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
26 * This file drives the GLSL IR -> LIR translation, contains the
27 * optimizations on the LIR, and drives the generation of native code
33 #include <sys/types.h>
35 #include "main/macros.h"
36 #include "main/shaderobj.h"
37 #include "main/uniforms.h"
38 #include "main/fbobject.h"
39 #include "program/prog_parameter.h"
40 #include "program/prog_print.h"
41 #include "program/register_allocate.h"
42 #include "program/sampler.h"
43 #include "program/hash_table.h"
44 #include "brw_context.h"
49 #include "glsl/glsl_types.h"
50 #include "glsl/ir_print_visitor.h"
55 memset(this, 0, sizeof(*this));
56 this->opcode = BRW_OPCODE_NOP;
57 this->conditional_mod = BRW_CONDITIONAL_NONE;
59 this->dst = reg_undef;
60 this->src[0] = reg_undef;
61 this->src[1] = reg_undef;
62 this->src[2] = reg_undef;
70 fs_inst::fs_inst(enum opcode opcode)
73 this->opcode = opcode;
76 fs_inst::fs_inst(enum opcode opcode, fs_reg dst)
79 this->opcode = opcode;
83 assert(dst.reg_offset >= 0);
86 fs_inst::fs_inst(enum opcode opcode, fs_reg dst, fs_reg src0)
89 this->opcode = opcode;
94 assert(dst.reg_offset >= 0);
95 if (src[0].file == GRF)
96 assert(src[0].reg_offset >= 0);
99 fs_inst::fs_inst(enum opcode opcode, fs_reg dst, fs_reg src0, fs_reg src1)
102 this->opcode = opcode;
108 assert(dst.reg_offset >= 0);
109 if (src[0].file == GRF)
110 assert(src[0].reg_offset >= 0);
111 if (src[1].file == GRF)
112 assert(src[1].reg_offset >= 0);
115 fs_inst::fs_inst(enum opcode opcode, fs_reg dst,
116 fs_reg src0, fs_reg src1, fs_reg src2)
119 this->opcode = opcode;
126 assert(dst.reg_offset >= 0);
127 if (src[0].file == GRF)
128 assert(src[0].reg_offset >= 0);
129 if (src[1].file == GRF)
130 assert(src[1].reg_offset >= 0);
131 if (src[2].file == GRF)
132 assert(src[2].reg_offset >= 0);
137 fs_visitor::op(fs_reg dst, fs_reg src0) \
139 return new(mem_ctx) fs_inst(BRW_OPCODE_##op, dst, src0); \
144 fs_visitor::op(fs_reg dst, fs_reg src0, fs_reg src1) \
146 return new(mem_ctx) fs_inst(BRW_OPCODE_##op, dst, src0, src1); \
165 /** Gen4 predicated IF. */
167 fs_visitor::IF(uint32_t predicate)
169 fs_inst *inst = new(mem_ctx) fs_inst(BRW_OPCODE_IF);
170 inst->predicate = predicate;
174 /** Gen6+ IF with embedded comparison. */
176 fs_visitor::IF(fs_reg src0, fs_reg src1, uint32_t condition)
178 assert(intel->gen >= 6);
179 fs_inst *inst = new(mem_ctx) fs_inst(BRW_OPCODE_IF,
180 reg_null_d, src0, src1);
181 inst->conditional_mod = condition;
186 * CMP: Sets the low bit of the destination channels with the result
187 * of the comparison, while the upper bits are undefined, and updates
188 * the flag register with the packed 16 bits of the result.
191 fs_visitor::CMP(fs_reg dst, fs_reg src0, fs_reg src1, uint32_t condition)
195 /* Take the instruction:
197 * CMP null<d> src0<f> src1<f>
199 * Original gen4 does type conversion to the destination type before
200 * comparison, producing garbage results for floating point comparisons.
201 * gen5 does the comparison on the execution type (resolved source types),
202 * so dst type doesn't matter. gen6 does comparison and then uses the
203 * result as if it was the dst type with no conversion, which happens to
204 * mostly work out for float-interpreted-as-int since our comparisons are
207 if (intel->gen == 4) {
208 dst.type = src0.type;
209 if (dst.file == FIXED_HW_REG)
210 dst.fixed_hw_reg.type = dst.type;
213 resolve_ud_negate(&src0);
214 resolve_ud_negate(&src1);
216 inst = new(mem_ctx) fs_inst(BRW_OPCODE_CMP, dst, src0, src1);
217 inst->conditional_mod = condition;
223 fs_inst::equals(fs_inst *inst)
225 return (opcode == inst->opcode &&
226 dst.equals(inst->dst) &&
227 src[0].equals(inst->src[0]) &&
228 src[1].equals(inst->src[1]) &&
229 src[2].equals(inst->src[2]) &&
230 saturate == inst->saturate &&
231 predicate == inst->predicate &&
232 conditional_mod == inst->conditional_mod &&
233 mlen == inst->mlen &&
234 base_mrf == inst->base_mrf &&
235 sampler == inst->sampler &&
236 target == inst->target &&
238 header_present == inst->header_present &&
239 shadow_compare == inst->shadow_compare &&
240 offset == inst->offset);
244 fs_inst::regs_written()
249 /* The SINCOS and INT_DIV_QUOTIENT_AND_REMAINDER math functions return 2,
250 * but we don't currently use them...nor do we have an opcode for them.
257 fs_inst::overwrites_reg(const fs_reg ®)
259 return (reg.file == dst.file &&
260 reg.reg == dst.reg &&
261 reg.reg_offset >= dst.reg_offset &&
262 reg.reg_offset < dst.reg_offset + regs_written());
268 return (opcode == SHADER_OPCODE_TEX ||
269 opcode == FS_OPCODE_TXB ||
270 opcode == SHADER_OPCODE_TXD ||
271 opcode == SHADER_OPCODE_TXF ||
272 opcode == SHADER_OPCODE_TXL ||
273 opcode == SHADER_OPCODE_TXS);
279 return (opcode == SHADER_OPCODE_RCP ||
280 opcode == SHADER_OPCODE_RSQ ||
281 opcode == SHADER_OPCODE_SQRT ||
282 opcode == SHADER_OPCODE_EXP2 ||
283 opcode == SHADER_OPCODE_LOG2 ||
284 opcode == SHADER_OPCODE_SIN ||
285 opcode == SHADER_OPCODE_COS ||
286 opcode == SHADER_OPCODE_INT_QUOTIENT ||
287 opcode == SHADER_OPCODE_INT_REMAINDER ||
288 opcode == SHADER_OPCODE_POW);
294 memset(this, 0, sizeof(*this));
298 /** Generic unset register constructor. */
302 this->file = BAD_FILE;
305 /** Immediate value constructor. */
306 fs_reg::fs_reg(float f)
310 this->type = BRW_REGISTER_TYPE_F;
314 /** Immediate value constructor. */
315 fs_reg::fs_reg(int32_t i)
319 this->type = BRW_REGISTER_TYPE_D;
323 /** Immediate value constructor. */
324 fs_reg::fs_reg(uint32_t u)
328 this->type = BRW_REGISTER_TYPE_UD;
332 /** Fixed brw_reg Immediate value constructor. */
333 fs_reg::fs_reg(struct brw_reg fixed_hw_reg)
336 this->file = FIXED_HW_REG;
337 this->fixed_hw_reg = fixed_hw_reg;
338 this->type = fixed_hw_reg.type;
342 fs_reg::equals(const fs_reg &r) const
344 return (file == r.file &&
346 reg_offset == r.reg_offset &&
348 negate == r.negate &&
350 memcmp(&fixed_hw_reg, &r.fixed_hw_reg,
351 sizeof(fixed_hw_reg)) == 0 &&
357 fs_reg::is_zero() const
362 return type == BRW_REGISTER_TYPE_F ? imm.f == 0.0 : imm.i == 0;
366 fs_reg::is_one() const
371 return type == BRW_REGISTER_TYPE_F ? imm.f == 1.0 : imm.i == 1;
375 fs_visitor::type_size(const struct glsl_type *type)
377 unsigned int size, i;
379 switch (type->base_type) {
382 case GLSL_TYPE_FLOAT:
384 return type->components();
385 case GLSL_TYPE_ARRAY:
386 return type_size(type->fields.array) * type->length;
387 case GLSL_TYPE_STRUCT:
389 for (i = 0; i < type->length; i++) {
390 size += type_size(type->fields.structure[i].type);
393 case GLSL_TYPE_SAMPLER:
394 /* Samplers take up no register space, since they're baked in at
399 assert(!"not reached");
405 fs_visitor::fail(const char *format, ...)
