2 * Copyright © 2010 Intel Corporation
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
20 * FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS
24 /** @file brw_fs_emit.cpp
26 * This file supports emitting code from the FS LIR to the actual
27 * native instructions.
31 #include "main/macros.h"
32 #include "brw_context.h"
37 #include "glsl/ir_print_visitor.h"
40 fs_visitor::generate_fb_write(fs_inst
*inst
)
42 GLboolean eot
= inst
->eot
;
43 struct brw_reg implied_header
;
45 /* Header is 2 regs, g0 and g1 are the contents. g0 will be implied
48 brw_push_insn_state(p
);
49 brw_set_mask_control(p
, BRW_MASK_DISABLE
);
50 brw_set_compression_control(p
, BRW_COMPRESSION_NONE
);
52 if (inst
->header_present
) {
53 if (intel
->gen
>= 6) {
54 brw_set_compression_control(p
, BRW_COMPRESSION_COMPRESSED
);
56 retype(brw_message_reg(inst
->base_mrf
), BRW_REGISTER_TYPE_UD
),
57 retype(brw_vec8_grf(0, 0), BRW_REGISTER_TYPE_UD
));
58 brw_set_compression_control(p
, BRW_COMPRESSION_NONE
);
60 if (inst
->target
> 0) {
61 /* Set the render target index for choosing BLEND_STATE. */
62 brw_MOV(p
, retype(brw_vec1_reg(BRW_MESSAGE_REGISTER_FILE
,
64 BRW_REGISTER_TYPE_UD
),
65 brw_imm_ud(inst
->target
));
68 implied_header
= brw_null_reg();
70 implied_header
= retype(brw_vec8_grf(0, 0), BRW_REGISTER_TYPE_UW
);
73 brw_message_reg(inst
->base_mrf
+ 1),
77 implied_header
= brw_null_reg();
80 brw_pop_insn_state(p
);
90 inst
->header_present
);
93 /* Computes the integer pixel x,y values from the origin.
95 * This is the basis of gl_FragCoord computation, but is also used
96 * pre-gen6 for computing the deltas from v0 for computing
100 fs_visitor::generate_pixel_xy(struct brw_reg dst
, bool is_x
)
102 struct brw_reg g1_uw
= retype(brw_vec1_grf(1, 0), BRW_REGISTER_TYPE_UW
);
104 struct brw_reg deltas
;
107 src
= stride(suboffset(g1_uw
, 4), 2, 4, 0);
108 deltas
= brw_imm_v(0x10101010);
110 src
= stride(suboffset(g1_uw
, 5), 2, 4, 0);
111 deltas
= brw_imm_v(0x11001100);
114 if (c
->dispatch_width
== 16) {
118 /* We do this 8 or 16-wide, but since the destination is UW we
119 * don't do compression in the 16-wide case.
121 brw_push_insn_state(p
);
122 brw_set_compression_control(p
, BRW_COMPRESSION_NONE
);
123 brw_ADD(p
, dst
, src
, deltas
);
124 brw_pop_insn_state(p
);
128 fs_visitor::generate_linterp(fs_inst
*inst
,
129 struct brw_reg dst
, struct brw_reg
*src
)
131 struct brw_reg delta_x
= src
[0];
132 struct brw_reg delta_y
= src
[1];
133 struct brw_reg interp
= src
[2];
136 delta_y
.nr
== delta_x
.nr
+ 1 &&
137 (intel
->gen
>= 6 || (delta_x
.nr
& 1) == 0)) {
138 brw_PLN(p
, dst
, interp
, delta_x
);
140 brw_LINE(p
, brw_null_reg(), interp
, delta_x
);
141 brw_MAC(p
, dst
, suboffset(interp
, 1), delta_y
);
146 fs_visitor::generate_math(fs_inst
*inst
,
147 struct brw_reg dst
