw_vec8_grf(src
[i
].reg
+ src
[i
].reg_offset
, 0);
76 brw_reg
= retype(brw_reg
, src
[i
].type
);
77 brw_reg
.dw1
.bits
.swizzle
= src
[i
].swizzle
;
79 brw_reg
= brw_abs(brw_reg
);
81 brw_reg
= negate(brw_reg
);
85 switch (src
[i
].type
) {
86 case BRW_REGISTER_TYPE_F
:
87 brw_reg
= brw_imm_f(src
[i
].imm
.f
);
89 case BRW_REGISTER_TYPE_D
:
90 brw_reg
= brw_imm_d(src
[i
].imm
.i
);
92 case BRW_REGISTER_TYPE_UD
:
93 brw_reg
= brw_imm_ud(src
[i
].imm
.u
);
96 assert(!"not reached");
97 brw_reg
= brw_null_reg();
103 brw_reg
= stride(brw_vec4_grf(prog_data
->dispatch_grf_start_reg
+
104 (src
[i
].reg
+ src
[i
].reg_offset
) / 2,
105 ((src
[i
].reg
+ src
[i
].reg_offset
) % 2) * 4),
107 brw_reg
= retype(brw_reg
, src
[i
].type
);
108 brw_reg
.dw1
.bits
.swizzle
= src
[i
].swizzle
;
110 brw_reg
= brw_abs(brw_reg
);
112 brw_reg
= negate(brw_reg
);
114 /* This should have been moved to pull constants. */
115 assert(!src
[i
].reladdr
);
119 brw_reg
= src
[i
].fixed_hw_reg
;
123 /* Probably unused. */
124 brw_reg
= brw_null_reg();
128 assert(!"not reached");
129 brw_reg
= brw_null_reg();
136 vec4_generator::vec4_generator(struct brw_context
*brw
,
137 struct gl_shader_program
*shader_prog
,
138 struct gl_program
*prog
,
139 struct brw_vec4_prog_data
*prog_data
,
142 : brw(brw
), shader_prog(shader_prog
), prog(prog
), prog_data(prog_data
),
143 mem_ctx(mem_ctx
), debug_flag(debug_flag
)
145 shader
= shader_prog
? shader_prog
->_LinkedShaders
[MESA_SHADER_VERTEX
] : NULL
;
147 p
= rzalloc(mem_ctx
, struct brw_compile
);
148 brw_init_compile(brw
, p
, mem_ctx
);
151 vec4_generator::~vec4_generator()
156 vec4_generator::mark_surface_used(unsigned surf_index
)
158 assert(surf_index
< BRW_MAX_VS_SURFACES
);
160 prog_data
->binding_table_size
= MAX2(prog_data
->binding_table_size
,
165 vec4_generator::generate_math1_gen4(vec4_instruction
*inst
,
171 brw_math_function(inst
->opcode
),
174 BRW_MATH_DATA_VECTOR
,
175 BRW_MATH_PRECISION_FULL
);
179 check_gen6_math_src_arg(struct brw_reg src
)
181 /* Source swizzles are ignored. */
184 assert(src
.dw1
.bits
.swizzle
== BRW_SWIZZLE_XYZW
);
188 vec4_generator::generate_math1_gen6(vec4_instruction
*inst
,
192 /* Can't do writemask because math can't be align16. */
193 assert(dst
.dw1
.bits
.writemask
== WRITEMASK_XYZW
);
194 check_gen6_math_src_arg(src
);
196 brw_set_access_mode(p
, BRW_ALIGN_1
);
199 brw_math_function(inst
->opcode
),
202 BRW_MATH_DATA_SCALAR
,
203 BRW_MATH_PRECISION_FULL
);
204 brw_set_access_mode(p
, BRW_ALIGN_16
);
208 vec4_generator::generate_math2_gen7(vec4_instruction
*inst
,
215 brw_math_function(inst
->opcode
),
220 vec4_generator::generate_math2_gen6(vec4_instruction
*inst
,
225 /* Can't do writemask because math can't be align16. */
226 assert(dst
.dw1
.bits
.writemask
== WRITEMASK_XYZW
);
227 /* Source swizzles are ignored. */
228 check_gen6_math_src_arg(src0