415 va_start(va, format);
416 msg = ralloc_vasprintf(mem_ctx, format, va);
418 msg = ralloc_asprintf(mem_ctx, "FS compile failed: %s\n", msg);
420 this->fail_msg = msg;
422 if (INTEL_DEBUG & DEBUG_WM) {
423 fprintf(stderr, "%s", msg);
428 fs_visitor::emit(enum opcode opcode)
430 return emit(fs_inst(opcode));
434 fs_visitor::emit(enum opcode opcode, fs_reg dst)
436 return emit(fs_inst(opcode, dst));
440 fs_visitor::emit(enum opcode opcode, fs_reg dst, fs_reg src0)
442 return emit(fs_inst(opcode, dst, src0));
446 fs_visitor::emit(enum opcode opcode, fs_reg dst, fs_reg src0, fs_reg src1)
448 return emit(fs_inst(opcode, dst, src0, src1));
452 fs_visitor::emit(enum opcode opcode, fs_reg dst,
453 fs_reg src0, fs_reg src1, fs_reg src2)
455 return emit(fs_inst(opcode, dst, src0, src1, src2));
459 fs_visitor::push_force_uncompressed()
461 force_uncompressed_stack++;
465 fs_visitor::pop_force_uncompressed()
467 force_uncompressed_stack--;
468 assert(force_uncompressed_stack >= 0);
472 fs_visitor::push_force_sechalf()
474 force_sechalf_stack++;
478 fs_visitor::pop_force_sechalf()
480 force_sechalf_stack--;
481 assert(force_sechalf_stack >= 0);
485 * Returns how many MRFs an FS opcode will write over.
487 * Note that this is not the 0 or 1 implied writes in an actual gen
488 * instruction -- the FS opcodes often generate MOVs in addition.
491 fs_visitor::implied_mrf_writes(fs_inst *inst)
496 switch (inst->opcode) {
497 case SHADER_OPCODE_RCP:
498 case SHADER_OPCODE_RSQ:
499 case SHADER_OPCODE_SQRT:
500 case SHADER_OPCODE_EXP2:
501 case SHADER_OPCODE_LOG2:
502 case SHADER_OPCODE_SIN:
503 case SHADER_OPCODE_COS:
504 return 1 * dispatch_width / 8;
505 case SHADER_OPCODE_POW:
506 case SHADER_OPCODE_INT_QUOTIENT:
507 case SHADER_OPCODE_INT_REMAINDER:
508 return 2 * dispatch_width / 8;
509 case SHADER_OPCODE_TEX:
511 case SHADER_OPCODE_TXD:
512 case SHADER_OPCODE_TXF:
513 case SHADER_OPCODE_TXL:
514 case SHADER_OPCODE_TXS:
516 case FS_OPCODE_FB_WRITE:
518 case FS_OPCODE_PULL_CONSTANT_LOAD:
519 case FS_OPCODE_UNSPILL:
521 case FS_OPCODE_SPILL:
524 assert(!"not reached");
530 fs_visitor::virtual_grf_alloc(int size)
532 if (virtual_grf_array_size <= virtual_grf_count) {
533 if (virtual_grf_array_size == 0)
534 virtual_grf_array_size = 16;
536 virtual_grf_array_size *= 2;
537 virtual_grf_sizes = reralloc(mem_ctx, virtual_grf_sizes, int,
538 virtual_grf_array_size);
540 virtual_grf_sizes[virtual_grf_count] = size;
541 return virtual_grf_count++;
544 /** Fixed HW reg constructor. */
545 fs_reg::fs_reg(enum register_file file, int reg)
550 this->type = BRW_REGISTER_TYPE_F;
553 /** Fixed HW reg constructor. */
554 fs_reg::fs_reg(enum register_file file, int reg, uint32_t type)
562 /** Automatic reg constructor. */
563 fs_reg::fs_reg(class fs_visitor *v, const struct glsl_type *type)
568 this->reg = v->virtual_grf_alloc(v->type_size(type));
569 this->reg_offset = 0;
570 this->type = brw_type_for_base_type(type);
574 fs_visitor::variable_storage(ir_variable *var)
576 return (fs_reg *)hash_table_find(this->variable_ht, var);
580 import_uniforms_callback(const void *key,
584 struct hash_table *dst_ht = (struct hash_table *)closure;
585 const fs_reg *reg = (const fs_reg *)data;
587 if (reg->file != UNIFORM)
590 hash_table_insert(dst_ht, data, key);
593 /* For 16-wide, we need to follow from the uniform setup of 8-wide dispatch.
594 * This brings in those uniform definitions
597 fs_visitor::import_uniforms(fs_visitor *v)
599 hash_table_call_foreach(v->variable_ht,
600 import_uniforms_callback,
602 this->params_remap = v->params_remap;
605 /* Our support for uniforms is piggy-backed on the struct
606 * gl_fragment_program, because that's where the values actually
607 * get stored, rather than in some global gl_shader_program uniform
611 fs_visitor::setup_uniform_values(int loc, const glsl_type *type)
613 unsigned int offset = 0;
615 if (type->is_matrix()) {
616 const glsl_type *column = glsl_type::get_instance(GLSL_TYPE_FLOAT,
617 type->vector_elements,
620 for (unsigned int i = 0; i < type->matrix_columns; i++) {
621 offset += setup_uniform_values(loc + offset, column);
627 switch (type->base_type) {
628 case GLSL_TYPE_FLOAT:
632 for (unsigned int i = 0; i < type->vector_elements; i++) {
633 unsigned int param = c->prog_data.nr_params++;
635 this->param_index[param] = loc;
636 this->param_offset[param] = i;
640 case GLSL_TYPE_STRUCT:
641 for (unsigned int i = 0; i < type->length; i++) {
642 offset += setup_uniform_values(loc + offset,
643 type->fields.structure[i].type);
647 case GLSL_TYPE_ARRAY:
648 for (unsigned int i = 0; i < type->length; i++) {
649 offset += setup_uniform_values(loc + offset, type->fields.array);
653 case GLSL_TYPE_SAMPLER:
654 /* The sampler takes up a slot, but we don't use any values from it. */
658 assert(!"not reached");
664 /* Our support for builtin uniforms is even scarier than non-builtin.
665 * It sits on top of the PROG_STATE_VAR parameters that are
666 * automatically updated from GL context state.
669 fs_visitor::setup_builtin_uniform_values(ir_variable *ir)
671 const ir_state_slot *const slots = ir->state_slots;
672 assert(ir->state_slots != NULL);
674 for (unsigned int i = 0; i < ir->num_state_slots; i++) {
675 /* This state reference has already been setup by ir_to_mesa, but we'll
676 * get the same index back here.
678 int index = _mesa_add_state_reference(this->fp->Base.Parameters,
679 (gl_state_index *)slots[i].tokens);
681 /* Add each of the unique swizzles of the element as a parameter.
682 * This'll end up matching the expected layout of the
683 * array/matrix/structure we're trying to fill in.