, struct brw_reg
*src
)
149 int op
= brw_math_function(inst
->opcode
);
151 if (intel
->gen
>= 6) {
152 assert(inst
->mlen
== 0);
154 if (inst
->opcode
== SHADER_OPCODE_POW
) {
155 brw_set_compression_control(p
, BRW_COMPRESSION_NONE
);
156 brw_math2(p
, dst
, op
, src
[0], src
[1]);
158 if (c
->dispatch_width
== 16) {
159 brw_set_compression_control(p
, BRW_COMPRESSION_2NDHALF
);
160 brw_math2(p
, sechalf(dst
), op
, sechalf(src
[0]), sechalf(src
[1]));
161 brw_set_compression_control(p
, BRW_COMPRESSION_COMPRESSED
);
164 brw_set_compression_control(p
, BRW_COMPRESSION_NONE
);
167 inst
->saturate
? BRW_MATH_SATURATE_SATURATE
:
168 BRW_MATH_SATURATE_NONE
,
170 BRW_MATH_DATA_VECTOR
,
171 BRW_MATH_PRECISION_FULL
);
173 if (c
->dispatch_width
== 16) {
174 brw_set_compression_control(p
, BRW_COMPRESSION_2NDHALF
);
175 brw_math(p
, sechalf(dst
),
177 inst
->saturate
? BRW_MATH_SATURATE_SATURATE
:
178 BRW_MATH_SATURATE_NONE
,
180 BRW_MATH_DATA_VECTOR
,
181 BRW_MATH_PRECISION_FULL
);
182 brw_set_compression_control(p
, BRW_COMPRESSION_COMPRESSED
);
185 } else /* gen <= 5 */{
186 assert(inst
->mlen
>= 1);
188 brw_set_compression_control(p
, BRW_COMPRESSION_NONE
);
191 inst
->saturate
? BRW_MATH_SATURATE_SATURATE
:
192 BRW_MATH_SATURATE_NONE
,
193 inst
->base_mrf
, src
[0],
194 BRW_MATH_DATA_VECTOR
,
195 BRW_MATH_PRECISION_FULL
);
197 if (c
->dispatch_width
== 16) {
198 brw_set_compression_control(p
, BRW_COMPRESSION_2NDHALF
);
199 brw_math(p
, sechalf(dst
),
201 inst
->saturate
? BRW_MATH_SATURATE_SATURATE
:
202 BRW_MATH_SATURATE_NONE
,
203 inst
->base_mrf
+ 1, sechalf(src
[0]),
204 BRW_MATH_DATA_VECTOR
,
205 BRW_MATH_PRECISION_FULL
);
207 brw_set_compression_control(p
, BRW_COMPRESSION_COMPRESSED
);
213 fs_visitor::generate_tex(fs_inst
*inst
, struct brw_reg dst
, struct brw_reg src
)
217 uint32_t simd_mode
= BRW_SAMPLER_SIMD_MODE_SIMD8
;
219 if (c
->dispatch_width
== 16)
220 simd_mode
= BRW_SAMPLER_SIMD_MODE_SIMD16
;
222 if (intel
->gen
>= 5) {
223 switch (inst
->opcode
) {
225 if (inst
->shadow_compare
) {
226 msg_type
= GEN5_SAMPLER_MESSAGE_SAMPLE_COMPARE
;
228 msg_type
= GEN5_SAMPLER_MESSAGE_SAMPLE
;
232 if (inst
->shadow_compare
) {
233 msg_type
= GEN5_SAMPLER_MESSAGE_SAMPLE_BIAS_COMPARE
;
235 msg_type
= GEN5_SAMPLER_MESSAGE_SAMPLE_BIAS
;
239 if (inst
->shadow_compare
) {
240 msg_type
= GEN5_SAMPLER_MESSAGE_SAMPLE_LOD_COMPARE
;
242 msg_type
= GEN5_SAMPLER_MESSAGE_SAMPLE_LOD
;
246 msg_type
= GEN5_SAMPLER_MESSAGE_SAMPLE_RESINFO
;
249 /* There is no sample_d_c message; comparisons are done manually */
250 msg_type
= GEN5_SAMPLER_MESSAGE_SAMPLE_DERIVS
;
253 msg_type
= GEN5_SAMPLER_MESSAGE_SAMPLE_LD
;
256 assert(!"not reached");
260 switch (inst
->opcode
) {
262 /* Note that G45 and older determines shadow compare and dispatch width
263 * from message length for most messages.