);
229 check_gen6_math_src_arg(src1
);
231 brw_set_access_mode(p
, BRW_ALIGN_1
);
234 brw_math_function(inst
->opcode
),
236 brw_set_access_mode(p
, BRW_ALIGN_16
);
240 vec4_generator::generate_math2_gen4(vec4_instruction
*inst
,
245 /* From the Ironlake PRM, Volume 4, Part 1, Section 6.1.13
248 * "Operand0[7]. For the INT DIV functions, this operand is the
251 * "Operand1[7]. For the INT DIV functions, this operand is the
254 bool is_int_div
= inst
->opcode
!= SHADER_OPCODE_POW
;
255 struct brw_reg
&op0
= is_int_div
? src1
: src0
;
256 struct brw_reg
&op1
= is_int_div
? src0
: src1
;
258 brw_push_insn_state(p
);
259 brw_set_saturate(p
, false);
260 brw_set_predicate_control(p
, BRW_PREDICATE_NONE
);
261 brw_MOV(p
, retype(brw_message_reg(inst
->base_mrf
+ 1), op1
.type
), op1
);
262 brw_pop_insn_state(p
);
266 brw_math_function(inst
->opcode
),
269 BRW_MATH_DATA_VECTOR
,
270 BRW_MATH_PRECISION_FULL
);
274 vec4_generator::generate_tex(vec4_instruction
*inst
,
281 switch (inst
->opcode
) {
282 case SHADER_OPCODE_TEX
:
283 case SHADER_OPCODE_TXL
:
284 if (inst
->shadow_compare
) {
285 msg_type
= GEN5_SAMPLER_MESSAGE_SAMPLE_LOD_COMPARE
;
287 msg_type
= GEN5_SAMPLER_MESSAGE_SAMPLE_LOD
;
290 case SHADER_OPCODE_TXD
:
291 if (inst
->shadow_compare
) {
292 /* Gen7.5+. Otherwise, lowered by brw_lower_texture_gradients(). */
293 assert(brw
->is_haswell
);
294 msg_type
= HSW_SAMPLER_MESSAGE_SAMPLE_DERIV_COMPARE
;
296 msg_type
= GEN5_SAMPLER_MESSAGE_SAMPLE_DERIVS
;
299 case SHADER_OPCODE_TXF
:
300 msg_type
= GEN5_SAMPLER_MESSAGE_SAMPLE_LD
;
302 case SHADER_OPCODE_TXF_MS
:
304 msg_type
= GEN7_SAMPLER_MESSAGE_SAMPLE_LD2DMS
;
306 msg_type
= GEN5_SAMPLER_MESSAGE_SAMPLE_LD
;
308 case SHADER_OPCODE_TXS
:
309 msg_type
= GEN5_SAMPLER_MESSAGE_SAMPLE_RESINFO
;
312 assert(!"should not get here: invalid VS texture opcode");
316 switch (inst
->opcode
) {
317 case SHADER_OPCODE_TEX
:
318 case SHADER_OPCODE_TXL
:
319 if (inst
->shadow_compare
) {
320 msg_type
= BRW_SAMPLER_MESSAGE_SIMD4X2_SAMPLE_LOD_COMPARE
;
321 assert(inst
->mlen
== 3);
323 msg_type
= BRW_SAMPLER_MESSAGE_SIMD4X2_SAMPLE_LOD
;
324 assert(inst
->mlen
== 2);
327 case SHADER_OPCODE_TXD
:
328 /* There is no sample_d_c message; comparisons are done manually. */
329 msg_type
= BRW_SAMPLER_MESSAGE_SIMD4X2_SAMPLE_GRADIENTS
;
330 assert(inst
->mlen
== 4);
332 case SHADER_OPCODE_TXF
:
333 msg_type
= BRW_SAMPLER_MESSAGE_SIMD4X2_LD
;
334 assert(inst
->mlen
== 2);
336 case SHADER_OPCODE_TXS
:
337 msg_type
= BRW_SAMPLER_MESSAGE_SIMD4X2_RESINFO
;
338 assert(inst
->mlen
== 2);
341 assert(!"should not get here: invalid VS texture opcode");
346 assert(msg_type
!= -1);
348 /* Load the message header if present. If there's a texture offset, we need
349 * to set it up explicitly and load the offset bitfield. Otherwise, we can
350 * use an implied move from g0 to the first message register.