686 for (unsigned int j = 0; j < 4; j++) {
687 int swiz = GET_SWZ(slots[i].swizzle, j);
688 if (swiz == last_swiz)
692 this->param_index[c->prog_data.nr_params] = index;
693 this->param_offset[c->prog_data.nr_params] = swiz;
694 c->prog_data.nr_params++;
700 fs_visitor::emit_fragcoord_interpolation(ir_variable *ir)
702 fs_reg *reg = new(this->mem_ctx) fs_reg(this, ir->type);
704 bool flip = !ir->origin_upper_left ^ c->key.render_to_fbo;
707 if (ir->pixel_center_integer) {
708 emit(MOV(wpos, this->pixel_x));
710 emit(ADD(wpos, this->pixel_x, fs_reg(0.5f)));
715 if (!flip && ir->pixel_center_integer) {
716 emit(MOV(wpos, this->pixel_y));
718 fs_reg pixel_y = this->pixel_y;
719 float offset = (ir->pixel_center_integer ? 0.0 : 0.5);
722 pixel_y.negate = true;
723 offset += c->key.drawable_height - 1.0;
726 emit(ADD(wpos, pixel_y, fs_reg(offset)));
731 if (intel->gen >= 6) {
732 emit(MOV(wpos, fs_reg(brw_vec8_grf(c->source_depth_reg, 0))));
734 emit(FS_OPCODE_LINTERP, wpos,
735 this->delta_x[BRW_WM_PERSPECTIVE_PIXEL_BARYCENTRIC],
736 this->delta_y[BRW_WM_PERSPECTIVE_PIXEL_BARYCENTRIC],
737 interp_reg(FRAG_ATTRIB_WPOS, 2));
741 /* gl_FragCoord.w: Already set up in emit_interpolation */
742 emit(BRW_OPCODE_MOV, wpos, this->wpos_w);
748 fs_visitor::emit_linterp(const fs_reg &attr, const fs_reg &interp,
749 glsl_interp_qualifier interpolation_mode,
752 brw_wm_barycentric_interp_mode barycoord_mode;
754 if (interpolation_mode == INTERP_QUALIFIER_SMOOTH)
755 barycoord_mode = BRW_WM_PERSPECTIVE_CENTROID_BARYCENTRIC;
757 barycoord_mode = BRW_WM_NONPERSPECTIVE_CENTROID_BARYCENTRIC;
759 if (interpolation_mode == INTERP_QUALIFIER_SMOOTH)
760 barycoord_mode = BRW_WM_PERSPECTIVE_PIXEL_BARYCENTRIC;
762 barycoord_mode = BRW_WM_NONPERSPECTIVE_PIXEL_BARYCENTRIC;
764 return emit(FS_OPCODE_LINTERP, attr,
765 this->delta_x[barycoord_mode],
766 this->delta_y[barycoord_mode], interp);
770 fs_visitor::emit_general_interpolation(ir_variable *ir)
772 fs_reg *reg = new(this->mem_ctx) fs_reg(this, ir->type);
773 reg->type = brw_type_for_base_type(ir->type->get_scalar_type());
776 unsigned int array_elements;
777 const glsl_type *type;
779 if (ir->type->is_array()) {
780 array_elements = ir->type->length;
781 if (array_elements == 0) {
782 fail("dereferenced array '%s' has length 0\n", ir->name);
784 type = ir->type->fields.array;
790 glsl_interp_qualifier interpolation_mode =
791 ir->determine_interpolation_mode(c->key.flat_shade);
793 int location = ir->location;
794 for (unsigned int i = 0; i < array_elements; i++) {
795 for (unsigned int j = 0; j < type->matrix_columns; j++) {
796 if (urb_setup[location] == -1) {
797 /* If there's no incoming setup data for this slot, don't
798 * emit interpolation for it.
800 attr.reg_offset += type->vector_elements;
805 if (interpolation_mode == INTERP_QUALIFIER_FLAT) {
806 /* Constant interpolation (flat shading) case. The SF has
807 * handed us defined values in only the constant offset
808 * field of the setup reg.
810 for (unsigned int k = 0; k < type->vector_elements; k++) {
811 struct brw_reg interp = interp_reg(location, k);
812 interp = suboffset(interp, 3);
813 interp.type = reg->type;
814 emit(FS_OPCODE_CINTERP, attr, fs_reg(interp));
818 /* Smooth/noperspective interpolation case. */
819 for (unsigned int k = 0; k < type->vector_elements; k++) {
820 /* FINISHME: At some point we probably want to push
821 * this farther by giving similar treatment to the
822 * other potentially constant components of the
823 * attribute, as well as making brw_vs_constval.c
824 * handle varyings other than gl_TexCoord.
826 if (location >= FRAG_ATTRIB_TEX0 &&
827 location <= FRAG_ATTRIB_TEX7 &&
828 k == 3 && !(c->key.proj_attrib_mask & (1 << location))) {
829 emit(BRW_OPCODE_MOV, attr, fs_reg(1.0f));
831 struct brw_reg interp = interp_reg(location, k);
832 emit_linterp(attr, fs_reg(interp), interpolation_mode,
834 if (brw->needs_unlit_centroid_workaround && ir->centroid) {
835 /* Get the pixel/sample mask into f0 so that we know
836 * which pixels are lit. Then, for each channel that is
837 * unlit, replace the centroid data with non-centroid
840 emit(FS_OPCODE_MOV_DISPATCH_TO_FLAGS, attr);
841 fs_inst *inst = emit_linterp(attr, fs_reg(interp),
842 interpolation_mode, false);
843 inst->predicate = BRW_PREDICATE_NORMAL;
844 inst->predicate_inverse = true;
846 if (intel->gen < 6) {
847 emit(BRW_OPCODE_MUL, attr, attr, this->pixel_w);
862 fs_visitor::emit_frontfacing_interpolation(ir_variable *ir)
864 fs_reg *reg = new(this->mem_ctx) fs_reg(this, ir->type);
866 /* The frontfacing comes in as a bit in the thread payload. */
867 if (intel->gen >= 6) {
868 emit(BRW_OPCODE_ASR, *reg,
869 fs_reg(retype(brw_vec1_grf(0, 0), BRW_REGISTER_TYPE_D)),
871 emit(BRW_OPCODE_NOT, *reg, *reg);
872 emit(BRW_OPCODE_AND, *reg, *reg, fs_reg(1));
874 struct brw_reg r1_6ud = retype(brw_vec1_grf(1, 6), BRW_REGISTER_TYPE_UD);
875 /* bit 31 is "primitive is back face", so checking < (1 << 31) gives
878 emit(CMP(*reg, fs_reg(r1_6ud), fs_reg(1u << 31), BRW_CONDITIONAL_L));
879 emit(BRW_OPCODE_AND, *reg, *reg, fs_reg(1u));
886 fs_visitor::emit_math(enum opcode opcode, fs_reg dst, fs_reg src)
889 case SHADER_OPCODE_RCP:
890 case SHADER_OPCODE_RSQ:
891 case SHADER_OPCODE_SQRT:
892 case SHADER_OPCODE_EXP2:
893 case SHADER_OPCODE_LOG2:
894 case SHADER_OPCODE_SIN:
895 case SHADER_OPCODE_COS:
898 assert(!"not reached: bad math opcode");
902 /* Can't do hstride == 0 args to gen6 math, so expand it out. We
903 * might be able to do better by doing execsize = 1 math and then
904 * expanding that result out, but we would need to be careful with
907 * Gen 6 hardware ignores source modifiers (negate and abs) on math
908 * instructions, so we also move to a temp to set those up.
910 if (intel->gen == 6 && (src.file == UNIFORM ||
913 fs_reg expanded = fs_reg(this, glsl_type::float_type);
914 emit(BRW_OPCODE_MOV, expanded, src);
918 fs_inst *inst = emit(opcode, dst, src);
920 if (intel->gen < 6) {
922 inst->mlen = dispatch_width / 8;
929 fs_visitor::emit_math(enum opcode opcode, fs_reg dst, fs_reg src0, fs_reg src1)
935 case SHADER_OPCODE_POW:
936 case SHADER_OPCODE_INT_QUOTIENT:
937 case SHADER_OPCODE_INT_REMAINDER:
940 assert(!"not reached: unsupported binary math opcode.");
944 if (intel->gen >= 7) {
945 inst = emit(opcode, dst, src0, src1);
946 } else if (intel->gen == 6) {
947 /* Can't do hstride == 0 args to gen6 math, so expand it out.
949 * The hardware ignores source modifiers (negate and abs) on math
950 * instructions, so we also move to a temp to set those up.
952 if (src0.file == UNIFORM || src0.abs || src0.negate) {
953 fs_reg expanded = fs_reg(this, glsl_type::float_type);
954 expanded.type = src0.type;
955 emit(BRW_OPCODE_MOV, expanded, src0);
959 if (src1.file == UNIFORM || src1.abs || src1.negate) {
960 fs_reg expanded = fs_reg(this, glsl_type::float_type);
961 expanded.type = src1.type;
962 emit(BRW_OPCODE_MOV, expanded, src1);
966 inst = emit(opcode, dst, src0, src1);
968 /* From the Ironlake PRM, Volume 4, Part 1, Section 6.1.13
971 * "Operand0[7]. For the INT DIV functions, this operand is the
974 * "Operand1[7]. For the INT DIV functions, this operand is the
977 bool is_int_div = opcode != SHADER_OPCODE_POW;
978 fs_reg &op0 = is_int_div ? src1 : src0;
979 fs_reg &op1 = is_int_div ? src0 : src1;
981 emit(BRW_OPCODE_MOV, fs_reg(MRF, base_mrf + 1, op1.type), op1);
982 inst = emit(opcode, dst, op0, reg_null_f);
984 inst->base_mrf = base_mrf;
985 inst->mlen = 2 * dispatch_width / 8;
991 * To be called after the last _mesa_add_state_reference() call, to
992 * set up prog_data.param[] for assign_curb_setup() and
993 * setup_pull_constants().
