#include "brw_vec4.h"
#include "brw_cfg.h"
#include "brw_vs.h"
+#include "brw_dead_control_flow.h"
extern "C" {
#include "main/macros.h"
if (type && (type->is_scalar() || type->is_vector() || type->is_matrix()))
this->swizzle = swizzle_for_size(type->vector_elements);
else
- this->swizzle = SWIZZLE_XYZW;
+ this->swizzle = BRW_SWIZZLE_XYZW;
}
/** Generic unset register constructor. */
this->file = IMM;
this->type = BRW_REGISTER_TYPE_F;
- this->imm.f = f;
+ this->fixed_hw_reg.dw1.f = f;
}
src_reg::src_reg(uint32_t u)
this->file = IMM;
this->type = BRW_REGISTER_TYPE_UD;
- this->imm.u = u;
+ this->fixed_hw_reg.dw1.ud = u;
}
src_reg::src_reg(int32_t i)
this->file = IMM;
this->type = BRW_REGISTER_TYPE_D;
- this->imm.i = i;
+ this->fixed_hw_reg.dw1.d = i;
+}
+
+src_reg::src_reg(struct brw_reg reg)
+{
+ init();
+
+ this->file = HW_REG;
+ this->fixed_hw_reg = reg;
+ this->type = reg.type;
}
src_reg::src_reg(dst_reg reg)
this->file = HW_REG;
this->fixed_hw_reg = reg;
+ this->type = reg.type;
}
dst_reg::dst_reg(src_reg reg)
}
bool
-vec4_visitor::can_do_source_mods(vec4_instruction *inst)
+vec4_instruction::can_do_source_mods(struct brw_context *brw)
{
- if (brw->gen == 6 && inst->is_math())
+ if (brw->gen == 6 && is_math())
+ return false;
+
+ if (is_send_from_grf())
return false;
- if (inst->is_send_from_grf())
+ if (!backend_instruction::can_do_source_mods())
return false;
return true;
return 1;
case VS_OPCODE_PULL_CONSTANT_LOAD:
return 2;
- case VS_OPCODE_SCRATCH_READ:
+ case SHADER_OPCODE_GEN4_SCRATCH_READ:
return 2;
- case VS_OPCODE_SCRATCH_WRITE:
+ case SHADER_OPCODE_GEN4_SCRATCH_WRITE:
return 3;
case GS_OPCODE_URB_WRITE:
case GS_OPCODE_THREAD_END:
case SHADER_OPCODE_TXL:
case SHADER_OPCODE_TXD:
case SHADER_OPCODE_TXF:
- case SHADER_OPCODE_TXF_MS:
+ case SHADER_OPCODE_TXF_CMS:
+ case SHADER_OPCODE_TXF_MCS:
case SHADER_OPCODE_TXS:
case SHADER_OPCODE_TG4:
+ case SHADER_OPCODE_TG4_OFFSET:
return inst->header_present ? 1 : 0;
+ case SHADER_OPCODE_UNTYPED_ATOMIC:
+ case SHADER_OPCODE_UNTYPED_SURFACE_READ:
+ return 0;
default:
- assert(!"not reached");
- return inst->mlen;
+ unreachable("not reached");
}
}
bool
-src_reg::equals(src_reg *r)
+src_reg::equals(const src_reg &r) const
+{
+ return (file == r.file &&
+ reg == r.reg &&
+ reg_offset == r.reg_offset &&
+ type == r.type &&
+ negate == r.negate &&
+ abs == r.abs &&
+ swizzle == r.swizzle &&
+ !reladdr && !r.reladdr &&
+ memcmp(&fixed_hw_reg, &r.fixed_hw_reg,
+ sizeof(fixed_hw_reg)) == 0);
+}
+
+static bool
+try_eliminate_instruction(vec4_instruction *inst, int new_writemask,
+ const struct brw_context *brw)
{
- return (file == r->file &&
- reg == r->reg &&
- reg_offset == r->reg_offset &&
- type == r->type &&
- negate == r->negate &&
- abs == r->abs &&
- swizzle == r->swizzle &&
- !reladdr && !r->reladdr &&
- memcmp(&fixed_hw_reg, &r->fixed_hw_reg,
- sizeof(fixed_hw_reg)) == 0 &&
- imm.u == r->imm.u);
+ if (inst->has_side_effects())
+ return false;
+
+ if (new_writemask == 0) {
+ /* Don't dead code eliminate instructions that write to the
+ * accumulator as a side-effect. Instead just set the destination
+ * to the null register to free it.