265 assert(c
->dispatch_width
== 8);
266 msg_type
= BRW_SAMPLER_MESSAGE_SIMD8_SAMPLE
;
267 if (inst
->shadow_compare
) {
268 assert(inst
->mlen
== 6);
270 assert(inst
->mlen
<= 4);
274 if (inst
->shadow_compare
) {
275 assert(inst
->mlen
== 6);
276 msg_type
= BRW_SAMPLER_MESSAGE_SIMD8_SAMPLE_BIAS_COMPARE
;
278 assert(inst
->mlen
== 9);
279 msg_type
= BRW_SAMPLER_MESSAGE_SIMD16_SAMPLE_BIAS
;
280 simd_mode
= BRW_SAMPLER_SIMD_MODE_SIMD16
;
284 if (inst
->shadow_compare
) {
285 assert(inst
->mlen
== 6);
286 msg_type
= BRW_SAMPLER_MESSAGE_SIMD8_SAMPLE_LOD_COMPARE
;
288 assert(inst
->mlen
== 9);
289 msg_type
= BRW_SAMPLER_MESSAGE_SIMD16_SAMPLE_LOD
;
290 simd_mode
= BRW_SAMPLER_SIMD_MODE_SIMD16
;
294 /* There is no sample_d_c message; comparisons are done manually */
295 assert(inst
->mlen
== 7 || inst
->mlen
== 10);
296 msg_type
= BRW_SAMPLER_MESSAGE_SIMD8_SAMPLE_GRADIENTS
;
299 assert(inst
->mlen
== 9);
300 msg_type
= BRW_SAMPLER_MESSAGE_SIMD16_LD
;
301 simd_mode
= BRW_SAMPLER_SIMD_MODE_SIMD16
;
304 assert(inst
->mlen
== 3);
305 msg_type
= BRW_SAMPLER_MESSAGE_SIMD16_RESINFO
;
306 simd_mode
= BRW_SAMPLER_SIMD_MODE_SIMD16
;
309 assert(!"not reached");
313 assert(msg_type
!= -1);
315 if (simd_mode
== BRW_SAMPLER_SIMD_MODE_SIMD16
) {
321 retype(dst
, BRW_REGISTER_TYPE_UW
),
324 SURF_INDEX_TEXTURE(inst
->sampler
),
331 inst
->header_present
,
336 /* For OPCODE_DDX and OPCODE_DDY, per channel of output we've got input
339 * arg0: ss0.tl ss0.tr ss0.bl ss0.br ss1.tl ss1.tr ss1.bl ss1.br
341 * and we're trying to produce:
344 * dst: (ss0.tr - ss0.tl) (ss0.tl - ss0.bl)
345 * (ss0.tr - ss0.tl) (ss0.tr - ss0.br)
346 * (ss0.br - ss0.bl) (ss0.tl - ss0.bl)
347 * (ss0.br - ss0.bl) (ss0.tr - ss0.br)
348 * (ss1.tr - ss1.tl) (ss1.tl - ss1.bl)
349 * (ss1.tr - ss1.tl) (ss1.tr - ss1.br)
350 * (ss1.br - ss1.bl) (ss1.tl - ss1.bl)
351 * (ss1.br - ss1.bl) (ss1.tr - ss1.br)
353 * and add another set of two more subspans if in 16-pixel dispatch mode.
355 * For DDX, it ends up being easy: width = 2, horiz=0 gets us the same result
356 * for each pair, and vertstride = 2 jumps us 2 elements after processing a
357 * pair. But for DDY, it's harder, as we want to produce the pairs swizzled
358 * between each other. We could probably do it like ddx and swizzle the right
359 * order later, but bail for now and just produce
360 * ((ss0.tl - ss0.bl)x4 (ss1.tl - ss1.bl)x4)
363 fs_visitor::generate_ddx(fs_inst
*inst
, struct brw_reg dst
, struct brw_reg src
)
365 struct brw_reg src0
= brw_reg(src
.file
, src
.nr
, 1,
367 BRW_VERTICAL_STRIDE_2
,
369 BRW_HORIZONTAL_STRIDE_0
,
370 BRW_SWIZZLE_XYZW
, WRITEMASK_XYZW
);
371 struct brw_reg src1
= brw_reg(src
.file
, src
.nr
, 0,
373 BRW_VERTICAL_STRIDE_2
,
375 BRW_HORIZONTAL_STRIDE_0
,
376 BRW_SWIZZLE_XYZW
, WRITEMASK_XYZW
);
377 brw_ADD(p
, dst
, src0
, negate(src1
));
381 fs_visitor::generate_ddy(fs_inst
*inst
, struct brw_reg dst
, struct brw_reg src
)
383 struct brw_reg src0
= brw_reg(src
.file
, src
.nr
, 0,
385 BRW_VERTICAL_STRIDE_4
,
387 BRW_HORIZONTAL_STRIDE_0
,
388 BRW_SWIZZLE_XYZW
, WRITEMASK_XYZW
);
389 struct brw_reg src1
= brw_reg(src
.file
, src
.nr
, 2,
391 BRW_VERTICAL_STRIDE_4
,
393 BRW_HORIZONTAL_STRIDE_0
,
394 BRW_SWIZZLE_XYZW
, WRITEMASK_XYZW
);
395 brw_ADD(p
, dst
, src0
, negate(src1
));
399 fs_visitor::generate_discard(fs_inst
*inst
)
401 struct brw_reg f0
= brw_flag_reg();
403 if (intel
->gen
>= 6) {
404 struct brw_reg g1
= retype(brw_vec1_grf(1, 7), BRW_REGISTER_TYPE_UW
);
405 struct brw_reg some_register
;
407 /* As of gen6, we no longer have the mask register to look at,
408 * so life gets a bit more complicated.
