*
*/
-#include "brw_fs_cfg.h"
+#include "brw_cfg.h"
#include "brw_fs_live_variables.h"
using namespace brw;
*/
/**
- * Sets up the use[] and def[] arrays.
+ * Sets up the use[] and def[] bitsets.
*
* The basic-block-level live variable analysis needs to know which
* variables get used before they're completely defined, and which
int ip = 0;
for (int b = 0; b < cfg->num_blocks; b++) {
- fs_bblock *block = cfg->blocks[b];
+ bblock_t *block = cfg->blocks[b];
assert(ip == block->start_ip);
if (b > 0)
assert(cfg->blocks[b - 1]->end_ip == ip - 1);
- for (fs_inst *inst = block->start;
+ for (fs_inst *inst = (fs_inst *)block->start;
inst != block->end->next;
inst = (fs_inst *)inst->next) {
if (inst->src[i].file == GRF) {
int reg = inst->src[i].reg;
- if (!bd[b].def[reg])
- bd[b].use[reg] = true;
+ if (!BITSET_TEST(bd[b].def, reg))
+ BITSET_SET(bd[b].use, reg);
}
}
* variable, and thus qualify for being in def[].
*/
if (inst->dst.file == GRF &&
- inst->regs_written() == v->virtual_grf_sizes[inst->dst.reg] &&
- !inst->predicated &&
- !inst->force_uncompressed &&
- !inst->force_sechalf) {
+ inst->regs_written == v->virtual_grf_sizes[inst->dst.reg] &&
+ !inst->is_partial_write()) {
int reg = inst->dst.reg;
- if (!bd[b].use[reg])
- bd[b].def[reg] = true;
+ if (!BITSET_TEST(bd[b].use, reg))
+ BITSET_SET(bd[b].def, reg);
}
ip++;
for (int b = 0; b < cfg->num_blocks; b++) {
/* Update livein */
- for (int i = 0; i < num_vars; i++) {
- if (bd[b].use[i] || (bd[b].liveout[i] && !bd[b].def[i])) {
- if (!bd[b].livein[i]) {
- bd[b].livein[i] = true;
- cont = true;
- }
+ for (int i = 0; i < bitset_words; i++) {
+ BITSET_WORD new_livein = (bd[b].use[i] |
+ (bd[b].liveout[i] & ~bd[b].def[i]));
+ if (new_livein & ~bd[b].livein[i]) {
+ bd[b].livein[i] |= new_livein;
+ cont = true;
}
}
/* Update liveout */
foreach_list(block_node, &cfg->blocks[b]->children) {
- fs_bblock_link *link = (fs_bblock_link *)block_node;
- fs_bblock *block = link->block;
-
- for (int i = 0; i < num_vars; i++) {
- if (bd[block->block_num].livein[i] && !bd[b].liveout[i]) {
- bd[b].liveout[i] = true;
- cont = true;
- }
+ bblock_link *link = (bblock_link *)block_node;
+ bblock_t *block = link->block;
+
+ for (int i = 0; i < bitset_words; i++) {
+ BITSET_WORD new_liveout = (bd[block->block_num].livein[i] &
+ ~bd[b].liveout[i]);
+ if (new_liveout) {
+ bd[b].liveout[i] |= new_liveout;
+ cont = true;
+ }
}
}
}
}
}
-fs_live_variables::fs_live_variables(fs_visitor *v, fs_cfg *cfg)
+fs_live_variables::fs_live_variables(fs_visitor *v, cfg_t *cfg)
: v(v), cfg(cfg)
{
mem_ctx = ralloc_context(cfg->mem_ctx);
- num_vars = v->virtual_grf_next;
+ num_vars = v->virtual_grf_count;
bd = rzalloc_array(mem_ctx, struct block_data, cfg->num_blocks);
- vars = rzalloc_array(mem_ctx, struct var, num_vars);
+ bitset_words = BITSET_WORDS(v->virtual_grf_count);
for (int i = 0; i < cfg->num_blocks; i++) {
- bd[i].def = rzalloc_array(mem_ctx, bool, num_vars);
- bd[i].use = rzalloc_array(mem_ctx, bool, num_vars);
- bd[i].livein = rzalloc_array(mem_ctx, bool, num_vars);
- bd[i].liveout = rzalloc_array(mem_ctx, bool, num_vars);
+ bd[i].def = rzalloc_array(mem_ctx, BITSET_WORD, bitset_words);
+ bd[i].use = rzalloc_array(mem_ctx, BITSET_WORD, bitset_words);
+ bd[i].livein = rzalloc_array(mem_ctx, BITSET_WORD, bitset_words);
+ bd[i].liveout = rzalloc_array(mem_ctx, BITSET_WORD, bitset_words);
}
setup_def_use();
void
fs_visitor::calculate_live_intervals()
{
- int num_vars = this->virtual_grf_next;
+ int num_vars = this->virtual_grf_count;
if (this->live_intervals_valid)
return;
- int *def = ralloc_array(mem_ctx, int, num_vars);
- int *use = ralloc_array(mem_ctx, int, num_vars);
- ralloc_free(this->virtual_grf_def);
- ralloc_free(this->virtual_grf_use);
- this->virtual_grf_def = def;
- this->virtual_grf_use = use;
+ int *start = ralloc_array(mem_ctx, int, num_vars);
+ int *end = ralloc_array(mem_ctx, int, num_vars);
+ ralloc_free(this->virtual_grf_start);
+ ralloc_free(this->virtual_grf_end);