352 if (inst
->texture_offset
) {
353 /* Explicitly set up the message header by copying g0 to the MRF. */
354 brw_push_insn_state(p
);
355 brw_set_mask_control(p
, BRW_MASK_DISABLE
);
356 brw_MOV(p
, retype(brw_message_reg(inst
->base_mrf
), BRW_REGISTER_TYPE_UD
),
357 retype(brw_vec8_grf(0, 0), BRW_REGISTER_TYPE_UD
));
359 /* Then set the offset bits in DWord 2. */
360 brw_set_access_mode(p
, BRW_ALIGN_1
);
362 retype(brw_vec1_reg(BRW_MESSAGE_REGISTER_FILE
, inst
->base_mrf
, 2),
363 BRW_REGISTER_TYPE_UD
),
364 brw_imm_uw(inst
->texture_offset
));
365 brw_pop_insn_state(p
);
366 } else if (inst
->header_present
) {
367 /* Set up an implied move from g0 to the MRF. */
368 src
= brw_vec8_grf(0, 0);
371 uint32_t return_format
;
374 case BRW_REGISTER_TYPE_D
:
375 return_format
= BRW_SAMPLER_RETURN_FORMAT_SINT32
;
377 case BRW_REGISTER_TYPE_UD
:
378 return_format
= BRW_SAMPLER_RETURN_FORMAT_UINT32
;
381 return_format
= BRW_SAMPLER_RETURN_FORMAT_FLOAT32
;
389 SURF_INDEX_VS_TEXTURE(inst
->sampler
),
392 1, /* response length */
394 inst
->header_present
,
395 BRW_SAMPLER_SIMD_MODE_SIMD4X2
,
398 mark_surface_used(SURF_INDEX_VS_TEXTURE(inst
->sampler
));
402 vec4_generator::generate_urb_write(vec4_instruction
*inst
)
405 brw_null_reg(), /* dest */
406 inst
->base_mrf
, /* starting mrf reg nr */
407 brw_vec8_grf(0, 0), /* src */
408 false, /* allocate */
411 0, /* response len */
413 inst
->eot
, /* writes complete */
414 inst
->offset
, /* urb destination offset */
415 BRW_URB_SWIZZLE_INTERLEAVE
);
419 vec4_generator::generate_oword_dual_block_offsets(struct brw_reg m1
,
420 struct brw_reg index
)
422 int second_vertex_offset
;
425 second_vertex_offset
= 1;
427 second_vertex_offset
= 16;
429 m1
= retype(m1
, BRW_REGISTER_TYPE_D
);
431 /* Set up M1 (message payload). Only the block offsets in M1.0 and
432 * M1.4 are used, and the rest are ignored.
434 struct brw_reg m1_0
= suboffset(vec1(m1
), 0);
435 struct brw_reg m1_4
= suboffset(vec1(m1
), 4);
436 struct brw_reg index_0
= suboffset(vec1(index
), 0);
437 struct brw_reg index_4
= suboffset(vec1(index
), 4);
439 brw_push_insn_state(p
);
440 brw_set_mask_control(p
, BRW_MASK_DISABLE
);
441 brw_set_access_mode(p
, BRW_ALIGN_1
);
443 brw_MOV(p
, m1_0
, index_0
);
445 if (index
.file
== BRW_IMMEDIATE_VALUE
) {
446 index_4
.dw1
.ud
+= second_vertex_offset
;
447 brw_MOV(p
, m1_4
, index_4
);
449 brw_ADD(p
, m1_4
, index_4
, brw_imm_d(second_vertex_offset
));
452 brw_pop_insn_state(p
);
456 vec4_generator::generate_unpack_flags(vec4_instruction
*inst
,
459 brw_push_insn_state(p
);
460 brw_set_mask_control(p
, BRW_MASK_DISABLE
);
461 brw_set_access_mode(p
, BRW_ALIGN_1
);
463 struct brw_reg flags
= brw_flag_reg(0, 0);
464 struct brw_reg dst_0
= suboffset(vec1(dst
), 0);
465 struct brw_reg dst_4
= suboffset(vec1(dst
), 4);
467 brw_AND(p
, dst_0
, flags
, brw_imm_ud(0x0f));
468 brw_AND(p
, dst_4
, flags
, brw_imm_ud(0xf0));
469 brw_SHR(p
, dst_4
, dst_4
, brw_imm_ud(4));
471 brw_pop_insn_state(p
);
475 vec4_generator::generate_scratch_read(vec4_instruction
*inst
,
477 struct brw_reg index
)
479 struct brw_reg header