411 /* Load the flag register with all ones. */
412 brw_push_insn_state(p
);
413 brw_set_mask_control(p
, BRW_MASK_DISABLE
);
414 brw_MOV(p
, f0
, brw_imm_uw(0xffff));
415 brw_pop_insn_state(p
);
417 /* Do a comparison that should always fail, to produce 0s in the flag
418 * reg where we have active channels.
420 some_register
= retype(brw_vec8_grf(0, 0), BRW_REGISTER_TYPE_UW
);
421 brw_CMP(p
, retype(brw_null_reg(), BRW_REGISTER_TYPE_UD
),
422 BRW_CONDITIONAL_NZ
, some_register
, some_register
);
424 /* Undo CMP's whacking of predication*/
425 brw_set_predicate_control(p
, BRW_PREDICATE_NONE
);
427 brw_push_insn_state(p
);
428 brw_set_mask_control(p
, BRW_MASK_DISABLE
);
429 brw_AND(p
, g1
, f0
, g1
);
430 brw_pop_insn_state(p
);
432 struct brw_reg g0
= retype(brw_vec1_grf(0, 0), BRW_REGISTER_TYPE_UW
);
434 brw_push_insn_state(p
);
435 brw_set_mask_control(p
, BRW_MASK_DISABLE
);
436 brw_set_compression_control(p
, BRW_COMPRESSION_NONE
);
438 /* Unlike the 965, we have the mask reg, so we just need
439 * somewhere to invert that (containing channels to be disabled)
440 * so it can be ANDed with the mask of pixels still to be
441 * written. Use the flag reg for consistency with gen6+.
443 brw_NOT(p
, f0
, brw_mask_reg(1)); /* IMASK */
444 brw_AND(p
, g0
, f0
, g0
);
446 brw_pop_insn_state(p
);
451 fs_visitor::generate_spill(fs_inst
*inst
, struct brw_reg src
)
453 assert(inst
->mlen
!= 0);
456 retype(brw_message_reg(inst
->base_mrf
+ 1), BRW_REGISTER_TYPE_UD
),
457 retype(src
, BRW_REGISTER_TYPE_UD
));
458 brw_oword_block_write_scratch(p
, brw_message_reg(inst
->base_mrf
), 1,
463 fs_visitor::generate_unspill(fs_inst
*inst
, struct brw_reg dst
)
465 assert(inst
->mlen
!= 0);
467 /* Clear any post destination dependencies that would be ignored by
468 * the block read. See the B-Spec for pre-gen5 send instruction.
470 * This could use a better solution, since texture sampling and
471 * math reads could potentially run into it as well -- anywhere
472 * that we have a SEND with a destination that is a register that
473 * was written but not read within the last N instructions (what's
474 * N? unsure). This is rare because of dead code elimination, but
477 if (intel
->gen
== 4 && !intel
->is_g4x
)
478 brw_MOV(p
, brw_null_reg(), dst
);
480 brw_oword_block_read_scratch(p
, dst
, brw_message_reg(inst
->base_mrf
), 1,
483 if (intel
->gen
== 4 && !intel
->is_g4x
) {
484 /* gen4 errata: destination from a send can't be used as a
485 * destination until it's been read. Just read it so we don't
488 brw_MOV(p
, brw_null_reg(), dst
);
493 fs_visitor::generate_pull_constant_load(fs_inst
*inst
, struct brw_reg dst
)
495 assert(inst
->mlen
!= 0);
497 /* Clear any post destination dependencies that would be ignored by
498 * the block read. See the B-Spec for pre-gen5 send instruction.
