*/
#include "brw_fs.h"
+#include "brw_cfg.h"
#include "glsl/glsl_types.h"
#include "glsl/ir_optimization.h"
}
this->grf_used = hw_reg_mapping[this->virtual_grf_count];
- foreach_list(node, &this->instructions) {
- fs_inst *inst = (fs_inst *)node;
-
+ foreach_block_and_inst(block, fs_inst, inst, cfg) {
assign_reg(hw_reg_mapping, &inst->dst, reg_width);
- assign_reg(hw_reg_mapping, &inst->src[0], reg_width);
- assign_reg(hw_reg_mapping, &inst->src[1], reg_width);
- assign_reg(hw_reg_mapping, &inst->src[2], reg_width);
+ for (i = 0; i < inst->sources; i++) {
+ assign_reg(hw_reg_mapping, &inst->src[i], reg_width);
+ }
}
if (this->grf_used >= max_grf) {
fail("Ran out of regs on trivial allocator (%d/%d)\n",
this->grf_used, max_grf);
+ } else {
+ this->virtual_grf_count = this->grf_used;
}
}
static void
-brw_alloc_reg_set(struct brw_context *brw, int reg_width)
+brw_alloc_reg_set(struct intel_screen *screen, int reg_width)
{
+ const struct brw_device_info *devinfo = screen->devinfo;
int base_reg_count = BRW_MAX_GRF / reg_width;
int index = reg_width - 1;
int class_count;
int class_sizes[BRW_MAX_MRF];
- if (brw->gen >= 7) {
+ if (devinfo->gen >= 7) {
for (class_count = 0; class_count < MAX_SAMPLER_MESSAGE_SIZE;
class_count++)
class_sizes[class_count] = class_count + 1;
ra_reg_count += base_reg_count - (class_sizes[i] - 1);
}
- uint8_t *ra_reg_to_grf = ralloc_array(brw, uint8_t, ra_reg_count);
- struct ra_regs *regs = ra_alloc_reg_set(brw, ra_reg_count);
- if (brw->gen >= 6)
+ uint8_t *ra_reg_to_grf = ralloc_array(screen, uint8_t, ra_reg_count);
+ struct ra_regs *regs = ra_alloc_reg_set(screen, ra_reg_count);
+ if (devinfo->gen >= 6)
ra_set_allocate_round_robin(regs);
- int *classes = ralloc_array(brw, int, class_count);
+ int *classes = ralloc_array(screen, int, class_count);
int aligned_pairs_class = -1;
/* Now, add the registers to their classes, and add the conflicts
/* Add a special class for aligned pairs, which we'll put delta_x/y
* in on gen5 so that we can do PLN.
*/
- if (brw->has_pln && reg_width == 1 && brw->gen < 6) {
+ if (devinfo->has_pln && reg_width == 1 && devinfo->gen < 6) {
aligned_pairs_class = ra_alloc_reg_class(regs);
for (int i = 0; i < pairs_reg_count; i++) {
ra_set_finalize(regs, NULL);
- brw->wm.reg_sets[index].regs = regs;
- for (unsigned i = 0; i < ARRAY_SIZE(brw->wm.reg_sets[index].classes); i++)
- brw->wm.reg_sets[index].classes[i] = -1;
+ screen->wm_reg_sets[index].regs = regs;
+ for (unsigned i = 0; i < ARRAY_SIZE(screen->wm_reg_sets[index].classes); i++)
+ screen->wm_reg_sets[index].classes[i] = -1;
for (int i = 0; i < class_count; i++)
- brw->wm.reg_sets[index].classes[class_sizes[i] - 1] = classes[i];
- brw->wm.reg_sets[index].ra_reg_to_grf = ra_reg_to_grf;
- brw->wm.reg_sets[index].aligned_pairs_class = aligned_pairs_class;
+ screen->wm_reg_sets[index].classes[class_sizes[i] - 1] = classes[i];
+ screen->wm_reg_sets[index].ra_reg_to_grf = ra_reg_to_grf;
+ screen->wm_reg_sets[index].aligned_pairs_class = aligned_pairs_class;
}
void
-brw_fs_alloc_reg_sets(struct brw_context *brw)
+brw_fs_alloc_reg_sets(struct intel_screen *screen)
{
- brw_alloc_reg_set(brw, 1);
- brw_alloc_reg_set(brw, 2);
+ brw_alloc_reg_set(screen, 1);
+ brw_alloc_reg_set(screen, 2);
}
-int
-count_to_loop_end(fs_inst *do_inst)
+static int
+count_to_loop_end(const bblock_t *block)
{
+ if (block->end()->opcode == BRW_OPCODE_WHILE)
+ return block->end_ip;
+
int depth = 1;
- int ip = 1;
- for (fs_inst *inst = (fs_inst *)do_inst->next;
+ /* Skip the first block, since we don't want to count the do the calling
+ * function found.