= brw_vec8_grf(0, 0);
481 gen6_resolve_implied_move(p
, &header
, inst
->base_mrf
);
483 generate_oword_dual_block_offsets(brw_message_reg(inst
->base_mrf
+ 1),
489 msg_type
= GEN6_DATAPORT_READ_MESSAGE_OWORD_DUAL_BLOCK_READ
;
490 else if (brw
->gen
== 5 || brw
->is_g4x
)
491 msg_type
= G45_DATAPORT_READ_MESSAGE_OWORD_DUAL_BLOCK_READ
;
493 msg_type
= BRW_DATAPORT_READ_MESSAGE_OWORD_DUAL_BLOCK_READ
;
495 /* Each of the 8 channel enables is considered for whether each
498 struct brw_instruction
*send
= brw_next_insn(p
, BRW_OPCODE_SEND
);
499 brw_set_dest(p
, send
, dst
);
500 brw_set_src0(p
, send
, header
);
502 send
->header
.destreg__conditionalmod
= inst
->base_mrf
;
503 brw_set_dp_read_message(p
, send
,
504 255, /* binding table index: stateless access */
505 BRW_DATAPORT_OWORD_DUAL_BLOCK_1OWORD
,
507 BRW_DATAPORT_READ_TARGET_RENDER_CACHE
,
509 true, /* header_present */
514 vec4_generator::generate_scratch_write(vec4_instruction
*inst
,
517 struct brw_reg index
)
519 struct brw_reg header
= brw_vec8_grf(0, 0);
522 /* If the instruction is predicated, we'll predicate the send, not
525 brw_set_predicate_control(p
, false);
527 gen6_resolve_implied_move(p
, &header
, inst
->base_mrf
);
529 generate_oword_dual_block_offsets(brw_message_reg(inst
->base_mrf
+ 1),
533 retype(brw_message_reg(inst
->base_mrf
+ 2), BRW_REGISTER_TYPE_D
),
534 retype(src
, BRW_REGISTER_TYPE_D
));
539 msg_type
= GEN7_DATAPORT_WRITE_MESSAGE_OWORD_DUAL_BLOCK_WRITE
;
540 else if (brw
->gen
== 6)
541 msg_type
= GEN6_DATAPORT_WRITE_MESSAGE_OWORD_DUAL_BLOCK_WRITE
;
543 msg_type
= BRW_DATAPORT_WRITE_MESSAGE_OWORD_DUAL_BLOCK_WRITE
;
545 brw_set_predicate_control(p
, inst
->predicate
);
547 /* Pre-gen6, we have to specify write commits to ensure ordering
548 * between reads and writes within a thread. Afterwards, that's
549 * guaranteed and write commits only matter for inter-thread
553 write_commit
= false;
555 /* The visitor set up our destination register to be g0. This
556 * means that when the next read comes along, we will end up
557 * reading from g0 and causing a block on the write commit. For
558 * write-after-read, we are relying on the value of the previous
559 * read being used (and thus blocking on completion) before our
560 * write is executed. This means we have to be careful in
561 * instruction scheduling to not violate this assumption.
566 /* Each of the 8 channel enables is considered for whether each
569 struct brw_instruction
*send
= brw_next_insn(p
, BRW_OPCODE_SEND
);
570 brw_set_dest(p
, send
, dst
);
571 brw_set_src0(p
, send
, header
);
573 send
->header
.destreg__conditionalmod
= inst
->base_mrf
;
574 brw_set_dp_write_message(p
, send
,
575 255, /* binding table index: stateless access */
576 BRW_DATAPORT_OWORD_DUAL_BLOCK_1OWORD
,
579 true, /* header present */
580 false, /* not a render target write */
581 write_commit
, /* rlen */
587 vec4_generator::generate_pull_constant_load(vec4_instruction
*inst
,
589 struct brw_reg index
,
590 struct brw_reg offset
)
592 assert(brw
->gen
<= 7);
593 assert(index
.file
== BRW_IMMEDIATE_VALUE
&&