+ */
+ if (inst->writes_accumulator || inst->writes_flag()) {
+ inst->dst = dst_reg(retype(brw_null_reg(), inst->dst.type));
+ } else {
+ inst->remove();
+ }
+
+ return true;
+ } else if (inst->dst.writemask != new_writemask) {
+ switch (inst->opcode) {
+ case SHADER_OPCODE_TXF_CMS:
+ case SHADER_OPCODE_GEN4_SCRATCH_READ:
+ case VS_OPCODE_PULL_CONSTANT_LOAD:
+ case VS_OPCODE_PULL_CONSTANT_LOAD_GEN7:
+ break;
+ default:
+ /* Do not set a writemask on Gen6 for math instructions, those are
+ * executed using align1 mode that does not support a destination mask.
+ */
+ if (!(brw->gen == 6 && inst->is_math()) && !inst->is_tex()) {
+ inst->dst.writemask = new_writemask;
+ return true;
+ }
+ }
+ }
+
+ return false;
}
/**
- * Must be called after calculate_live_intervales() to remove unused
+ * Must be called after calculate_live_intervals() to remove unused
* writes to registers -- register allocation will fail otherwise
* because something deffed but not used won't be considered to
* interfere with other regs.
vec4_visitor::dead_code_eliminate()
{
bool progress = false;
- int pc = 0;
+ int pc = -1;
calculate_live_intervals();
- foreach_list_safe(node, &this->instructions) {
- vec4_instruction *inst = (vec4_instruction *)node;
+ foreach_in_list_safe(vec4_instruction, inst, &instructions) {
+ pc++;
+
+ bool inst_writes_flag = false;
+ if (inst->dst.file != GRF) {
+ if (inst->dst.is_null() && inst->writes_flag()) {
+ inst_writes_flag = true;
+ } else {
+ continue;
+ }
+ }
if (inst->dst.file == GRF) {
- assert(this->virtual_grf_end[inst->dst.reg] >= pc);
- if (this->virtual_grf_end[inst->dst.reg] == pc) {
- inst->remove();
- progress = true;
+ int write_mask = inst->dst.writemask;
+
+ for (int c = 0; c < 4; c++) {
+ if (write_mask & (1 << c)) {
+ assert(this->virtual_grf_end[inst->dst.reg * 4 + c] >= pc);
+ if (this->virtual_grf_end[inst->dst.reg * 4 + c] == pc) {
+ write_mask &= ~(1 << c);
+ }
+ }
+ }
+
+ progress = try_eliminate_instruction(inst, write_mask, brw) ||
+ progress;
+ }
+
+ if (inst->predicate || inst->prev == NULL)
+ continue;
+
+ int dead_channels;
+ if (inst_writes_flag) {
+/* Arbitrarily chosen, other than not being an xyzw writemask. */
+#define FLAG_WRITEMASK (1 << 5)
+ dead_channels = inst->reads_flag() ? 0 : FLAG_WRITEMASK;
+ } else {
+ dead_channels = inst->dst.writemask;
+
+ for (int i = 0; i < 3; i++) {
+ if (inst->src[i].file != GRF ||
+ inst->src[i].reg != inst->dst.reg)
+ continue;
+
+ for (int j = 0; j < 4; j++) {
+ int swiz = BRW_GET_SWZ(inst->src[i].swizzle, j);
+ dead_channels &= ~(1 << swiz);
+ }
}
}
- pc++;
+ for (exec_node *node = inst->prev, *prev = node->prev;
+ prev != NULL && dead_channels != 0;
+ node = prev, prev = prev->prev) {
+ vec4_instruction *scan_inst = (vec4_instruction *)node;
+
+ if (scan_inst->is_control_flow())
+ break;
+
+ if (inst_writes_flag) {
+ if (scan_inst->dst.is_null() && scan_inst->writes_flag()) {
+ scan_inst->remove();
+ progress = true;
+ continue;
+ } else if (scan_inst->reads_flag()) {
+ break;
+ }
+ }
+
+ if (inst->dst.file == scan_inst->dst.file &&
+ inst->dst.reg == scan_inst->dst.reg &&
+ inst->dst.reg_offset == scan_inst->dst.reg_offset) {
+ int new_writemask = scan_inst->dst.writemask & ~dead_channels;
+
+ progress = try_eliminate_instruction(scan_inst, new_writemask, brw) ||
+ progress;
+ }
+
+ for (int i = 0; i < 3; i++) {
+ if (scan_inst->src[i].file != inst->dst.file ||
+ scan_inst->src[i].reg != inst->dst.reg)
+ continue;
+
+ for (int j = 0; j < 4; j++) {
+ int swiz = BRW_GET_SWZ(scan_inst->src[i].swizzle, j);
+ dead_channels &= ~(1 << swiz);
+ }
+ }
+ }
}
if (progress)
- live_intervals_valid = false;
+ invalidate_live_intervals();
return progress;
}
* vector. The goal is to make elimination of unused uniform
* components easier later.