500 * This could use a better solution, since texture sampling and
501 * math reads could potentially run into it as well -- anywhere
502 * that we have a SEND with a destination that is a register that
503 * was written but not read within the last N instructions (what's
504 * N? unsure). This is rare because of dead code elimination, but
507 if (intel
->gen
== 4 && !intel
->is_g4x
)
508 brw_MOV(p
, brw_null_reg(), dst
);
510 brw_oword_block_read(p
, dst
, brw_message_reg(inst
->base_mrf
),
511 inst
->offset
, SURF_INDEX_FRAG_CONST_BUFFER
);
513 if (intel
->gen
== 4 && !intel
->is_g4x
) {
514 /* gen4 errata: destination from a send can't be used as a
515 * destination until it's been read. Just read it so we don't
518 brw_MOV(p
, brw_null_reg(), dst
);
522 static struct brw_reg
523 brw_reg_from_fs_reg(fs_reg
*reg
)
525 struct brw_reg brw_reg
;
531 if (reg
->smear
== -1) {
532 brw_reg
= brw_vec8_reg(reg
->file
, reg
->reg
, 0);
534 brw_reg
= brw_vec1_reg(reg
->file
, reg
->reg
, reg
->smear
);
536 brw_reg
= retype(brw_reg
, reg
->type
);
538 brw_reg
= sechalf(brw_reg
);
542 case BRW_REGISTER_TYPE_F
:
543 brw_reg
= brw_imm_f(reg
->imm
.f
);
545 case BRW_REGISTER_TYPE_D
:
546 brw_reg
= brw_imm_d(reg
->imm
.i
);
548 case BRW_REGISTER_TYPE_UD
:
549 brw_reg
= brw_imm_ud(reg
->imm
.u
);
552 assert(!"not reached");
553 brw_reg
= brw_null_reg();
558 brw_reg
= reg
->fixed_hw_reg
;
561 /* Probably unused. */
562 brw_reg
= brw_null_reg();
565 assert(!"not reached");
566 brw_reg
= brw_null_reg();
569 assert(!"not reached");
570 brw_reg
= brw_null_reg();
574 brw_reg
= brw_abs(brw_reg
);
576 brw_reg
= negate(brw_reg
);
582 fs_visitor::generate_code()
584 int last_native_inst
= p
->nr_insn
;
585 const char *last_annotation_string
= NULL
;
586 ir_instruction
*last_annotation_ir
= NULL
;
588 int loop_stack_array_size
= 16;
589 int loop_stack_depth
= 0;
590 brw_instruction
**loop_stack
=
591 rzalloc_array(this->mem_ctx
, brw_instruction
*, loop_stack_array_size
);
592 int *if_depth_in_loop
=
593 rzalloc_array(this->mem_ctx
, int, loop_stack_array_size
);
596 if (unlikely(INTEL_DEBUG
& DEBUG_WM
)) {
597 printf("Native code for fragment shader %d (%d-wide dispatch):\n",
598 prog
->Name
, c
->dispatch_width
);
601 foreach_list(node
, &this->instructions
) {
602 fs_inst
*inst
= (fs_inst
*)node
;
603 struct brw_reg src
[3], dst
;
605 if (unlikely(INTEL_DEBUG
& DEBUG_WM
)) {
606 if (last_annotation_ir
!= inst
->ir
) {
607 last_annotation_ir
= inst
->ir
;
608 if (last_annotation_ir
) {
610 last_annotation_ir
->print();
614 if (last_annotation_string
!= inst
->annotation
) {
615 last_annotation_string
= inst
->annotation
;
616 if (last_annotation_string
)
617 printf(" %s\n", last_annotation_string
);
621 for (unsigned int i
= 0; i
< 3; i
++) {
622 src
[i
] = brw_reg_from_fs_reg(&inst
->src
[i
]);
624 dst