+ this->virtual_grf_start = start;
+ this->virtual_grf_end = end;
for (int i = 0; i < num_vars; i++) {
- def[i] = MAX_INSTRUCTION;
- use[i] = -1;
+ start[i] = MAX_INSTRUCTION;
+ end[i] = -1;
}
/* Start by setting up the intervals with no knowledge of control
for (unsigned int i = 0; i < 3; i++) {
if (inst->src[i].file == GRF) {
int reg = inst->src[i].reg;
-
- use[reg] = ip;
+ int end_ip = ip;
+
+ /* In most cases, a register can be written over safely by the
+ * same instruction that is its last use. For a single
+ * instruction, the sources are dereferenced before writing of the
+ * destination starts (naturally). This gets more complicated for
+ * simd16, because the instruction:
+ *
+ * mov(16) g4<1>F g4<8,8,1>F g6<8,8,1>F
+ *
+ * is actually decoded in hardware as:
+ *
+ * mov(8) g4<1>F g4<8,8,1>F g6<8,8,1>F
+ * mov(8) g5<1>F g5<8,8,1>F g7<8,8,1>F
+ *
+ * Which is safe. However, if we have uniform accesses
+ * happening, we get into trouble:
+ *
+ * mov(8) g4<1>F g4<0,1,0>F g6<8,8,1>F
+ * mov(8) g5<1>F g4<0,1,0>F g7<8,8,1>F
+ *
+ * Now our destination for the first instruction overwrote the
+ * second instruction's src0, and we get garbage for those 8
+ * pixels. There's a similar issue for the pre-gen6
+ * pixel_x/pixel_y, which are registers of 16-bit values and thus
+ * would get stomped by the first decode as well.
+ */
+ if (dispatch_width == 16 && (inst->src[i].smear >= 0 ||
+ (this->pixel_x.reg == reg ||
+ this->pixel_y.reg == reg))) {
+ end_ip++;
+ }
+
+ start[reg] = MIN2(start[reg], ip);
+ end[reg] = MAX2(end[reg], end_ip);
}
}
if (inst->dst.file == GRF) {
int reg = inst->dst.reg;
- def[reg] = MIN2(def[reg], ip);
+ start[reg] = MIN2(start[reg], ip);
+ end[reg] = MAX2(end[reg], ip);
}
ip++;
}
/* Now, extend those intervals using our analysis of control flow. */
- fs_cfg cfg(this);
+ cfg_t cfg(this);
fs_live_variables livevars(this, &cfg);
for (int b = 0; b < cfg.num_blocks; b++) {
for (int i = 0; i < num_vars; i++) {
- if (livevars.bd[b].livein[i]) {
- def[i] = MIN2(def[i], cfg.blocks[b]->start_ip);
- use[i] = MAX2(use[i], cfg.blocks[b]->start_ip);
+ if (BITSET_TEST(livevars.bd[b].livein, i)) {
+ start[i] = MIN2(start[i], cfg.blocks[b]->start_ip);
+ end[i] = MAX2(end[i], cfg.blocks[b]->start_ip);
}
- if (livevars.bd[b].liveout[i]) {
- def[i] = MIN2(def[i], cfg.blocks[b]->end_ip);
- use[i] = MAX2(use[i], cfg.blocks[b]->end_ip);
+ if (BITSET_TEST(livevars.bd[b].liveout, i)) {
+ start[i] = MIN2(start[i], cfg.blocks[b]->end_ip);
+ end[i] = MAX2(end[i], cfg.blocks[b]->end_ip);
}
}
}
bool
fs_visitor::virtual_grf_interferes(int a, int b)
{
- int a_def = this->virtual_grf_def[a], a_use = this->virtual_grf_use[a];
- int b_def = this->virtual_grf_def[b], b_use = this->virtual_grf_use[b];
-
- /* If there's dead code (def but not use), it would break our test
- * unless we consider it used.
- */
- if ((a_use == -1 && a_def != MAX_INSTRUCTION) ||
- (b_use == -1 && b_def != MAX_INSTRUCTION)) {
- return true;
- }
-
- int start = MAX2(a_def, b_def);
- int end = MIN2(a_use, b_use);
-
- /* If the register is used to store 16 values of less than float
- * size (only the case for pixel_[xy]), then we can't allocate
- * another dword-sized thing to that register that would be used in
- * the same instruction. This is because when the GPU decodes (for
- * example):
- *
- * (declare (in ) vec4 gl_FragCoord@0x97766a0)
- * add(16) g6<1>F g6<8,8,1>UW 0.5F { align1 compr };
- *
- * it's actually processed as:
- * add(8) g6<1>F g6<8,8,1>UW 0.5F { align1 };
- * add(8) g7<1>F g6.8<8,8,1>UW 0.5F { align1 sechalf };
- *
- * so our second half values in g6 got overwritten in the first
- * half.
- */
- if (c->dispatch_width == 16 && (this->pixel_x.reg == a ||
- this->pixel_x.reg == b ||
- this->pixel_y.reg == a ||
- this->pixel_y.reg == b)) {
- return start <= end;
- }
-
- return start < end;
+ return !(virtual_grf_end[a] <= virtual_grf_start[b] ||
+ virtual_grf_end[b] <= virtual_grf_start[a]);
}
projects / mesa.git / blobdiff