*/
- foreach_list(node, &this->instructions) {
- vec4_instruction *inst = (vec4_instruction *)node;
-
+ foreach_in_list(vec4_instruction, inst, &instructions) {
for (int i = 0 ; i < 3; i++) {
if (inst->src[i].file != UNIFORM)
continue;
* expect unused vector elements when we've moved array access out
* to pull constants, and from some GLSL code generators like wine.
*/
- foreach_list(node, &this->instructions) {
- vec4_instruction *inst = (vec4_instruction *)node;
-
+ foreach_in_list(vec4_instruction, inst, &instructions) {
for (int i = 0 ; i < 3; i++) {
if (inst->src[i].file != UNIFORM)
continue;
* push constants.
*/
for (int src = 0; src < uniforms; src++) {
+ assert(src < uniform_array_size);
int size = this->uniform_vector_size[src];
if (!uniform_used[src]) {
/* Move the references to the data */
for (int j = 0; j < size; j++) {
- prog_data->param[dst * 4 + new_chan[src] + j] =
- prog_data->param[src * 4 + j];
+ stage_prog_data->param[dst * 4 + new_chan[src] + j] =
+ stage_prog_data->param[src * 4 + j];
}
this->uniform_vector_size[dst] += size;
this->uniforms = new_uniform_count;
/* Now, update the instructions for our repacked uniforms. */
- foreach_list(node, &this->instructions) {
- vec4_instruction *inst = (vec4_instruction *)node;
-
+ foreach_in_list(vec4_instruction, inst, &instructions) {
for (int i = 0 ; i < 3; i++) {
int src = inst->src[i].reg;
}
}
-bool
-src_reg::is_zero() const
-{
- if (file != IMM)
- return false;
-
- if (type == BRW_REGISTER_TYPE_F) {
- return imm.f == 0.0;
- } else {
- return imm.i == 0;
- }
-}
-
-bool
-src_reg::is_one() const
-{
- if (file != IMM)
- return false;
-
- if (type == BRW_REGISTER_TYPE_F) {
- return imm.f == 1.0;
- } else {
- return imm.i == 1;
- }
-}
-
/**
* Does algebraic optimizations (0 * a = 0, 1 * a = a, a + 0 = a).
*
{
bool progress = false;
- foreach_list(node, &this->instructions) {
- vec4_instruction *inst = (vec4_instruction *)node;
-
+ foreach_in_list(vec4_instruction, inst, &instructions) {
switch (inst->opcode) {
case BRW_OPCODE_ADD:
if (inst->src[1].is_zero()) {
inst->src[0] = src_reg(0u);
break;
default:
- assert(!"not reached");
- inst->src[0] = src_reg(0.0f);
- break;
+ unreachable("not reached");
}
inst->src[1] = src_reg();
progress = true;
}
if (progress)
- this->live_intervals_valid = false;
+ invalidate_live_intervals();
return progress;
}
/* Only allow 32 registers (256 uniform components) as push constants,
* which is the limit on gen6.
+ *
+ * If changing this value, note the limitation about total_regs in
+ * brw_curbe.c.
*/
int max_uniform_components = 32 * 8;
if (this->uniforms * 4 <= max_uniform_components)
pull_constant_loc[i / 4] = -1;
if (i >= max_uniform_components) {
- const float **values = &prog_data->param[i];
+ const gl_constant_value **values = &stage_prog_data->param[i];
/* Try to find an existing copy of this uniform in the pull
* constants if it was part of an array access already.