996 fs_visitor::setup_paramvalues_refs()
998 if (dispatch_width != 8)
1001 /* Set up the pointers to ParamValues now that that array is finalized. */
1002 for (unsigned int i = 0; i < c->prog_data.nr_params; i++) {
1003 c->prog_data.param[i] =
1004 (const float *)fp->Base.Parameters->ParameterValues[this->param_index[i]] +
1005 this->param_offset[i];
1010 fs_visitor::assign_curb_setup()
1012 c->prog_data.curb_read_length = ALIGN(c->prog_data.nr_params, 8) / 8;
1013 if (dispatch_width == 8) {
1014 c->prog_data.first_curbe_grf = c->nr_payload_regs;
1016 c->prog_data.first_curbe_grf_16 = c->nr_payload_regs;
1019 /* Map the offsets in the UNIFORM file to fixed HW regs. */
1020 foreach_list(node, &this->instructions) {
1021 fs_inst *inst = (fs_inst *)node;
1023 for (unsigned int i = 0; i < 3; i++) {
1024 if (inst->src[i].file == UNIFORM) {
1025 int constant_nr = inst->src[i].reg + inst->src[i].reg_offset;
1026 struct brw_reg brw_reg = brw_vec1_grf(c->nr_payload_regs +
1030 inst->src[i].file = FIXED_HW_REG;
1031 inst->src[i].fixed_hw_reg = retype(brw_reg, inst->src[i].type);
1038 fs_visitor::calculate_urb_setup()
1040 for (unsigned int i = 0; i < FRAG_ATTRIB_MAX; i++) {
1045 /* Figure out where each of the incoming setup attributes lands. */
1046 if (intel->gen >= 6) {
1047 for (unsigned int i = 0; i < FRAG_ATTRIB_MAX; i++) {
1048 if (fp->Base.InputsRead & BITFIELD64_BIT(i)) {
1049 urb_setup[i] = urb_next++;
1053 /* FINISHME: The sf doesn't map VS->FS inputs for us very well. */
1054 for (unsigned int i = 0; i < VERT_RESULT_MAX; i++) {
1055 /* Point size is packed into the header, not as a general attribute */
1056 if (i == VERT_RESULT_PSIZ)
1059 if (c->key.vp_outputs_written & BITFIELD64_BIT(i)) {
1060 int fp_index = _mesa_vert_result_to_frag_attrib((gl_vert_result) i);
1062 /* The back color slot is skipped when the front color is
1063 * also written to. In addition, some slots can be
1064 * written in the vertex shader and not read in the
1065 * fragment shader. So the register number must always be
1066 * incremented, mapped or not.
1069 urb_setup[fp_index] = urb_next;
1075 * It's a FS only attribute, and we did interpolation for this attribute
1076 * in SF thread. So, count it here, too.
1078 * See compile_sf_prog() for more info.
1080 if (fp->Base.InputsRead & BITFIELD64_BIT(FRAG_ATTRIB_PNTC))
1081 urb_setup[FRAG_ATTRIB_PNTC] = urb_next++;
1084 /* Each attribute is 4 setup channels, each of which is half a reg. */
1085 c->prog_data.urb_read_length = urb_next * 2;
1089 fs_visitor::assign_urb_setup()
1091 int urb_start = c->nr_payload_regs + c->prog_data.curb_read_length;
1093 /* Offset all the urb_setup[] index by the actual position of the
1094 * setup regs, now that the location of the constants has been chosen.
1096 foreach_list(node, &this->instructions) {
1097 fs_inst *inst = (fs_inst *)node;
1099 if (inst->opcode == FS_OPCODE_LINTERP) {
1100 assert(inst->src[2].file == FIXED_HW_REG);
1101 inst->src[2].fixed_hw_reg.nr += urb_start;
1104 if (inst->opcode == FS_OPCODE_CINTERP) {
1105 assert(inst->src[0].file == FIXED_HW_REG);
1106 inst->src[0].fixed_hw_reg.nr += urb_start;
1110 this->first_non_payload_grf = urb_start + c->prog_data.urb_read_length;
1114 * Split large virtual GRFs into separate components if we can.
1116 * This is mostly duplicated with what brw_fs_vector_splitting does,
1117 * but that's really conservative because it's afraid of doing
1118 * splitting that doesn't result in real progress after the rest of
1119 * the optimization phases, which would cause infinite looping in
1120 * optimization. We can do it once here, safely. This also has the
1121 * opportunity to split interpolated values, or maybe even uniforms,
1122 * which we don't have at the IR level.
1124 * We want to split, because virtual GRFs are what we register
1125 * allocate and spill (due to contiguousness requirements for some
1126 * instructions), and they're what we naturally generate in the
1127 * codegen process, but most virtual GRFs don't actually need to be
1128 * contiguous sets of GRFs. If we split, we'll end up with reduced
1129 * live intervals and better dead code elimination and coalescing.
1132 fs_visitor::split_virtual_grfs()
1134 int num_vars = this->virtual_grf_count;
1135 bool split_grf[num_vars];
1136 int new_virtual_grf[num_vars];
1138 /* Try to split anything > 0 sized. */
1139 for (int i = 0; i < num_vars; i++) {
1140 if (this->virtual_grf_sizes[i] != 1)
1141 split_grf[i] = true;
1143 split_grf[i] = false;
1147 this->delta_x[BRW_WM_PERSPECTIVE_PIXEL_BARYCENTRIC].file == GRF) {
1148 /* PLN opcodes rely on the delta_xy being contiguous. We only have to
1149 * check this for BRW_WM_PERSPECTIVE_PIXEL_BARYCENTRIC, because prior to
1150 * Gen6, that was the only supported interpolation mode, and since Gen6,
1151 * delta_x and delta_y are in fixed hardware registers.
1153 split_grf[this->delta_x[BRW_WM_PERSPECTIVE_PIXEL_BARYCENTRIC].reg] =
1157 foreach_list(node, &this->instructions) {
1158 fs_inst *inst = (fs_inst *)node;
1160 /* If there's a SEND message that requires contiguous destination
1161 * registers, no splitting is allowed.
1163 if (inst->regs_written() > 1) {
1164 split_grf[inst->dst.reg] = false;
1168 /* Allocate new space for split regs. Note that the virtual
1169 * numbers will be contiguous.
1171 for (int i = 0; i < num_vars; i++) {
1173 new_virtual_grf[i] = virtual_grf_alloc(1);
1174 for (int j = 2; j < this->virtual_grf_sizes[i]; j++) {
1175 int reg = virtual_grf_alloc(1);
1176 assert(reg == new_virtual_grf[i] + j - 1);
1179 this->virtual_grf_sizes[i] = 1;
1183 foreach_list(node, &this->instructions) {
1184 fs_inst *inst = (fs_inst *)node;
1186 if (inst->dst.file == GRF &&
1187 split_grf[inst->dst.reg] &&
1188 inst->dst.reg_offset != 0) {
1189 inst->dst.reg = (new_virtual_grf[inst->dst.reg] +
1190 inst->dst.reg_offset - 1);
1191 inst->dst.reg_offset = 0;
1193 for (int i = 0; i < 3; i++) {
1194 if (inst->src[i].file == GRF &&
1195 split_grf[inst->src[i].reg] &&
1196 inst->src[i].reg_offset != 0) {
1197 inst->src[i].reg = (new_virtual_grf[inst->src[i].reg] +
1198 inst->src[i].reg_offset - 1);
1199 inst->src[i].reg_offset = 0;
1203 this->live_intervals_valid = false;
1207 * Remove unused virtual GRFs and compact the virtual_grf_* arrays.