594 index
.type
== BRW_REGISTER_TYPE_UD
);
595 uint32_t surf_index
= index
.dw1
.ud
;
597 struct brw_reg header
= brw_vec8_grf(0, 0);
599 gen6_resolve_implied_move(p
, &header
, inst
->base_mrf
);
601 brw_MOV(p
, retype(brw_message_reg(inst
->base_mrf
+ 1), BRW_REGISTER_TYPE_D
),
607 msg_type
= GEN6_DATAPORT_READ_MESSAGE_OWORD_DUAL_BLOCK_READ
;
608 else if (brw
->gen
== 5 || brw
->is_g4x
)
609 msg_type
= G45_DATAPORT_READ_MESSAGE_OWORD_DUAL_BLOCK_READ
;
611 msg_type
= BRW_DATAPORT_READ_MESSAGE_OWORD_DUAL_BLOCK_READ
;
613 /* Each of the 8 channel enables is considered for whether each
616 struct brw_instruction
*send
= brw_next_insn(p
, BRW_OPCODE_SEND
);
617 brw_set_dest(p
, send
, dst
);
618 brw_set_src0(p
, send
, header
);
620 send
->header
.destreg__conditionalmod
= inst
->base_mrf
;
621 brw_set_dp_read_message(p
, send
,
623 BRW_DATAPORT_OWORD_DUAL_BLOCK_1OWORD
,
625 BRW_DATAPORT_READ_TARGET_DATA_CACHE
,
627 true, /* header_present */
630 mark_surface_used(surf_index
);
634 vec4_generator::generate_pull_constant_load_gen7(vec4_instruction
*inst
,
636 struct brw_reg surf_index
,
637 struct brw_reg offset
)
639 assert(surf_index
.file
== BRW_IMMEDIATE_VALUE
&&
640 surf_index
.type
== BRW_REGISTER_TYPE_UD
);
642 brw_instruction
*insn
= brw_next_insn(p
, BRW_OPCODE_SEND
);
643 brw_set_dest(p
, insn
, dst
);
644 brw_set_src0(p
, insn
, offset
);
645 brw_set_sampler_message(p
, insn
,
647 0, /* LD message ignores sampler unit */
648 GEN5_SAMPLER_MESSAGE_SAMPLE_LD
,
651 false, /* no header */
652 BRW_SAMPLER_SIMD_MODE_SIMD4X2
,
655 mark_surface_used(surf_index
.dw1
.ud
);
659 * Generate assembly for a Vec4 IR instruction.
661 * \param instruction The Vec4 IR instruction to generate code for.
662 * \param dst The destination register.
663 * \param src An array of up to three source registers.
666 vec4_generator::generate_vec4_instruction(vec4_instruction
*instruction
,
670 vec4_instruction
*inst
= (vec4_instruction
*) instruction
;
672 switch (inst
->opcode
) {
674 brw_MOV(p
, dst
, src
[0]);
677 brw_ADD(p
, dst
, src
[0], src
[1]);
680 brw_MUL(p
, dst
, src
[0], src
[1]);
682 case BRW_OPCODE_MACH
:
683 brw_set_acc_write_control(p
, 1);
684 brw_MACH(p
, dst
, src
[0], src
[1]);
685 brw_set_acc_write_control(p
, 0);
689 brw_MAD(p
, dst
, src
[0], src
[1], src
[2]);
693 brw_FRC(p
, dst
, src
[0]);
695 case BRW_OPCODE_RNDD
:
696 brw_RNDD(p
, dst
, src
[0]);
698 case BRW_OPCODE_RNDE
:
699 brw_RNDE(p
, dst
, src
[0]);
701 case BRW_OPCODE_RNDZ
:
702 brw_RNDZ(p
, dst
, src
[0]);
706 brw_AND(p
, dst
, src
[0], src
[1]);
709 brw_OR(p
, dst
, src
[0], src
[1]);
712 brw_XOR(p
, dst
, src
[0], src
[1]);
715 brw_NOT(p
, dst
, src
[0]);
718 brw_ASR(p
, dst
, src
[0], src
[1]);
721 brw_SHR(p
, dst
, src
[0], src
[1]);
724 brw_SHL(p
, dst
, src
[0], src
[1]);
728 brw_CMP(p
, dst
, inst
->conditional_mod
, src
[0], src
[1]);
731 brw_SEL(p
, dst
, src
[0], src
[1]);
735 brw_DPH(p
, dst
, src
[0], src
[1]);
739 brw_DP4(p
, dst
, src
[0], src
[1]);
743 brw_DP3(p
, dst
, src
[0], src
[1]);
747 brw_DP2(p
, dst
, src
[0], src
[1]);
750 case BRW_OPCODE_F32TO16
:
751 brw_F32TO16(p