*/
- for (unsigned int j = 0; j < prog_data->nr_pull_params; j += 4) {
+ for (unsigned int j = 0; j < stage_prog_data->nr_pull_params; j += 4) {
int matches;
for (matches = 0; matches < 4; matches++) {
- if (prog_data->pull_param[j + matches] != values[matches])
+ if (stage_prog_data->pull_param[j + matches] != values[matches])
break;
}
}
if (pull_constant_loc[i / 4] == -1) {
- assert(prog_data->nr_pull_params % 4 == 0);
- pull_constant_loc[i / 4] = prog_data->nr_pull_params / 4;
+ assert(stage_prog_data->nr_pull_params % 4 == 0);
+ pull_constant_loc[i / 4] = stage_prog_data->nr_pull_params / 4;
for (int j = 0; j < 4; j++) {
- prog_data->pull_param[prog_data->nr_pull_params++] = values[j];
+ stage_prog_data->pull_param[stage_prog_data->nr_pull_params++] =
+ values[j];
}
}
}
/* Now actually rewrite usage of the things we've moved to pull
* constants.
*/
- foreach_list_safe(node, &this->instructions) {
- vec4_instruction *inst = (vec4_instruction *)node;
-
+ foreach_in_list_safe(vec4_instruction, inst, &instructions) {
for (int i = 0 ; i < 3; i++) {
if (inst->src[i].file != UNIFORM ||
pull_constant_loc[inst->src[i].reg] == -1)
vec4_instruction *last_mrf_write[BRW_MAX_GRF];
uint8_t mrf_channels_written[BRW_MAX_GRF];
- cfg_t cfg(this);
+ calculate_cfg();
assert(prog_data->total_grf ||
!"Must be called after register allocation");
- for (int i = 0; i < cfg.num_blocks; i++) {
- bblock_t *bblock = cfg.blocks[i];
- vec4_instruction *inst;
+ for (int i = 0; i < cfg->num_blocks; i++) {
+ bblock_t *bblock = cfg->blocks[i];
memset(last_grf_write, 0, sizeof(last_grf_write));
memset(last_mrf_write, 0, sizeof(last_mrf_write));
- for (inst = (vec4_instruction *)bblock->start;
- inst != (vec4_instruction *)bblock->end->next;
- inst = (vec4_instruction *)inst->next) {
+ foreach_inst_in_block (vec4_instruction, inst, bblock) {
/* If we read from a register that we were doing dependency control
* on, don't do dependency control across the read.
*/
continue;
}
+ /* Dependency control does not work well over math instructions.
+ */
+ if (inst->is_math()) {
+ memset(last_grf_write, 0, sizeof(last_grf_write));
+ memset(last_mrf_write, 0, sizeof(last_mrf_write));
+ continue;
+ }
+
/* Now, see if we can do dependency control for this instruction
* against a previous one writing to its destination.
*/
calculate_live_intervals();
- foreach_list_safe(node, &this->instructions) {
- vec4_instruction *inst = (vec4_instruction *)node;
-
+ foreach_in_list_safe(vec4_instruction, inst, &instructions) {
int ip = next_ip;
next_ip++;
/* Can't coalesce this GRF if someone else was going to
* read it later.
*/
- if (this->virtual_grf_end[inst->src[0].reg] > ip)
+ if (this->virtual_grf_end[inst->src[0].reg * 4 + 0] > ip ||
+ this->virtual_grf_end[inst->src[0].reg * 4 + 1] > ip ||
+ this->virtual_grf_end[inst->src[0].reg * 4 + 2] > ip ||
+ this->virtual_grf_end[inst->src[0].reg * 4 + 3] > ip)
continue;
/* We need to check interference with the final destination between this
}
if (progress)
- live_intervals_valid = false;
+ invalidate_live_intervals();
return progress;
}
/* Check that the instructions are compatible with the registers we're trying
* to split.
*/
- foreach_list(node, &this->instructions) {
- vec4_instruction *inst = (vec4_instruction *)node;
-
+ foreach_in_list(vec4_instruction, inst, &instructions) {
/* If there's a SEND message loading from a GRF on gen7+, it needs to be
* contiguous.