1209 * During code generation, we create tons of temporary variables, many of
1210 * which get immediately killed and are never used again. Yet, in later
1211 * optimization and analysis passes, such as compute_live_intervals, we need
1212 * to loop over all the virtual GRFs. Compacting them can save a lot of
1216 fs_visitor::compact_virtual_grfs()
1218 /* Mark which virtual GRFs are used, and count how many. */
1219 int remap_table[this->virtual_grf_count];
1220 memset(remap_table, -1, sizeof(remap_table));
1222 foreach_list(node, &this->instructions) {
1223 const fs_inst *inst = (const fs_inst *) node;
1225 if (inst->dst.file == GRF)
1226 remap_table[inst->dst.reg] = 0;
1228 for (int i = 0; i < 3; i++) {
1229 if (inst->src[i].file == GRF)
1230 remap_table[inst->src[i].reg] = 0;
1234 /* In addition to registers used in instructions, fs_visitor keeps
1235 * direct references to certain special values which must be patched:
1237 fs_reg *special[] = {
1238 &frag_depth, &pixel_x, &pixel_y, &pixel_w, &wpos_w, &dual_src_output,
1239 &outputs[0], &outputs[1], &outputs[2], &outputs[3],
1240 &outputs[4], &outputs[5], &outputs[6], &outputs[7],
1241 &delta_x[0], &delta_x[1], &delta_x[2],
1242 &delta_x[3], &delta_x[4], &delta_x[5],
1243 &delta_y[0], &delta_y[1], &delta_y[2],
1244 &delta_y[3], &delta_y[4], &delta_y[5],
1246 STATIC_ASSERT(BRW_WM_BARYCENTRIC_INTERP_MODE_COUNT == 6);
1247 STATIC_ASSERT(BRW_MAX_DRAW_BUFFERS == 8);
1249 /* Treat all special values as used, to be conservative */
1250 for (unsigned i = 0; i < ARRAY_SIZE(special); i++) {
1251 if (special[i]->file == GRF)
1252 remap_table[special[i]->reg] = 0;
1255 /* Compact the GRF arrays. */
1257 for (int i = 0; i < this->virtual_grf_count; i++) {
1258 if (remap_table[i] != -1) {
1259 remap_table[i] = new_index;
1260 virtual_grf_sizes[new_index] = virtual_grf_sizes[i];
1261 if (live_intervals_valid) {
1262 virtual_grf_use[new_index] = virtual_grf_use[i];
1263 virtual_grf_def[new_index] = virtual_grf_def[i];
1269 this->virtual_grf_count = new_index;
1271 /* Patch all the instructions to use the newly renumbered registers */
1272 foreach_list(node, &this->instructions) {
1273 fs_inst *inst = (fs_inst *) node;
1275 if (inst->dst.file == GRF)
1276 inst->dst.reg = remap_table[inst->dst.reg];
1278 for (int i = 0; i < 3; i++) {
1279 if (inst->src[i].file == GRF)
1280 inst->src[i].reg = remap_table[inst->src[i].reg];
1284 /* Patch all the references to special values */
1285 for (unsigned i = 0; i < ARRAY_SIZE(special); i++) {
1286 if (special[i]->file == GRF && remap_table[special[i]->reg] != -1)
1287 special[i]->reg = remap_table[special[i]->reg];
1292 fs_visitor::remove_dead_constants()
1294 if (dispatch_width == 8) {
1295 this->params_remap = ralloc_array(mem_ctx, int, c->prog_data.nr_params);
1297 for (unsigned int i = 0; i < c->prog_data.nr_params; i++)
1298 this->params_remap[i] = -1;
1300 /* Find which params are still in use. */
1301 foreach_list(node, &this->instructions) {
1302 fs_inst *inst = (fs_inst *)node;
1304 for (int i = 0; i < 3; i++) {
1305 int constant_nr = inst->src[i].reg + inst->src[i].reg_offset;
1307 if (inst->src[i].file != UNIFORM)
1310 assert(constant_nr < (int)c->prog_data.nr_params);
1312 /* For now, set this to non-negative. We'll give it the
1313 * actual new number in a moment, in order to keep the
1314 * register numbers nicely ordered.
1316 this->params_remap[constant_nr] = 0;
1320 /* Figure out what the new numbers for the params will be. At some
1321 * point when we're doing uniform array access, we're going to want
1322 * to keep the distinction between .reg and .reg_offset, but for
1323 * now we don't care.
1325 unsigned int new_nr_params = 0;
1326 for (unsigned int i = 0; i < c->prog_data.nr_params; i++) {
1327 if (this->params_remap[i] != -1) {
1328 this->params_remap[i] = new_nr_params++;
1332 /* Update the list of params to be uploaded to match our new numbering. */
1333 for (unsigned int i = 0; i < c->prog_data.nr_params; i++) {
1334 int remapped = this->params_remap[i];
1339 /* We've already done setup_paramvalues_refs() so no need to worry
1340 * about param_index and param_offset.
1342 c->prog_data.param[remapped] = c->prog_data.param[i];
1345 c->prog_data.nr_params = new_nr_params;
1347 /* This should have been generated in the 8-wide pass already. */
1348 assert(this->params_remap);
1351 /* Now do the renumbering of the shader to remove unused params. */
1352 foreach_list(node, &this->instructions) {
1353 fs_inst *inst = (fs_inst *)node;
1355 for (int i = 0; i < 3; i++) {
1356 int constant_nr = inst->src[i].reg + inst->src[i].reg_offset;
1358 if (inst->src[i].file != UNIFORM)
1361 assert(this->params_remap[constant_nr] != -1);
1362 inst->src[i].reg = this->params_remap[constant_nr];
1363 inst->src[i].reg_offset = 0;
1371 * Choose accesses from the UNIFORM file to demote to using the pull
1374 * We allow a fragment shader to have more than the specified minimum
1375 * maximum number of fragment shader uniform components (64). If
1376 * there are too many of these, they'd fill up all of register space.
1377 * So, this will push some of them out to the pull constant buffer and
1378 * update the program to load them.
1381 fs_visitor::setup_pull_constants()
1383 /* Only allow 16 registers (128 uniform components) as push constants. */
1384 unsigned int max_uniform_components = 16 * 8;
1385 if (c->prog_data.nr_params <= max_uniform_components)
1388 if (dispatch_width == 16) {
1389 fail("Pull constants not supported in 16-wide\n");
1393 /* Just demote the end of the list. We could probably do better
1394 * here, demoting things that are rarely used in the program first.
1396 int pull_uniform_base = max_uniform_components;
1397 int pull_uniform_count = c->prog_data.nr_params - pull_uniform_base;
1399 foreach_list(node, &this->instructions) {
1400 fs_inst *inst = (fs_inst *)node;
1402 for (int i = 0; i < 3; i++) {
1403 if (inst->src[i].file != UNIFORM)
1406 int uniform_nr = inst->src[i].reg + inst->src[i].reg_offset;
1407 if (uniform_nr < pull_uniform_base)
1410 fs_reg dst = fs_reg(this, glsl_type::float_type);
1411 fs_reg index = fs_reg((unsigned)SURF_INDEX_FRAG_CONST_BUFFER);
1412 fs_reg offset = fs_reg((unsigned)(((uniform_nr -
1413 pull_uniform_base) * 4) & ~15));
1414 fs_inst *pull = new(mem_ctx) fs_inst(FS_OPCODE_PULL_CONSTANT_LOAD,
1415 dst, index, offset);
1416 pull->ir = inst->ir;
1417 pull->annotation = inst->annotation;
1418 pull->base_mrf = 14;
1421 inst->insert_before(pull);
1423 inst->src[i].file = GRF;
1424 inst->src[i].reg = dst.reg;
1425 inst->src[i].reg_offset = 0;
1426 inst->src[i].smear = (uniform_nr - pull_uniform_base) & 3;
1430 for (int i = 0; i < pull_uniform_count; i++) {
1431 c->prog_data.pull_param[i] = c->prog_data.param[pull_uniform_base + i];
1433 c->prog_data.nr_params -= pull_uniform_count;
1434 c->prog_data.nr_pull_params = pull_uniform_count;
1438 fs_visitor::opt_algebraic()
1440 bool progress = false;
1442 foreach_list(node, &this->instructions) {
1443 fs_inst *inst = (fs_inst *)node;
1445 switch (inst->opcode) {
1446 case BRW_OPCODE_MUL:
1447 if (inst->src[1].file != IMM)
1451 if (inst->src[1].is_one()) {
1452 inst->opcode = BRW_OPCODE_MOV;
1453 inst->src[1] = reg_undef;
1459 if (inst->src[1].is_zero()) {
1460 inst->opcode = BRW_OPCODE_MOV;
1461 inst->src[0] = inst->src[1];
1462 inst->src[1] = reg_undef;
1468 case BRW_OPCODE_ADD:
1469 if (inst->src[1].file != IMM)
1473 if (inst->src[1].is_zero()) {
1474 inst->opcode = BRW_OPCODE_MOV;
1475 inst->src[1] = reg_undef;
1489 * Must be called after calculate_live_intervales() to remove unused
1490 * writes to registers -- register allocation will fail otherwise
1491 * because something deffed but not used won't be considered to
1492 * interfere with other regs.