+ */
+ for (block = block->next();
depth > 0;
- inst = (fs_inst *)inst->next) {
- switch (inst->opcode) {
- case BRW_OPCODE_DO:
+ block = block->next()) {
+ if (block->start()->opcode == BRW_OPCODE_DO)
depth++;
- break;
- case BRW_OPCODE_WHILE:
+ if (block->end()->opcode == BRW_OPCODE_WHILE) {
depth--;
- break;
- default:
- break;
+ if (depth == 0)
+ return block->end_ip;
}
- ip++;
}
- return ip;
+ unreachable("not reached");
}
/**
*
* The layout of the payload registers is:
*
- * 0..nr_payload_regs-1: fixed function setup (including bary coordinates).
- * nr_payload_regs..nr_payload_regs+curb_read_lengh-1: uniform data
- * nr_payload_regs+curb_read_lengh..first_non_payload_grf-1: setup coefficients.
+ * 0..payload.num_regs-1: fixed function setup (including bary coordinates).
+ * payload.num_regs..payload.num_regs+curb_read_lengh-1: uniform data
+ * payload.num_regs+curb_read_lengh..first_non_payload_grf-1: setup coefficients.
*
* And we have payload_node_count nodes covering these registers in order
* (note that in SIMD16, a node is two registers).
int payload_last_use_ip[payload_node_count];
memset(payload_last_use_ip, 0, sizeof(payload_last_use_ip));
int ip = 0;
- foreach_list(node, &this->instructions) {
- fs_inst *inst = (fs_inst *)node;
-
+ foreach_block_and_inst(block, fs_inst, inst, cfg) {
switch (inst->opcode) {
case BRW_OPCODE_DO:
loop_depth++;
* the end now.
*/
if (loop_depth == 1)
- loop_end_ip = ip + count_to_loop_end(inst);
+ loop_end_ip = count_to_loop_end(block);
break;
case BRW_OPCODE_WHILE:
loop_depth--;
* assign_curbe_setup(), and interpolation uses fixed hardware regs from
* the start (see interp_reg()).