, dst
, src
[0]);
754 case BRW_OPCODE_F16TO32
:
755 brw_F16TO32(p
, dst
, src
[0]);
759 brw_LRP(p
, dst
, src
[0], src
[1], src
[2]);
762 case BRW_OPCODE_BFREV
:
763 /* BFREV only supports UD type for src and dst. */
764 brw_BFREV(p
, retype(dst
, BRW_REGISTER_TYPE_UD
),
765 retype(src
[0], BRW_REGISTER_TYPE_UD
));
768 /* FBH only supports UD type for dst. */
769 brw_FBH(p
, retype(dst
, BRW_REGISTER_TYPE_UD
), src
[0]);
772 /* FBL only supports UD type for dst. */
773 brw_FBL(p
, retype(dst
, BRW_REGISTER_TYPE_UD
), src
[0]);
775 case BRW_OPCODE_CBIT
:
776 /* CBIT only supports UD type for dst. */
777 brw_CBIT(p
, retype(dst
, BRW_REGISTER_TYPE_UD
), src
[0]);
781 brw_BFE(p
, dst
, src
[0], src
[1], src
[2]);
784 case BRW_OPCODE_BFI1
:
785 brw_BFI1(p
, dst
, src
[0], src
[1]);
787 case BRW_OPCODE_BFI2
:
788 brw_BFI2(p
, dst
, src
[0], src
[1], src
[2]);
792 if (inst
->src
[0].file
!= BAD_FILE
) {
793 /* The instruction has an embedded compare (only allowed on gen6) */
794 assert(brw
->gen
== 6);
795 gen6_IF(p
, inst
->conditional_mod
, src
[0], src
[1]);
797 struct brw_instruction
*brw_inst
= brw_IF(p
, BRW_EXECUTE_8
);
798 brw_inst
->header
.predicate_control
= inst
->predicate
;
802 case BRW_OPCODE_ELSE
:
805 case BRW_OPCODE_ENDIF
:
810 brw_DO(p
, BRW_EXECUTE_8
);
813 case BRW_OPCODE_BREAK
:
815 brw_set_predicate_control(p
, BRW_PREDICATE_NONE
);
817 case BRW_OPCODE_CONTINUE
:
818 /* FINISHME: We need to write the loop instruction support still. */
823 brw_set_predicate_control(p
, BRW_PREDICATE_NONE
);
826 case BRW_OPCODE_WHILE
:
830 case SHADER_OPCODE_RCP
:
831 case SHADER_OPCODE_RSQ
:
832 case SHADER_OPCODE_SQRT
:
833 case SHADER_OPCODE_EXP2
:
834 case SHADER_OPCODE_LOG2
:
835 case SHADER_OPCODE_SIN
:
836 case SHADER_OPCODE_COS
:
838 generate_math1_gen6(inst
, dst
, src
[0]);
840 /* Also works for Gen7. */
841 generate_math1_gen4(inst
, dst
, src
[0]);
845 case SHADER_OPCODE_POW
:
846 case SHADER_OPCODE_INT_QUOTIENT
:
847 case SHADER_OPCODE_INT_REMAINDER
:
849 generate_math2_gen7(inst
, dst
, src
[0], src
[1]);
850 } else if (brw
->gen
== 6) {
851 generate_math2_gen6(inst
, dst
, src
[0], src
[1]);
853 generate_math2_gen4(inst
, dst
, src
[0], src
[1]);
857 case SHADER_OPCODE_TEX
:
858 case SHADER_OPCODE_TXD
:
859 case SHADER_OPCODE_TXF
:
860 case SHADER_OPCODE_TXF_MS
:
861 case SHADER_OPCODE_TXL
:
862 case SHADER_OPCODE_TXS
:
863 generate_tex(inst
, dst
, src
[0]);
866 case VS_OPCODE_URB_WRITE
:
867 generate_urb_write(inst
);
870 case VS_OPCODE_SCRATCH_READ
:
871 generate_scratch_read(inst
, dst
, src
[0]);
874 case VS_OPCODE_SCRATCH_WRITE
:
875 generate_scratch_write(inst
, dst
, src
[0], src
[1]);
878 case VS_OPCODE_PULL_CONSTANT_LOAD
:
879 generate_pull_constant_load(inst
, dst
, src
[0], src
[1]);
882 case VS_OPCODE_PULL_CONSTANT_LOAD_GEN7
:
883 generate_pull_constant_load_gen7(inst
, dst
, src
[0], src
[1]);
886 case SHADER_OPCODE_SHADER_TIME_ADD
:
887 brw_shader_time_add(p
, src
[0], SURF_INDEX_VS_SHADER_TIME
);
888 mark_surface_used(SURF_INDEX_VS_SHADER_TIME
);
891 case VS_OPCODE_UNPACK_FLAGS_SIMD4X2
:
892 generate_unpack_flags(inst