1495 fs_visitor::dead_code_eliminate()
1497 bool progress = false;
1500 calculate_live_intervals();
1502 foreach_list_safe(node, &this->instructions) {
1503 fs_inst *inst = (fs_inst *)node;
1505 if (inst->dst.file == GRF && this->virtual_grf_use[inst->dst.reg] <= pc) {
1514 live_intervals_valid = false;
1520 * Implements a second type of register coalescing: This one checks if
1521 * the two regs involved in a raw move don't interfere, in which case
1522 * they can both by stored in the same place and the MOV removed.
1525 fs_visitor::register_coalesce_2()
1527 bool progress = false;
1529 calculate_live_intervals();
1531 foreach_list_safe(node, &this->instructions) {
1532 fs_inst *inst = (fs_inst *)node;
1534 if (inst->opcode != BRW_OPCODE_MOV ||
1537 inst->src[0].file != GRF ||
1538 inst->src[0].negate ||
1540 inst->src[0].smear != -1 ||
1541 inst->dst.file != GRF ||
1542 inst->dst.type != inst->src[0].type ||
1543 virtual_grf_sizes[inst->src[0].reg] != 1 ||
1544 virtual_grf_interferes(inst->dst.reg, inst->src[0].reg)) {
1548 int reg_from = inst->src[0].reg;
1549 assert(inst->src[0].reg_offset == 0);
1550 int reg_to = inst->dst.reg;
1551 int reg_to_offset = inst->dst.reg_offset;
1553 foreach_list_safe(node, &this->instructions) {
1554 fs_inst *scan_inst = (fs_inst *)node;
1556 if (scan_inst->dst.file == GRF &&
1557 scan_inst->dst.reg == reg_from) {
1558 scan_inst->dst.reg = reg_to;
1559 scan_inst->dst.reg_offset = reg_to_offset;
1561 for (int i = 0; i < 3; i++) {
1562 if (scan_inst->src[i].file == GRF &&
1563 scan_inst->src[i].reg == reg_from) {
1564 scan_inst->src[i].reg = reg_to;
1565 scan_inst->src[i].reg_offset = reg_to_offset;
1571 live_intervals_valid = false;
1580 fs_visitor::register_coalesce()
1582 bool progress = false;
1586 foreach_list_safe(node, &this->instructions) {
1587 fs_inst *inst = (fs_inst *)node;
1589 /* Make sure that we dominate the instructions we're going to
1590 * scan for interfering with our coalescing, or we won't have
1591 * scanned enough to see if anything interferes with our
1592 * coalescing. We don't dominate the following instructions if
1593 * we're in a loop or an if block.
1595 switch (inst->opcode) {
1599 case BRW_OPCODE_WHILE:
1605 case BRW_OPCODE_ENDIF:
1611 if (loop_depth || if_depth)
1614 if (inst->opcode != BRW_OPCODE_MOV ||
1617 inst->dst.file != GRF || (inst->src[0].file != GRF &&
1618 inst->src[0].file != UNIFORM)||
1619 inst->dst.type != inst->src[0].type)
1622 bool has_source_modifiers = inst->src[0].abs || inst->src[0].negate;
1624 /* Found a move of a GRF to a GRF. Let's see if we can coalesce
1625 * them: check for no writes to either one until the exit of the
1628 bool interfered = false;
1630 for (fs_inst *scan_inst = (fs_inst *)inst->next;
1631 !scan_inst->is_tail_sentinel();
1632 scan_inst = (fs_inst *)scan_inst->next) {
1633 if (scan_inst->dst.file == GRF) {
1634 if (scan_inst->overwrites_reg(inst->dst) ||
1635 scan_inst->overwrites_reg(inst->src[0])) {
1641 /* The gen6 MATH instruction can't handle source modifiers or
1642 * unusual register regions, so avoid coalescing those for
1643 * now. We should do something more specific.
1645 if (intel->gen >= 6 &&
1646 scan_inst->is_math() &&
1647 (has_source_modifiers || inst->src[0].file == UNIFORM)) {
1652 /* The accumulator result appears to get used for the
1653 * conditional modifier generation. When negating a UD
1654 * value, there is a 33rd bit generated for the sign in the
1655 * accumulator value, so now you can't check, for example,
1656 * equality with a 32-bit value. See piglit fs-op-neg-uint.
1658 if (scan_inst->conditional_mod &&
1659 inst->src[0].negate &&
1660 inst->src[0].type == BRW_REGISTER_TYPE_UD) {
1669 /* Rewrite the later usage to point at the source of the move to
1672 for (fs_inst *scan_inst = inst;
1673 !scan_inst->is_tail_sentinel();
1674 scan_inst = (fs_inst *)scan_inst->next) {
1675 for (int i = 0; i < 3; i++) {
1676 if (scan_inst->src[i].file == GRF &&
1677 scan_inst->src[i].reg == inst->dst.reg &&
1678 scan_inst->src[i].reg_offset == inst->dst.reg_offset) {
1679 fs_reg new_src = inst->src[0];
1680 if (scan_inst->src[i].abs) {
1684 new_src.negate ^= scan_inst->src[i].negate;
1685 scan_inst->src[i] = new_src;
1695 live_intervals_valid = false;
1702 fs_visitor::compute_to_mrf()
1704 bool progress = false;
1707 calculate_live_intervals();
1709 foreach_list_safe(node, &this->instructions) {
1710 fs_inst *inst = (fs_inst *)node;
1715 if (inst->opcode != BRW_OPCODE_MOV ||
1717 inst->dst.file != MRF || inst->src[0].file != GRF ||
1718 inst->dst.type != inst->src[0].type ||
1719 inst->src[0].abs || inst->src[0].negate || inst->src[0].smear != -1)
1722 /* Work out which hardware MRF registers are written by this
1725 int mrf_low = inst->dst.reg & ~BRW_MRF_COMPR4;
1727 if (inst->dst.reg & BRW_MRF_COMPR4) {
1728 mrf_high = mrf_low + 4;
1729 } else if (dispatch_width == 16 &&
1730 (!inst->force_uncompressed && !inst->force_sechalf)) {
1731 mrf_high = mrf_low + 1;
1736 /* Can't compute-to-MRF this GRF if someone else was going to
1739 if (this->virtual_grf_use[inst->src[0].reg] > ip)
1742 /* Found a move of a GRF to a MRF. Let's see if we can go
1743 * rewrite the thing that made this GRF to write into the MRF.
1746 for (scan_inst = (fs_inst *)inst->prev;
1747 scan_inst->prev != NULL;
1748 scan_inst = (fs_inst *)scan_inst->prev) {
1749 if (scan_inst->dst.file == GRF &&
1750 scan_inst->dst.reg == inst->src[0].reg) {
1751 /* Found the last thing to write our reg we want to turn
1752 * into a compute-to-MRF.
1755 /* SENDs can only write to GRFs, so no compute-to-MRF. */
1756 if (scan_inst->mlen) {
1760 /* If it's predicated, it (probably) didn't populate all
1761 * the channels. We might be able to rewrite everything
1762 * that writes that reg, but it would require smarter
1763 * tracking to delay the rewriting until complete success.
1765 if (scan_inst->predicate)
1768 /* If it's half of register setup and not the same half as
1769 * our MOV we're trying to remove, bail for now.
1771 if (scan_inst->force_uncompressed != inst->force_uncompressed ||
1772 scan_inst->force_sechalf != inst->force_sechalf) {
1776 /* SEND instructions can't have MRF as a destination. */
1777 if (scan_inst->mlen)
1780 if (intel->gen >= 6) {
1781 /* gen6 math instructions must have the destination be
1782 * GRF, so no compute-to-MRF for them.