*/
- for (int i = 0; i < 3; i++) {
+ for (int i = 0; i < inst->sources; i++) {
if (inst->src[i].file == HW_REG &&
inst->src[i].fixed_hw_reg.file == BRW_GENERAL_REGISTER_FILE) {
int node_nr = inst->src[i].fixed_hw_reg.nr / reg_width;
memset(mrf_used, 0, BRW_MAX_MRF * sizeof(bool));
- foreach_list(node, &this->instructions) {
- fs_inst *inst = (fs_inst *)node;
-
+ foreach_block_and_inst(block, fs_inst, inst, cfg) {
if (inst->dst.file == MRF) {
int reg = inst->dst.reg & ~BRW_MRF_COMPR4;
mrf_used[reg] = true;
bool
fs_visitor::assign_regs(bool allow_spilling)
{
+ struct intel_screen *screen = brw->intelScreen;
/* Most of this allocation was written for a reg_width of 1
* (dispatch_width == 8). In extending to SIMD16, the code was
* left in place and it was converted to have the hardware
int hw_reg_mapping[this->virtual_grf_count];
int payload_node_count = (ALIGN(this->first_non_payload_grf, reg_width) /
reg_width);
- int rsi = reg_width - 1; /* Which brw->wm.reg_sets[] to use */
+ int rsi = reg_width - 1; /* Which screen->wm_reg_sets[] to use */
calculate_live_intervals();
int node_count = this->virtual_grf_count;
int first_mrf_hack_node = node_count;
if (brw->gen >= 7)
node_count += BRW_MAX_GRF - GEN7_MRF_HACK_START;
- struct ra_graph *g = ra_alloc_interference_graph(brw->wm.reg_sets[rsi].regs,
+ struct ra_graph *g = ra_alloc_interference_graph(screen->wm_reg_sets[rsi].regs,
node_count);
for (int i = 0; i < this->virtual_grf_count; i++) {
unsigned size = this->virtual_grf_sizes[i];
int c;
- assert(size <= ARRAY_SIZE(brw->wm.reg_sets[rsi].classes) &&
+ assert(size <= ARRAY_SIZE(screen->wm_reg_sets[rsi].classes) &&
"Register allocation relies on split_virtual_grfs()");
- c = brw->wm.reg_sets[rsi].classes[size - 1];
+ c = screen->wm_reg_sets[rsi].classes[size - 1];
/* Special case: on pre-GEN6 hardware that supports PLN, the
* second operand of a PLN instruction needs to be an
* any other interpolation modes). So all we need to do is find
* that register and set it to the appropriate class.
*/
- if (brw->wm.reg_sets[rsi].aligned_pairs_class >= 0 &&
+ if (screen->wm_reg_sets[rsi].aligned_pairs_class >= 0 &&
+ this->delta_x[BRW_WM_PERSPECTIVE_PIXEL_BARYCENTRIC].file == GRF &&
this->delta_x[BRW_WM_PERSPECTIVE_PIXEL_BARYCENTRIC].reg == i) {
- c = brw->wm.reg_sets[rsi].aligned_pairs_class;
+ c = screen->wm_reg_sets[rsi].aligned_pairs_class;
}
ra_set_node_class(g, i, c);
}
}
- if (!ra_allocate_no_spills(g)) {
+ if (!ra_allocate(g)) {
/* Failed to allocate registers. Spill a reg, and the caller will
* loop back into here to try again.
*/
if (reg == -1) {
fail("no register to spill:\n");
- dump_instructions();
+ dump_instructions(NULL);
} else if (allow_spilling) {
spill_reg(reg);
}
for (int i = 0; i < this->virtual_grf_count; i++) {
int reg = ra_get_node_reg(g, i);
- hw_reg_mapping[i] = brw->wm.reg_sets[rsi].ra_reg_to_grf[reg] * reg_width;
+ hw_reg_mapping[i] = screen->wm_reg_sets[rsi].ra_reg_to_grf[reg] * reg_width;
this->grf_used = MAX2(this->grf_used,
hw_reg_mapping[i] + this->virtual_grf_sizes[i] *
reg_width);
}
- foreach_list(node, &this->instructions) {
- fs_inst *inst = (fs_inst *)node;
-
+ foreach_block_and_inst(block, fs_inst, inst, cfg) {
assign_reg(hw_reg_mapping, &inst->dst, reg_width);
- assign_reg(hw_reg_mapping, &inst->src[0], reg_width);
- assign_reg(hw_reg_mapping, &inst->src[1], reg_width);
- assign_reg(hw_reg_mapping, &inst->src[2], reg_width);
+ for (int i = 0; i < inst->sources; i++) {
+ assign_reg(hw_reg_mapping, &inst->src[i], reg_width);
+ }
}
+ this->virtual_grf_count = this->grf_used;
+
ralloc_free(g);
return true;
}
void
-fs_visitor::emit_unspill(fs_inst *inst, fs_reg dst, uint32_t spill_offset,
- int count)
+fs_visitor::emit_unspill(bblock_t *block, fs_inst *inst, fs_reg dst,
+ uint32_t spill_offset, int count)
{
for (int i = 0; i < count; i++) {
/* The gen7 descriptor-based offset is 12 bits of HWORD units. */
unspill_inst->base_mrf = 14;
unspill_inst->mlen = 1; /* header contains offset */
}
- inst->insert_before(unspill_inst);
+ inst->insert_before(block, unspill_inst);
- dst.reg_offset++;
+ dst = offset(dst, 1);
spill_offset += dispatch_width * sizeof(float);
}
}
* spill/unspill we'll have to do, and guess that the insides of
* loops run 10 times.