*/
this->virtual_grf_sizes[i] = 1;
}
- foreach_list(node, &this->instructions) {
- vec4_instruction *inst = (vec4_instruction *)node;
-
+ foreach_in_list(vec4_instruction, inst, &instructions) {
if (inst->dst.file == GRF && split_grf[inst->dst.reg] &&
inst->dst.reg_offset != 0) {
inst->dst.reg = (new_virtual_grf[inst->dst.reg] +
}
}
}
- this->live_intervals_valid = false;
+ invalidate_live_intervals();
}
void
vec4_visitor::dump_instruction(backend_instruction *be_inst)
+{
+ dump_instruction(be_inst, stderr);
+}
+
+void
+vec4_visitor::dump_instruction(backend_instruction *be_inst, FILE *file)
{
vec4_instruction *inst = (vec4_instruction *)be_inst;
- printf("%s ", brw_instruction_name(inst->opcode));
+ if (inst->predicate) {
+ fprintf(file, "(%cf0) ",
+ inst->predicate_inverse ? '-' : '+');
+ }
+
+ fprintf(file, "%s", brw_instruction_name(inst->opcode));
+ if (inst->conditional_mod) {
+ fprintf(file, "%s", conditional_modifier[inst->conditional_mod]);
+ }
+ fprintf(file, " ");
switch (inst->dst.file) {
case GRF:
- printf("vgrf%d.%d", inst->dst.reg, inst->dst.reg_offset);
+ fprintf(file, "vgrf%d.%d", inst->dst.reg, inst->dst.reg_offset);
break;
case MRF:
- printf("m%d", inst->dst.reg);
+ fprintf(file, "m%d", inst->dst.reg);
+ break;
+ case HW_REG:
+ if (inst->dst.fixed_hw_reg.file == BRW_ARCHITECTURE_REGISTER_FILE) {
+ switch (inst->dst.fixed_hw_reg.nr) {
+ case BRW_ARF_NULL:
+ fprintf(file, "null");
+ break;
+ case BRW_ARF_ADDRESS:
+ fprintf(file, "a0.%d", inst->dst.fixed_hw_reg.subnr);
+ break;
+ case BRW_ARF_ACCUMULATOR:
+ fprintf(file, "acc%d", inst->dst.fixed_hw_reg.subnr);
+ break;
+ case BRW_ARF_FLAG:
+ fprintf(file, "f%d.%d", inst->dst.fixed_hw_reg.nr & 0xf,
+ inst->dst.fixed_hw_reg.subnr);
+ break;
+ default:
+ fprintf(file, "arf%d.%d", inst->dst.fixed_hw_reg.nr & 0xf,
+ inst->dst.fixed_hw_reg.subnr);
+ break;
+ }
+ } else {
+ fprintf(file, "hw_reg%d", inst->dst.fixed_hw_reg.nr);
+ }
+ if (inst->dst.fixed_hw_reg.subnr)
+ fprintf(file, "+%d", inst->dst.fixed_hw_reg.subnr);
break;
case BAD_FILE:
- printf("(null)");
+ fprintf(file, "(null)");
break;
default:
- printf("???");
+ fprintf(file, "???");
break;
}
if (inst->dst.writemask != WRITEMASK_XYZW) {
- printf(".");
+ fprintf(file, ".");
if (inst->dst.writemask & 1)
- printf("x");
+ fprintf(file, "x");
if (inst->dst.writemask & 2)
- printf("y");
+ fprintf(file, "y");
if (inst->dst.writemask & 4)
- printf("z");
+ fprintf(file, "z");
if (inst->dst.writemask & 8)
- printf("w");
+ fprintf(file, "w");
}
- printf(", ");
+ fprintf(file, ":%s, ", brw_reg_type_letters(inst->dst.type));
- for (int i = 0; i < 3; i++) {
+ for (int i = 0; i < 3 && inst->src[i].file != BAD_FILE; i++) {
+ if (inst->src[i].negate)
+ fprintf(file, "-");
+ if (inst->src[i].abs)
+ fprintf(file, "|");
switch (inst->src[i].file) {
case GRF:
- printf("vgrf%d", inst->src[i].reg);
+ fprintf(file, "vgrf%d", inst->src[i].reg);
break;
case ATTR:
- printf("attr%d", inst->src[i].reg);
+ fprintf(file, "attr%d", inst->src[i].reg);
break;
case UNIFORM:
- printf("u%d", inst->src[i].reg);
+ fprintf(file, "u%d", inst->src[i].reg);
break;
case IMM:
switch (inst->src[i].type) {
case BRW_REGISTER_TYPE_F:
- printf("%fF", inst->src[i].imm.f);
+ fprintf(file, "%fF", inst->src[i].fixed_hw_reg.dw1.f);
break;