= brw_reg_from_fs_reg(&inst
->dst
);
626 brw_set_conditionalmod(p
, inst
->conditional_mod
);
627 brw_set_predicate_control(p
, inst
->predicated
);
628 brw_set_predicate_inverse(p
, inst
->predicate_inverse
);
629 brw_set_saturate(p
, inst
->saturate
);
631 if (inst
->force_uncompressed
|| c
->dispatch_width
== 8) {
632 brw_set_compression_control(p
, BRW_COMPRESSION_NONE
);
633 } else if (inst
->force_sechalf
) {
634 brw_set_compression_control(p
, BRW_COMPRESSION_2NDHALF
);
636 brw_set_compression_control(p
, BRW_COMPRESSION_COMPRESSED
);
639 switch (inst
->opcode
) {
641 brw_MOV(p
, dst
, src
[0]);
644 brw_ADD(p
, dst
, src
[0], src
[1]);
647 brw_MUL(p
, dst
, src
[0], src
[1]);
649 case BRW_OPCODE_MACH
:
650 brw_set_acc_write_control(p
, 1);
651 brw_MACH(p
, dst
, src
[0], src
[1]);
652 brw_set_acc_write_control(p
, 0);
656 brw_FRC(p
, dst
, src
[0]);
658 case BRW_OPCODE_RNDD
:
659 brw_RNDD(p
, dst
, src
[0]);
661 case BRW_OPCODE_RNDE
:
662 brw_RNDE(p
, dst
, src
[0]);
664 case BRW_OPCODE_RNDZ
:
665 brw_RNDZ(p
, dst
, src
[0]);
669 brw_AND(p
, dst
, src
[0], src
[1]);
672 brw_OR(p
, dst
, src
[0], src
[1]);
675 brw_XOR(p
, dst
, src
[0], src
[1]);
678 brw_NOT(p
, dst
, src
[0]);
681 brw_ASR(p
, dst
, src
[0], src
[1]);
684 brw_SHR(p
, dst
, src
[0], src
[1]);
687 brw_SHL(p
, dst
, src
[0], src
[1]);
691 brw_CMP(p
, dst
, inst
->conditional_mod
, src
[0], src
[1]);
694 brw_SEL(p
, dst
, src
[0], src
[1]);
698 if (inst
->src
[0].file
!= BAD_FILE
) {
699 /* The instruction has an embedded compare (only allowed on gen6) */
700 assert(intel
->gen
== 6);
701 gen6_IF(p
, inst
->conditional_mod
, src
[0], src
[1]);
703 brw_IF(p
, c
->dispatch_width
== 16 ? BRW_EXECUTE_16
: BRW_EXECUTE_8
);
705 if_depth_in_loop
[loop_stack_depth
]++;
708 case BRW_OPCODE_ELSE
:
711 case BRW_OPCODE_ENDIF
:
713 if_depth_in_loop
[loop_stack_depth
]--;
717 loop_stack
[loop_stack_depth
++] = brw_DO(p
, BRW_EXECUTE_8
);
718 if (loop_stack_array_size
<= loop_stack_depth
) {
719 loop_stack_array_size
*= 2;
720 loop_stack
= reralloc(this->mem_ctx
, loop_stack
, brw_instruction
*,
721 loop_stack_array_size
);
722 if_depth_in_loop
= reralloc(this->mem_ctx
, if_depth_in_loop
, int,
723 loop_stack_array_size
);
725 if_depth_in_loop
[loop_stack_depth
] = 0;
728 case BRW_OPCODE_BREAK
:
729 brw_BREAK(p
, if_depth_in_loop
[loop_stack_depth
]);
730 brw_set_predicate_control(p
, BRW_PREDICATE_NONE
);
732 case BRW_OPCODE_CONTINUE
:
733 /* FINISHME: We need to write the loop instruction support still. */
735 gen6_CONT(p
, loop_stack
[loop_stack_depth
- 1]);
737 brw_CONT(p
, if_depth_in_loop
[loop_stack_depth
]);
738 brw_set_predicate_control(p
, BRW_PREDICATE_NONE
);
741 case BRW_OPCODE_WHILE
: {
742 struct brw_instruction
*inst0
, *inst1
;
748 assert(loop_stack_depth
> 0);
750 inst0
= inst1
= brw_WHILE(p
, loop_stack
[loop_stack_depth
]);