*/
- foreach_list(node, &this->instructions) {
- fs_inst *inst = (fs_inst *)node;
-
- for (unsigned int i = 0; i < 3; i++) {
+ foreach_block_and_inst(block, fs_inst, inst, cfg) {
+ for (unsigned int i = 0; i < inst->sources; i++) {
if (inst->src[i].file == GRF) {
spill_costs[inst->src[i].reg] += loop_scale;
* loading pull constants, so spilling them is unlikely to reduce
* register pressure anyhow.
*/
- if (inst->src[i].smear >= 0) {
+ if (!inst->src[i].is_contiguous()) {
no_spill[inst->src[i].reg] = true;
}
}
if (inst->dst.file == GRF) {
spill_costs[inst->dst.reg] += inst->regs_written * loop_scale;
- if (inst->dst.smear >= 0) {
+ if (!inst->dst.is_contiguous()) {
no_spill[inst->dst.reg] = true;
}
}
{
int reg_size = dispatch_width * sizeof(float);
int size = virtual_grf_sizes[spill_reg];
- unsigned int spill_offset = c->last_scratch;
+ unsigned int spill_offset = last_scratch;
assert(ALIGN(spill_offset, 16) == spill_offset); /* oword read/write req. */
int spill_base_mrf = dispatch_width > 8 ? 13 : 14;
spilled_any_registers = true;
}
- c->last_scratch += size * reg_size;
+ last_scratch += size * reg_size;
/* Generate spill/unspill instructions for the objects being
* spilled. Right now, we spill or unspill the whole thing to a
* virtual grf of the same size. For most instructions, though, we
* could just spill/unspill the GRF being accessed.
*/
- foreach_list(node, &this->instructions) {
- fs_inst *inst = (fs_inst *)node;
-
- for (unsigned int i = 0; i < 3; i++) {
+ foreach_block_and_inst (block, fs_inst, inst, cfg) {
+ for (unsigned int i = 0; i < inst->sources; i++) {
if (inst->src[i].file == GRF &&
inst->src[i].reg == spill_reg) {
int regs_read = inst->regs_read(this, i);
inst->src[i].reg = unspill_dst.reg;
inst->src[i].reg_offset = 0;
- emit_unspill(inst, unspill_dst, subset_spill_offset, regs_read);
+ emit_unspill(block, inst, unspill_dst, subset_spill_offset,
+ regs_read);
}
}
*/
if (inst->predicate || inst->force_uncompressed ||
inst->force_sechalf || inst->dst.subreg_offset) {
- emit_unspill(inst, spill_src, subset_spill_offset,
+ emit_unspill(block, inst, spill_src, subset_spill_offset,
inst->regs_written);
}
fs_inst *spill_inst =
new(mem_ctx) fs_inst(SHADER_OPCODE_GEN4_SCRATCH_WRITE,
reg_null_f, spill_src);
- spill_src.reg_offset++;
+ spill_src = offset(spill_src, 1);
spill_inst->offset = subset_spill_offset + chan * reg_size;
spill_inst->ir = inst->ir;
spill_inst->annotation = inst->annotation;
spill_inst->mlen = 1 + dispatch_width / 8; /* header, value */
spill_inst->base_mrf = spill_base_mrf;
- inst->insert_after(spill_inst);
+ inst->insert_after(block, spill_inst);
}
}
}