case BRW_REGISTER_TYPE_D:
- printf("%dD", inst->src[i].imm.i);
+ fprintf(file, "%dD", inst->src[i].fixed_hw_reg.dw1.d);
break;
case BRW_REGISTER_TYPE_UD:
- printf("%uU", inst->src[i].imm.u);
+ fprintf(file, "%uU", inst->src[i].fixed_hw_reg.dw1.ud);
break;
default:
- printf("???");
+ fprintf(file, "???");
break;
}
break;
+ case HW_REG:
+ if (inst->src[i].fixed_hw_reg.negate)
+ fprintf(file, "-");
+ if (inst->src[i].fixed_hw_reg.abs)
+ fprintf(file, "|");
+ if (inst->src[i].fixed_hw_reg.file == BRW_ARCHITECTURE_REGISTER_FILE) {
+ switch (inst->src[i].fixed_hw_reg.nr) {
+ case BRW_ARF_NULL:
+ fprintf(file, "null");
+ break;
+ case BRW_ARF_ADDRESS:
+ fprintf(file, "a0.%d", inst->src[i].fixed_hw_reg.subnr);
+ break;
+ case BRW_ARF_ACCUMULATOR:
+ fprintf(file, "acc%d", inst->src[i].fixed_hw_reg.subnr);
+ break;
+ case BRW_ARF_FLAG:
+ fprintf(file, "f%d.%d", inst->src[i].fixed_hw_reg.nr & 0xf,
+ inst->src[i].fixed_hw_reg.subnr);
+ break;
+ default:
+ fprintf(file, "arf%d.%d", inst->src[i].fixed_hw_reg.nr & 0xf,
+ inst->src[i].fixed_hw_reg.subnr);
+ break;
+ }
+ } else {
+ fprintf(file, "hw_reg%d", inst->src[i].fixed_hw_reg.nr);
+ }
+ if (inst->src[i].fixed_hw_reg.subnr)
+ fprintf(file, "+%d", inst->src[i].fixed_hw_reg.subnr);
+ if (inst->src[i].fixed_hw_reg.abs)
+ fprintf(file, "|");
+ break;
case BAD_FILE:
- printf("(null)");
+ fprintf(file, "(null)");
break;
default:
- printf("???");
+ fprintf(file, "???");
break;
}
- if (inst->src[i].reg_offset)
- printf(".%d", inst->src[i].reg_offset);
+ /* Don't print .0; and only VGRFs have reg_offsets and sizes */
+ if (inst->src[i].reg_offset != 0 &&
+ inst->src[i].file == GRF &&
+ virtual_grf_sizes[inst->src[i].reg] != 1)
+ fprintf(file, ".%d", inst->src[i].reg_offset);
+
+ if (inst->src[i].file != IMM) {
+ static const char *chans[4] = {"x", "y", "z", "w"};
+ fprintf(file, ".");
+ for (int c = 0; c < 4; c++) {
+ fprintf(file, "%s", chans[BRW_GET_SWZ(inst->src[i].swizzle, c)]);
+ }
+ }
+
+ if (inst->src[i].abs)
+ fprintf(file, "|");
- static const char *chans[4] = {"x", "y", "z", "w"};
- printf(".");
- for (int c = 0; c < 4; c++) {
- printf("%s", chans[BRW_GET_SWZ(inst->src[i].swizzle, c)]);
+ if (inst->src[i].file != IMM) {
+ fprintf(file, ":%s", brw_reg_type_letters(inst->src[i].type));
}
- if (i < 3)
- printf(", ");
+ if (i < 2 && inst->src[i + 1].file != BAD_FILE)
+ fprintf(file, ", ");
}
- printf("\n");
+ fprintf(file, "\n");
+}
+
+
+static inline struct brw_reg
+attribute_to_hw_reg(int attr, bool interleaved)
+{
+ if (interleaved)
+ return stride(brw_vec4_grf(attr / 2, (attr % 2) * 4), 0, 4, 1);
+ else
+ return brw_vec8_grf(attr, 0);
}
+
/**
* Replace each register of type ATTR in this->instructions with a reference
* to a fixed HW register.
+ *
+ * If interleaved is true, then each attribute takes up half a register, with
+ * register N containing attribute 2*N in its first half and attribute 2*N+1
+ * in its second half (this corresponds to the payload setup used by geometry
+ * shaders in "single" or "dual instanced" dispatch mode). If interleaved is
+ * false, then each attribute takes up a whole register, with register N
+ * containing attribute N (this corresponds to the payload setup used by
+ * vertex shaders, and by geometry shaders in "dual object" dispatch mode).