751 if (intel
->gen
< 6) {
752 /* patch all the BREAK/CONT instructions from last BGNLOOP */
753 while (inst0
> loop_stack
[loop_stack_depth
]) {
755 if (inst0
->header
.opcode
== BRW_OPCODE_BREAK
&&
756 inst0
->bits3
.if_else
.jump_count
== 0) {
757 inst0
->bits3
.if_else
.jump_count
= br
* (inst1
- inst0
+ 1);
759 else if (inst0
->header
.opcode
== BRW_OPCODE_CONTINUE
&&
760 inst0
->bits3
.if_else
.jump_count
== 0) {
761 inst0
->bits3
.if_else
.jump_count
= br
* (inst1
- inst0
);
768 case SHADER_OPCODE_RCP
:
769 case SHADER_OPCODE_RSQ
:
770 case SHADER_OPCODE_SQRT
:
771 case SHADER_OPCODE_EXP2
:
772 case SHADER_OPCODE_LOG2
:
773 case SHADER_OPCODE_POW
:
774 case SHADER_OPCODE_SIN
:
775 case SHADER_OPCODE_COS
:
776 generate_math(inst
, dst
, src
);
778 case FS_OPCODE_PIXEL_X
:
779 generate_pixel_xy(dst
, true);
781 case FS_OPCODE_PIXEL_Y
:
782 generate_pixel_xy(dst
, false);
784 case FS_OPCODE_CINTERP
:
785 brw_MOV(p
, dst
, src
[0]);
787 case FS_OPCODE_LINTERP
:
788 generate_linterp(inst
, dst
, src
);
796 generate_tex(inst
, dst
, src
[0]);
798 case FS_OPCODE_DISCARD
:
799 generate_discard(inst
);
802 generate_ddx(inst
, dst
, src
[0]);
805 generate_ddy(inst
, dst
, src
[0]);
808 case FS_OPCODE_SPILL
:
809 generate_spill(inst
, src
[0]);
812 case FS_OPCODE_UNSPILL
:
813 generate_unspill(inst
, dst
);
816 case FS_OPCODE_PULL_CONSTANT_LOAD
:
817 generate_pull_constant_load(inst
, dst
);
820 case FS_OPCODE_FB_WRITE
:
821 generate_fb_write(inst
);
824 if (inst
->opcode
< (int)ARRAY_SIZE(brw_opcodes
)) {
825 _mesa_problem(ctx
, "Unsupported opcode `%s' in FS",
826 brw_opcodes
[inst
->opcode
].name
);
828 _mesa_problem(ctx
, "Unsupported opcode %d in FS", inst
->opcode
);
830 fail("unsupported opcode in FS\n");
833 if (unlikely(INTEL_DEBUG
& DEBUG_WM
)) {
834 for (unsigned int i
= last_native_inst
; i
< p
->nr_insn
; i
++) {
836 printf("0x%08x 0x%08x 0x%08x 0x%08x ",
837 ((uint32_t *)&p
->store
[i
])[3],
838 ((uint32_t *)&p
->store
[i
])[2],
839 ((uint32_t *)&p
->store
[i
])[1],
840 ((uint32_t *)&p
->store
[i
])[0]);
842 brw_disasm(stdout
, &p
->store
[i
], intel
->gen
);
846 last_native_inst
= p
->nr_insn
;
849 if (unlikely(INTEL_DEBUG
& DEBUG_WM
)) {
853 ralloc_free(loop_stack
);
854 ralloc_free(if_depth_in_loop
);
858 /* OK, while the INTEL_DEBUG=wm above is very nice for debugging FS
859 * emit issues, it doesn't get the jump distances into the output,
860 * which is often something we want to debug. So this is here in
861 * case you're doing that.
864 if (unlikely(INTEL_DEBUG
& DEBUG_WM
)) {
865 for (unsigned int i
= 0; i
< p
->nr_insn
; i
++) {
866 printf("0x%08x 0x%08x 0x%08x 0x%08x ",
867 ((uint32_t *)&p
->store
[i
])[3],
868 ((uint32_t *)&p
->store
[i
])[2],
869 ((uint32_t *)&p
->store
[i
])[1],
870 ((uint32_t *)&p
->store
[i
])[0]);
871 brw_disasm(stdout
, &p
->store
[i
], intel
->gen
);