1784 if (scan_inst->is_math()) {
1789 if (scan_inst->dst.reg_offset == inst->src[0].reg_offset) {
1790 /* Found the creator of our MRF's source value. */
1791 scan_inst->dst.file = MRF;
1792 scan_inst->dst.reg = inst->dst.reg;
1793 scan_inst->saturate |= inst->saturate;
1800 /* We don't handle flow control here. Most computation of
1801 * values that end up in MRFs are shortly before the MRF
1804 if (scan_inst->opcode == BRW_OPCODE_DO ||
1805 scan_inst->opcode == BRW_OPCODE_WHILE ||
1806 scan_inst->opcode == BRW_OPCODE_ELSE ||
1807 scan_inst->opcode == BRW_OPCODE_ENDIF) {
1811 /* You can't read from an MRF, so if someone else reads our
1812 * MRF's source GRF that we wanted to rewrite, that stops us.
1814 bool interfered = false;
1815 for (int i = 0; i < 3; i++) {
1816 if (scan_inst->src[i].file == GRF &&
1817 scan_inst->src[i].reg == inst->src[0].reg &&
1818 scan_inst->src[i].reg_offset == inst->src[0].reg_offset) {
1825 if (scan_inst->dst.file == MRF) {
1826 /* If somebody else writes our MRF here, we can't
1827 * compute-to-MRF before that.
1829 int scan_mrf_low = scan_inst->dst.reg & ~BRW_MRF_COMPR4;
1832 if (scan_inst->dst.reg & BRW_MRF_COMPR4) {
1833 scan_mrf_high = scan_mrf_low + 4;
1834 } else if (dispatch_width == 16 &&
1835 (!scan_inst->force_uncompressed &&
1836 !scan_inst->force_sechalf)) {
1837 scan_mrf_high = scan_mrf_low + 1;
1839 scan_mrf_high = scan_mrf_low;
1842 if (mrf_low == scan_mrf_low ||
1843 mrf_low == scan_mrf_high ||
1844 mrf_high == scan_mrf_low ||
1845 mrf_high == scan_mrf_high) {
1850 if (scan_inst->mlen > 0) {
1851 /* Found a SEND instruction, which means that there are
1852 * live values in MRFs from base_mrf to base_mrf +
1853 * scan_inst->mlen - 1. Don't go pushing our MRF write up
1856 if (mrf_low >= scan_inst->base_mrf &&
1857 mrf_low < scan_inst->base_mrf + scan_inst->mlen) {
1860 if (mrf_high >= scan_inst->base_mrf &&
1861 mrf_high < scan_inst->base_mrf + scan_inst->mlen) {
1869 live_intervals_valid = false;
1875 * Walks through basic blocks, looking for repeated MRF writes and
1876 * removing the later ones.
1879 fs_visitor::remove_duplicate_mrf_writes()
1881 fs_inst *last_mrf_move[16];
1882 bool progress = false;
1884 /* Need to update the MRF tracking for compressed instructions. */
1885 if (dispatch_width == 16)
1888 memset(last_mrf_move, 0, sizeof(last_mrf_move));
1890 foreach_list_safe(node, &this->instructions) {
1891 fs_inst *inst = (fs_inst *)node;
1893 switch (inst->opcode) {
1895 case BRW_OPCODE_WHILE:
1897 case BRW_OPCODE_ELSE:
1898 case BRW_OPCODE_ENDIF:
1899 memset(last_mrf_move, 0, sizeof(last_mrf_move));
1905 if (inst->opcode == BRW_OPCODE_MOV &&
1906 inst->dst.file == MRF) {
1907 fs_inst *prev_inst = last_mrf_move[inst->dst.reg];
1908 if (prev_inst && inst->equals(prev_inst)) {
1915 /* Clear out the last-write records for MRFs that were overwritten. */
1916 if (inst->dst.file == MRF) {
1917 last_mrf_move[inst->dst.reg] = NULL;
1920 if (inst->mlen > 0) {
1921 /* Found a SEND instruction, which will include two or fewer
1922 * implied MRF writes. We could do better here.
1924 for (int i = 0; i < implied_mrf_writes(inst); i++) {
1925 last_mrf_move[inst->base_mrf + i] = NULL;
1929 /* Clear out any MRF move records whose sources got overwritten. */
1930 if (inst->dst.file == GRF) {
1931 for (unsigned int i = 0; i < Elements(last_mrf_move); i++) {
1932 if (last_mrf_move[i] &&
1933 last_mrf_move[i]->src[0].reg == inst->dst.reg) {
1934 last_mrf_move[i] = NULL;
1939 if (inst->opcode == BRW_OPCODE_MOV &&
1940 inst->dst.file == MRF &&
1941 inst->src[0].file == GRF &&
1943 last_mrf_move[inst->dst.reg] = inst;
1948 live_intervals_valid = false;
1954 fs_visitor::dump_instruction(fs_inst *inst)
1956 if (inst->opcode < ARRAY_SIZE(opcode_descs) &&
1957 opcode_descs[inst->opcode].name) {
1958 printf("%s", opcode_descs[inst->opcode].name);
1960 printf("op%d", inst->opcode);
1966 switch (inst->dst.file) {
1968 printf("vgrf%d", inst->dst.reg);
1969 if (inst->dst.reg_offset)
1970 printf("+%d", inst->dst.reg_offset);
1973 printf("m%d", inst->dst.reg);
1979 printf("***u%d***", inst->dst.reg);
1987 for (int i = 0; i < 3; i++) {
1988 if (inst->src[i].negate)
1990 if (inst->src[i].abs)
1992 switch (inst->src[i].file) {
1994 printf("vgrf%d", inst->src[i].reg);
1995 if (inst->src[i].reg_offset)
1996 printf("+%d", inst->src[i].reg_offset);
1999 printf("***m%d***", inst->src[i].reg);
2002 printf("u%d", inst->src[i].reg);
2003 if (inst->src[i].reg_offset)
2004 printf(".%d", inst->src[i].reg_offset);
2013 if (inst->src[i].abs)
2022 if (inst->force_uncompressed)
2025 if (inst->force_sechalf)
2032 fs_visitor::dump_instructions()
2035 foreach_list(node, &this->instructions) {
2036 fs_inst *inst = (fs_inst *)node;
2037 printf("%d: ", ip++);
2038 dump_instruction(inst);
2043 * Possibly returns an instruction that set up @param reg.
2045 * Sometimes we want to take the result of some expression/variable
2046 * dereference tree and rewrite the instruction generating the result
2047 * of the tree. When processing the tree, we know that the
2048 * instructions generated are all writing temporaries that are dead
2049 * outside of this tree. So, if we have some instructions that write
2050 * a temporary, we're free to point that temp write somewhere else.
2052 * Note that this doesn't guarantee that the instruction generated
2053 * only reg -- it might be the size=4 destination of a texture instruction.
2056 fs_visitor::get_instruction_generating_reg(fs_inst *start,
2062 end->force_uncompressed ||
2063 end->force_sechalf ||
2064 !reg.equals(end->dst)) {
2072 fs_visitor::setup_payload_gen6()
2074 struct intel_context *intel = &brw->intel;
2076 (fp->Base.InputsRead & (1 << FRAG_ATTRIB_WPOS)) != 0;
2077 unsigned barycentric_interp_modes = c->prog_data.barycentric_interp_modes;
2079 assert(intel->gen >= 6);
2081 /* R0-1: masks, pixel X/Y coordinates. */
2082 c->nr_payload_regs = 2;
2083 /* R2: only for 32-pixel dispatch.*/
2085 /* R3-26: barycentric interpolation coordinates. These appear in the
2086 * same order that they appear in the brw_wm_barycentric_interp_mode
2087 * enum. Each set of coordinates occupies 2 registers if dispatch width
2088 * == 8 and 4 registers if dispatch width == 16. Coordinates only
2089 * appear if they were enabled using the "Barycentric Interpolation
2090 * Mode" bits in WM_STATE.