*/
void
-vec4_visitor::lower_attributes_to_hw_regs(const int *attribute_map)
+vec4_visitor::lower_attributes_to_hw_regs(const int *attribute_map,
+ bool interleaved)
{
- foreach_list(node, &this->instructions) {
- vec4_instruction *inst = (vec4_instruction *)node;
-
+ foreach_in_list(vec4_instruction, inst, &instructions) {
/* We have to support ATTR as a destination for GL_FIXED fixup. */
if (inst->dst.file == ATTR) {
int grf = attribute_map[inst->dst.reg + inst->dst.reg_offset];
*/
assert(grf != 0);
- struct brw_reg reg = brw_vec8_grf(grf, 0);
+ struct brw_reg reg = attribute_to_hw_reg(grf, interleaved);
reg.type = inst->dst.type;
reg.dw1.bits.writemask = inst->dst.writemask;
*/
assert(grf != 0);
- struct brw_reg reg = brw_vec8_grf(grf, 0);
+ struct brw_reg reg = attribute_to_hw_reg(grf, interleaved);
reg.dw1.bits.swizzle = inst->src[i].swizzle;
reg.type = inst->src[i].type;
if (inst->src[i].abs)
nr_attributes++;
}
- lower_attributes_to_hw_regs(attribute_map);
+ lower_attributes_to_hw_regs(attribute_map, false /* interleaved */);
/* The BSpec says we always have to read at least one thing from
* the VF, and it appears that the hardware wedges otherwise.
int
vec4_visitor::setup_uniforms(int reg)
{
- prog_data->dispatch_grf_start_reg = reg;
+ prog_data->base.dispatch_grf_start_reg = reg;
/* The pre-gen6 VS requires that some push constants get loaded no
* matter what, or the GPU would hang.
*/
if (brw->gen < 6 && this->uniforms == 0) {
+ assert(this->uniforms < this->uniform_array_size);
this->uniform_vector_size[this->uniforms] = 1;
- prog_data->param = reralloc(NULL, prog_data->param, const float *, 4);
+ stage_prog_data->param =
+ reralloc(NULL, stage_prog_data->param, const gl_constant_value *, 4);
for (unsigned int i = 0; i < 4; i++) {
unsigned int slot = this->uniforms * 4 + i;
- static float zero = 0.0;
- prog_data->param[slot] = &zero;
+ static gl_constant_value zero = { .f = 0.0 };
+ stage_prog_data->param[slot] = &zero;
}
this->uniforms++;
reg += ALIGN(uniforms, 2) / 2;
}
- prog_data->nr_params = this->uniforms * 4;
+ stage_prog_data->nr_params = this->uniforms * 4;
- prog_data->curb_read_length = reg - prog_data->dispatch_grf_start_reg;
+ prog_data->curb_read_length = reg - prog_data->base.dispatch_grf_start_reg;
return reg;
}
*/
emit(ADD(diff, src_reg(diff), src_reg(-2u)));
- emit_shader_time_write(ST_VS, src_reg(diff));
- emit_shader_time_write(ST_VS_WRITTEN, src_reg(1u));
+ emit_shader_time_write(st_base, src_reg(diff));
+ emit_shader_time_write(st_written, src_reg(1u));
emit(BRW_OPCODE_ELSE);
- emit_shader_time_write(ST_VS_RESET, src_reg(1u));
+ emit_shader_time_write(st_reset, src_reg(1u));
emit(BRW_OPCODE_ENDIF);
}
if (INTEL_DEBUG & DEBUG_SHADER_TIME)
emit_shader_time_begin();
+ assign_common_binding_table_offsets(0);
+
emit_prolog();
/* Generate VS IR for main(). (the visitor only descends into
* functions called "main").