, dst
);
896 if (inst
->opcode
< (int) ARRAY_SIZE(opcode_descs
)) {
897 _mesa_problem(ctx
, "Unsupported opcode in `%s' in VS\n",
898 opcode_descs
[inst
->opcode
].name
);
900 _mesa_problem(ctx
, "Unsupported opcode %d in VS", inst
->opcode
);
907 vec4_generator::generate_code(exec_list
*instructions
)
909 int last_native_insn_offset
= 0;
910 const char *last_annotation_string
= NULL
;
911 const void *last_annotation_ir
= NULL
;
913 if (unlikely(debug_flag
)) {
915 printf("Native code for vertex shader %d:\n", shader_prog
->Name
);
917 printf("Native code for vertex program %d:\n", prog
->Id
);
921 foreach_list(node
, instructions
) {
922 vec4_instruction
*inst
= (vec4_instruction
*)node
;
923 struct brw_reg src
[3], dst
;
925 if (unlikely(debug_flag
)) {
926 if (last_annotation_ir
!= inst
->ir
) {
927 last_annotation_ir
= inst
->ir
;
928 if (last_annotation_ir
) {
931 ((ir_instruction
*) last_annotation_ir
)->print();
933 const prog_instruction
*vpi
;
934 vpi
= (const prog_instruction
*) inst
->ir
;
935 printf("%d: ", (int)(vpi
- prog
->Instructions
));
936 _mesa_fprint_instruction_opt(stdout
, vpi
, 0,
937 PROG_PRINT_DEBUG
, NULL
);
942 if (last_annotation_string
!= inst
->annotation
) {
943 last_annotation_string
= inst
->annotation
;
944 if (last_annotation_string
)
945 printf(" %s\n", last_annotation_string
);
949 for (unsigned int i
= 0; i
< 3; i
++) {
950 src
[i
] = inst
->get_src(this->prog_data
, i
);
952 dst
= inst
->get_dst();
954 brw_set_conditionalmod(p
, inst
->conditional_mod
);
955 brw_set_predicate_control(p
, inst
->predicate
);
956 brw_set_predicate_inverse(p
, inst
->predicate_inverse
);
957 brw_set_saturate(p
, inst
->saturate
);
958 brw_set_mask_control(p
, inst
->force_writemask_all
);
960 unsigned pre_emit_nr_insn
= p
->nr_insn
;
962 generate_vec4_instruction(inst
, dst
, src
);
964 if (inst
->no_dd_clear
|| inst
->no_dd_check
) {
965 assert(p
->nr_insn
== pre_emit_nr_insn
+ 1 ||
966 !"no_dd_check or no_dd_clear set for IR emitting more "
967 "than 1 instruction");
969 struct brw_instruction
*last
= &p
->store
[pre_emit_nr_insn
];
971 if (inst
->no_dd_clear
)
972 last
->header
.dependency_control
|= BRW_DEPENDENCY_NOTCLEARED
;
973 if (inst
->no_dd_check
)
974 last
->header
.dependency_control
|= BRW_DEPENDENCY_NOTCHECKED
;
977 if (unlikely(debug_flag
)) {
978 brw_dump_compile(p
, stdout
,
979 last_native_insn_offset
, p
->next_insn_offset
);
982 last_native_insn_offset
= p
->next_insn_offset
;
985 if (unlikely(debug_flag
)) {
991 /* OK, while the INTEL_DEBUG=vs above is very nice for debugging VS
992 * emit issues, it doesn't get the jump distances into the output,
993 * which is often something we want to debug. So this is here in
994 * case you're doing that.
996 if (0 && unlikely(debug_flag
)) {
997 brw_dump_compile(p
, stdout
, 0, p
->next_insn_offset
);
1002 vec4_generator::generate_assembly(exec_list
*instructions
,
1003 unsigned *assembly_size
)
1005 brw_set_access_mode(p
, BRW_ALIGN_16
);
1006 generate_code(instructions
);
1007 return brw_get_program(p
, assembly_size
);
1010 } /* namespace brw */