2092 for (int i = 0; i < BRW_WM_BARYCENTRIC_INTERP_MODE_COUNT; ++i) {
2093 if (barycentric_interp_modes & (1 << i)) {
2094 c->barycentric_coord_reg[i] = c->nr_payload_regs;
2095 c->nr_payload_regs += 2;
2096 if (dispatch_width == 16) {
2097 c->nr_payload_regs += 2;
2102 /* R27: interpolated depth if uses source depth */
2104 c->source_depth_reg = c->nr_payload_regs;
2105 c->nr_payload_regs++;
2106 if (dispatch_width == 16) {
2107 /* R28: interpolated depth if not 8-wide. */
2108 c->nr_payload_regs++;
2111 /* R29: interpolated W set if GEN6_WM_USES_SOURCE_W. */
2113 c->source_w_reg = c->nr_payload_regs;
2114 c->nr_payload_regs++;
2115 if (dispatch_width == 16) {
2116 /* R30: interpolated W if not 8-wide. */
2117 c->nr_payload_regs++;
2120 /* R31: MSAA position offsets. */
2121 /* R32-: bary for 32-pixel. */
2122 /* R58-59: interp W for 32-pixel. */
2124 if (fp->Base.OutputsWritten & BITFIELD64_BIT(FRAG_RESULT_DEPTH)) {
2125 c->source_depth_to_render_target = true;
2132 uint32_t orig_nr_params = c->prog_data.nr_params;
2134 if (intel->gen >= 6)
2135 setup_payload_gen6();
2137 setup_payload_gen4();
2142 calculate_urb_setup();
2144 emit_interpolation_setup_gen4();
2146 emit_interpolation_setup_gen6();
2148 /* Generate FS IR for main(). (the visitor only descends into
2149 * functions called "main").
2152 foreach_list(node, &*shader->ir) {
2153 ir_instruction *ir = (ir_instruction *)node;
2155 this->result = reg_undef;
2159 emit_fragment_program_code();
2167 split_virtual_grfs();
2169 setup_paramvalues_refs();
2170 setup_pull_constants();
2176 compact_virtual_grfs();
2178 progress = remove_duplicate_mrf_writes() || progress;
2180 progress = opt_algebraic() || progress;
2181 progress = opt_cse() || progress;
2182 progress = opt_copy_propagate() || progress;
2183 progress = dead_code_eliminate() || progress;
2184 progress = register_coalesce() || progress;
2185 progress = register_coalesce_2() || progress;
2186 progress = compute_to_mrf() || progress;
2189 remove_dead_constants();
2191 schedule_instructions();
2193 assign_curb_setup();
2197 /* Debug of register spilling: Go spill everything. */
2198 for (int i = 0; i < virtual_grf_count; i++) {
2204 assign_regs_trivial();
2206 while (!assign_regs()) {
2212 assert(force_uncompressed_stack == 0);
2213 assert(force_sechalf_stack == 0);
2218 if (dispatch_width == 8) {
2219 c->prog_data.reg_blocks = brw_register_blocks(grf_used);
2221 c->prog_data.reg_blocks_16 = brw_register_blocks(grf_used);
2223 /* Make sure we didn't try to sneak in an extra uniform */
2224 assert(orig_nr_params == c->prog_data.nr_params);
2225 (void) orig_nr_params;
2232 brw_wm_fs_emit(struct brw_context *brw, struct brw_wm_compile *c,
2233 struct gl_fragment_program *fp,
2234 struct gl_shader_program *prog,
2235 unsigned *final_assembly_size)
2237 struct intel_context *intel = &brw->intel;
2238 bool start_busy = false;
2239 float start_time = 0;
2241 if (unlikely(INTEL_DEBUG & DEBUG_PERF)) {
2242 start_busy = (intel->batch.last_bo &&
2243 drm_intel_bo_busy(intel->batch.last_bo));
2244 start_time = get_time();
2247 struct brw_shader *shader = NULL;
2249 shader = (brw_shader *) prog->_LinkedShaders[MESA_SHADER_FRAGMENT];
2251 if (unlikely(INTEL_DEBUG & DEBUG_WM)) {
2253 printf("GLSL IR for native fragment shader %d:\n", prog->Name);
2254 _mesa_print_ir(shader->ir, NULL);
2257 printf("ARB_fragment_program %d ir for native fragment shader\n",
2259 _mesa_print_program(&fp->Base);
2263 /* Now the main event: Visit the shader IR and generate our FS IR for it.
2265 fs_visitor v(brw, c, prog, fp, 8);
2267 prog->LinkStatus = false;
2268 ralloc_strcat(&prog->InfoLog, v.fail_msg);
2270 _mesa_problem(NULL, "Failed to compile fragment shader: %s\n",
2276 exec_list *simd16_instructions = NULL;
2277 fs_visitor v2(brw, c, prog, fp, 16);
2278 if (intel->gen >= 5 && c->prog_data.nr_pull_params == 0) {
2279 v2.import_uniforms(&v);
2281 perf_debug("16-wide shader failed to compile, falling back to "
2282 "8-wide at a 10-20%% performance cost: %s", v2.fail_msg);
2284 simd16_instructions = &v2.instructions;
2288 c->prog_data.dispatch_width = 8;
2290 fs_generator g(brw, c, prog, fp, v.dual_src_output.file != BAD_FILE);
2291 const unsigned *generated = g.generate_assembly(&v.instructions,
2292 simd16_instructions,
2293 final_assembly_size);
2295 if (unlikely(INTEL_DEBUG & DEBUG_PERF) && shader) {
2296 if (shader->compiled_once)
2297 brw_wm_debug_recompile(brw, prog, &c->key);
2298 shader->compiled_once = true;
2300 if (start_busy && !drm_intel_bo_busy(intel->batch.last_bo)) {
2301 perf_debug("FS compile took %.03f ms and stalled the GPU\n",
2302 (get_time() - start_time) * 1000);
2310 brw_fs_precompile(struct gl_context *ctx, struct gl_shader_program *prog)
2312 struct brw_context *brw = brw_context(ctx);
2313 struct intel_context *intel = &brw->intel;
2314 struct brw_wm_prog_key key;
2316 if (!prog->_LinkedShaders[MESA_SHADER_FRAGMENT])
2319 struct gl_fragment_program *fp = (struct gl_fragment_program *)
2320 prog->_LinkedShaders[MESA_SHADER_FRAGMENT]->Program;
2321 struct brw_fragment_program *bfp = brw_fragment_program(fp);
2322 bool program_uses_dfdy = fp->UsesDFdy;
2324 memset(&key, 0, sizeof(key));
2326 if (intel->gen < 6) {
2328 key.iz_lookup |= IZ_PS_KILL_ALPHATEST_BIT;
2330 if (fp->Base.OutputsWritten & BITFIELD64_BIT(FRAG_RESULT_DEPTH))
2331 key.iz_lookup |= IZ_PS_COMPUTES_DEPTH_BIT;
2333 /* Just assume depth testing. */
2334 key.iz_lookup |= IZ_DEPTH_TEST_ENABLE_BIT;
2335 key.iz_lookup |= IZ_DEPTH_WRITE_ENABLE_BIT;
2338 if (prog->Name != 0)
2339 key.proj_attrib_mask = 0xffffffff;
2342 key.vp_outputs_written |= BITFIELD64_BIT(FRAG_ATTRIB_WPOS);
2344 for (int i = 0; i < FRAG_ATTRIB_MAX; i++) {
2345 if (!(fp->Base.InputsRead & BITFIELD64_BIT(i)))
2348 if (prog->Name == 0)
2349 key.proj_attrib_mask |= 1 << i;
2351 if (intel->gen < 6) {
2352 int vp_index = _mesa_vert_result_to_frag_attrib((gl_vert_result) i);
2355 key.vp_outputs_written |= BITFIELD64_BIT(vp_index);
2359 key.clamp_fragment_color = true;
2361 for (int i = 0; i < MAX_SAMPLERS; i++) {
2362 if (fp->Base.ShadowSamplers & (1 << i)) {
2363 /* Assume DEPTH_TEXTURE_MODE is the default: X, X, X, 1 */
2364 key.tex.swizzles[i] =
2365 MAKE_SWIZZLE4(SWIZZLE_X, SWIZZLE_X, SWIZZLE_X, SWIZZLE_ONE);
2367 /* Color sampler: assume no swizzling. */
2368 key.tex.swizzles[i] = SWIZZLE_XYZW;
2372 if (fp->Base.InputsRead & FRAG_BIT_WPOS) {
2373 key.drawable_height = ctx->DrawBuffer->Height;
2376 if ((fp->Base.InputsRead & FRAG_BIT_WPOS) || program_uses_dfdy) {
2377 key.render_to_fbo = _mesa_is_user_fbo(ctx->DrawBuffer);
2380 key.nr_color_regions = 1;
2382 key.program_string_id = bfp->id;
2384 uint32_t old_prog_offset = brw->wm.prog_offset;
2385 struct brw_wm_prog_data *old_prog_data = brw->wm.prog_data;
2387 bool success = do_wm_prog(brw, prog, bfp, &key);
2389 brw->wm.prog_offset = old_prog_offset;
2390 brw->wm.prog_data = old_prog_data;