*/
if (shader) {
- visit_instructions(shader->ir);
+ visit_instructions(shader->base.ir);
} else {
emit_program_code();
}
base_ir = NULL;
- if (key->userclip_active && !key->uses_clip_distance)
+ if (key->userclip_active && !prog->UsesClipDistanceOut)
setup_uniform_clipplane_values();
emit_thread_end();
move_push_constants_to_pull_constants();
split_virtual_grfs();
+ const char *stage_name = stage == MESA_SHADER_GEOMETRY ? "gs" : "vs";
+
+#define OPT(pass, args...) do { \
+ pass_num++; \
+ bool this_progress = pass(args); \
+ \
+ if (unlikely(INTEL_DEBUG & DEBUG_OPTIMIZER) && this_progress) { \
+ char filename[64]; \
+ snprintf(filename, 64, "%s-%04d-%02d-%02d-" #pass, \
+ stage_name, shader_prog->Name, iteration, pass_num); \
+ \
+ backend_visitor::dump_instructions(filename); \
+ } \
+ \
+ progress = progress || this_progress; \
+ } while (false)
+
+
+ if (unlikely(INTEL_DEBUG & DEBUG_OPTIMIZER)) {
+ char filename[64];
+ snprintf(filename, 64, "%s-%04d-00-start",
+ stage_name, shader_prog->Name);
+
+ backend_visitor::dump_instructions(filename);
+ }
+
bool progress;
+ int iteration = 0;
do {
progress = false;
- progress = dead_code_eliminate() || progress;
- progress = opt_copy_propagation() || progress;
- progress = opt_algebraic() || progress;
- progress = opt_register_coalesce() || progress;
+ iteration++;
+ int pass_num = 0;
+
+ OPT(dead_code_eliminate);
+ OPT(dead_control_flow_eliminate, this);
+ OPT(opt_copy_propagation);
+ OPT(opt_algebraic);
+ OPT(opt_cse);
+ OPT(opt_register_coalesce);
} while (progress);
while (!reg_allocate()) {
if (failed)
- break;
+ return false;
}
opt_schedule_instructions();
unsigned *final_assembly_size)
{
bool start_busy = false;
- float start_time = 0;
+ double start_time = 0;
if (unlikely(brw->perf_debug)) {
start_busy = (brw->batch.last_bo &&
if (prog)
shader = (brw_shader *) prog->_LinkedShaders[MESA_SHADER_VERTEX];
- if (unlikely(INTEL_DEBUG & DEBUG_VS)) {
- if (prog) {
- printf("GLSL IR for native vertex shader %d:\n", prog->Name);
- _mesa_print_ir(shader->ir, NULL);
- printf("\n\n");
- } else {
- printf("ARB_vertex_program %d for native vertex shader\n",
- c->vp->program.Base.Id);
- _mesa_print_program(&c->vp->program.Base);
- }
- }
+ if (unlikely(INTEL_DEBUG & DEBUG_VS))
+ brw_dump_ir(brw, "vertex", prog, &shader->base, &c->vp->program.Base);
- vec4_vs_visitor v(brw, c, prog_data, prog, shader, mem_ctx);
+ vec4_vs_visitor v(brw, c, prog_data, prog, mem_ctx);
if (!v.run()) {
if (prog) {
prog->LinkStatus = false;
return NULL;
}
- vec4_generator g(brw, prog, &c->vp->program.Base, &prog_data->base, mem_ctx,
- INTEL_DEBUG & DEBUG_VS);
- const unsigned *generated =g.generate_assembly(&v.instructions,
- final_assembly_size);
+ const unsigned *assembly = NULL;
+ vec4_generator g(brw, prog, &c->vp->program.Base, &prog_data->base,
+ mem_ctx, INTEL_DEBUG & DEBUG_VS);
+ assembly = g.generate_assembly(&v.instructions, final_assembly_size);
if (unlikely(brw->perf_debug) && shader) {
if (shader->compiled_once) {
shader->compiled_once = true;
}
- return generated;
-}
-
-
-bool
-brw_vec4_prog_data_compare(const struct brw_vec4_prog_data *a,
- const struct brw_vec4_prog_data *b)
-{
- /* Compare all the struct up to the pointers. */
- if (memcmp(a, b, offsetof(struct brw_vec4_prog_data, param)))
- return false;
-
- if (memcmp(a->param, b->param, a->nr_params * sizeof(void *)))
- return false;
-
- if (memcmp(a->pull_param, b->pull_param, a->nr_pull_params * sizeof(void *)))
- return false;
-
- return true;
+ return assembly;
}
void
-brw_vec4_prog_data_free(const struct brw_vec4_prog_data *prog_data)
+brw_vec4_setup_prog_key_for_precompile(struct gl_context *ctx,
+ struct brw_vec4_prog_key *key,
+ GLuint id, struct gl_program *prog)
{
- ralloc_free((void *)prog_data->param);
- ralloc_free((void *)prog_data->pull_param);
+ key->program_string_id = id;
+ key->clamp_vertex_color = ctx->API == API_OPENGL_COMPAT;
+
+ unsigned sampler_count = _mesa_fls(prog->SamplersUsed);
+ for (unsigned i = 0; i < sampler_count; i++) {
+ if (prog->ShadowSamplers & (1 << i)) {
+ /* Assume DEPTH_TEXTURE_MODE is the default: X, X, X, 1 */
+ key->tex.swizzles[i] =
+ MAKE_SWIZZLE4(SWIZZLE_X, SWIZZLE_X, SWIZZLE_X, SWIZZLE_ONE);
+ } else {
+ /* Color sampler: assume no swizzling. */
+ key->tex.swizzles[i] = SWIZZLE_XYZW;
+ }
+ }
}
-
} /* extern "C" */