X-Git-Url: https://git.libre-soc.org/?a=blobdiff_plain;f=src%2Fmesa%2Fdrivers%2Fdri%2Fi965%2Fbrw_fs_live_variables.cpp;h=189a119025d8cd4571432565f06fa57abe3704ec;hb=8776b1b14b229d110f283f5da8c3c36261068ede;hp=099eeaa0bbb18359f82afd998de9d9d887a7f5a0;hpb=4b821a97b5fcdc4c530d5455c43196be09830322;p=mesa.git diff --git a/src/mesa/drivers/dri/i965/brw_fs_live_variables.cpp b/src/mesa/drivers/dri/i965/brw_fs_live_variables.cpp index 099eeaa0bbb..189a119025d 100644 --- a/src/mesa/drivers/dri/i965/brw_fs_live_variables.cpp +++ b/src/mesa/drivers/dri/i965/brw_fs_live_variables.cpp @@ -30,6 +30,8 @@ using namespace brw; +#define MAX_INSTRUCTION (1 << 30) + /** @file brw_fs_live_variables.cpp * * Support for calculating liveness information about virtual GRFs. @@ -46,10 +48,81 @@ using namespace brw; * VGRF is ultimately used or defined. Tracking individual components * allows us to easily assemble this information. * - * See Muchnik's Advanced Compiler Design and Implementation, section + * See Muchnick's Advanced Compiler Design and Implementation, section * 14.1 (p444). */ +void +fs_live_variables::setup_one_read(struct block_data *bd, fs_inst *inst, + int ip, const fs_reg ®) +{ + int var = var_from_reg(reg); + assert(var < num_vars); + + /* 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: + * + * add(16) g4<1>F g4<8,8,1>F g6<8,8,1>F + * + * is actually decoded in hardware as: + * + * add(8) g4<1>F g4<8,8,1>F g6<8,8,1>F + * add(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: + * + * add(8) g4<1>F g4<0,1,0>F g6<8,8,1>F + * add(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. + */ + int end_ip = ip; + if (inst->exec_size == 16 && (reg.stride == 0 || + reg.type == BRW_REGISTER_TYPE_UW || + reg.type == BRW_REGISTER_TYPE_W || + reg.type == BRW_REGISTER_TYPE_UB || + reg.type == BRW_REGISTER_TYPE_B)) { + end_ip++; + } + + start[var] = MIN2(start[var], ip); + end[var] = MAX2(end[var], end_ip); + + /* The use[] bitset marks when the block makes use of a variable (VGRF + * channel) without having completely defined that variable within the + * block. + */ + if (!BITSET_TEST(bd->def, var)) + BITSET_SET(bd->use, var); +} + +void +fs_live_variables::setup_one_write(struct block_data *bd, fs_inst *inst, + int ip, const fs_reg ®) +{ + int var = var_from_reg(reg); + assert(var < num_vars); + + start[var] = MIN2(start[var], ip); + end[var] = MAX2(end[var], ip); + + /* The def[] bitset marks when an initialization in a block completely + * screens off previous updates of that variable (VGRF channel). + */ + if (inst->dst.file == GRF && !inst->is_partial_write()) { + if (!BITSET_TEST(bd->use, var)) + BITSET_SET(bd->def, var); + } +} + /** * Sets up the use[] and def[] bitsets. * @@ -64,49 +137,53 @@ fs_live_variables::setup_def_use() { int ip = 0; - for (int b = 0; b < cfg->num_blocks; b++) { - bblock_t *block = cfg->blocks[b]; - + foreach_block (block, cfg) { assert(ip == block->start_ip); - if (b > 0) - assert(cfg->blocks[b - 1]->end_ip == ip - 1); + if (block->num > 0) + assert(cfg->blocks[block->num - 1]->end_ip == ip - 1); - for (fs_inst *inst = (fs_inst *)block->start; - inst != block->end->next; - inst = (fs_inst *)inst->next) { + struct block_data *bd = &block_data[block->num]; + foreach_inst_in_block(fs_inst, inst, block) { /* Set use[] for this instruction */ - for (unsigned int i = 0; i < 3; i++) { - if (inst->src[i].file != GRF) - continue; + for (unsigned int i = 0; i < inst->sources; i++) { + fs_reg reg = inst->src[i]; - int regs_read = 1; - /* We don't know how many components are read in a send-from-grf, - * so just assume "all of them." - */ - if (inst->is_send_from_grf()) - regs_read = v->virtual_grf_sizes[inst->src[i].reg]; - - for (int j = 0; j < regs_read; j++) { - int var = var_from_vgrf[inst->src[i].reg] + - inst->src[i].reg_offset + j; + if (reg.file != GRF) + continue; - if (!BITSET_TEST(bd[b].def, var)) - BITSET_SET(bd[b].use, var); + for (int j = 0; j < inst->regs_read(v, i); j++) { + setup_one_read(bd, inst, ip, reg); + reg.reg_offset++; } } + if (inst->reads_flag()) { + /* The vertical combination predicates read f0.0 and f0.1. */ + if (inst->predicate == BRW_PREDICATE_ALIGN1_ANYV || + inst->predicate == BRW_PREDICATE_ALIGN1_ALLV) { + assert(inst->flag_subreg == 0); + if (!BITSET_TEST(bd->flag_def, 1)) { + BITSET_SET(bd->flag_use, 1); + } + } + if (!BITSET_TEST(bd->flag_def, inst->flag_subreg)) { + BITSET_SET(bd->flag_use, inst->flag_subreg); + } + } - /* Check for unconditional writes to whole registers. These - * are the things that screen off preceding definitions of a - * variable, and thus qualify for being in def[]. - */ - if (inst->dst.file == GRF && !inst->is_partial_write()) { - int var = var_from_vgrf[inst->dst.reg] + inst->dst.reg_offset; + /* Set def[] for this instruction */ + if (inst->dst.file == GRF) { + fs_reg reg = inst->dst; for (int j = 0; j < inst->regs_written; j++) { - if (!BITSET_TEST(bd[b].use, var + j)) - BITSET_SET(bd[b].def, var + j); + setup_one_write(bd, inst, ip, reg); + reg.reg_offset++; } } + if (inst->writes_flag()) { + if (!BITSET_TEST(bd->flag_use, inst->flag_subreg)) { + BITSET_SET(bd->flag_def, inst->flag_subreg); + } + } ip++; } @@ -127,39 +204,79 @@ fs_live_variables::compute_live_variables() while (cont) { cont = false; - for (int b = 0; b < cfg->num_blocks; b++) { + foreach_block (block, cfg) { + struct block_data *bd = &block_data[block->num]; + /* Update livein */ 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; + BITSET_WORD new_livein = (bd->use[i] | + (bd->liveout[i] & + ~bd->def[i])); + if (new_livein & ~bd->livein[i]) { + bd->livein[i] |= new_livein; cont = true; } } + BITSET_WORD new_livein = (bd->flag_use[0] | + (bd->flag_liveout[0] & + ~bd->flag_def[0])); + if (new_livein & ~bd->flag_livein[0]) { + bd->flag_livein[0] |= new_livein; + cont = true; + } /* Update liveout */ - foreach_list(block_node, &cfg->blocks[b]->children) { - bblock_link *link = (bblock_link *)block_node; - bblock_t *block = link->block; + foreach_list_typed(bblock_link, child_link, link, &block->children) { + struct block_data *child_bd = &block_data[child_link->block->num]; for (int i = 0; i < bitset_words; i++) { - BITSET_WORD new_liveout = (bd[block->block_num].livein[i] & - ~bd[b].liveout[i]); + BITSET_WORD new_liveout = (child_bd->livein[i] & + ~bd->liveout[i]); if (new_liveout) { - bd[b].liveout[i] |= new_liveout; + bd->liveout[i] |= new_liveout; cont = true; } } + BITSET_WORD new_liveout = (child_bd->flag_livein[0] & + ~bd->flag_liveout[0]); + if (new_liveout) { + bd->flag_liveout[0] |= new_liveout; + cont = true; + } + } + } + } +} + +/** + * Extend the start/end ranges for each variable to account for the + * new information calculated from control flow. + */ +void +fs_live_variables::compute_start_end() +{ + foreach_block (block, cfg) { + struct block_data *bd = &block_data[block->num]; + + for (int i = 0; i < num_vars; i++) { + if (BITSET_TEST(bd->livein, i)) { + start[i] = MIN2(start[i], block->start_ip); + end[i] = MAX2(end[i], block->start_ip); } + + if (BITSET_TEST(bd->liveout, i)) { + start[i] = MIN2(start[i], block->end_ip); + end[i] = MAX2(end[i], block->end_ip); + } + } } } -fs_live_variables::fs_live_variables(fs_visitor *v, cfg_t *cfg) +fs_live_variables::fs_live_variables(fs_visitor *v, const cfg_t *cfg) : v(v), cfg(cfg) { - mem_ctx = ralloc_context(cfg->mem_ctx); + mem_ctx = ralloc_context(NULL); num_vgrfs = v->virtual_grf_count; num_vars = 0; @@ -176,18 +293,31 @@ fs_live_variables::fs_live_variables(fs_visitor *v, cfg_t *cfg) } } - bd = rzalloc_array(mem_ctx, struct block_data, cfg->num_blocks); + start = ralloc_array(mem_ctx, int, num_vars); + end = rzalloc_array(mem_ctx, int, num_vars); + for (int i = 0; i < num_vars; i++) { + start[i] = MAX_INSTRUCTION; + end[i] = -1; + } + + block_data= rzalloc_array(mem_ctx, struct block_data, cfg->num_blocks); bitset_words = BITSET_WORDS(num_vars); for (int i = 0; i < cfg->num_blocks; i++) { - 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); + block_data[i].def = rzalloc_array(mem_ctx, BITSET_WORD, bitset_words); + block_data[i].use = rzalloc_array(mem_ctx, BITSET_WORD, bitset_words); + block_data[i].livein = rzalloc_array(mem_ctx, BITSET_WORD, bitset_words); + block_data[i].liveout = rzalloc_array(mem_ctx, BITSET_WORD, bitset_words); + + block_data[i].flag_def[0] = 0; + block_data[i].flag_use[0] = 0; + block_data[i].flag_livein[0] = 0; + block_data[i].flag_liveout[0] = 0; } setup_def_use(); compute_live_variables(); + compute_start_end(); } fs_live_variables::~fs_live_variables() @@ -195,12 +325,11 @@ fs_live_variables::~fs_live_variables() ralloc_free(mem_ctx); } -#define MAX_INSTRUCTION (1 << 30) - void fs_visitor::invalidate_live_intervals() { - this->live_intervals_valid = false; + ralloc_free(live_intervals); + live_intervals = NULL; } /** @@ -212,100 +341,37 @@ fs_visitor::invalidate_live_intervals() void fs_visitor::calculate_live_intervals() { - if (this->live_intervals_valid) + if (this->live_intervals) return; int num_vgrfs = this->virtual_grf_count; - int *start = ralloc_array(mem_ctx, int, num_vgrfs); - int *end = ralloc_array(mem_ctx, int, num_vgrfs); ralloc_free(this->virtual_grf_start); ralloc_free(this->virtual_grf_end); - this->virtual_grf_start = start; - this->virtual_grf_end = end; + virtual_grf_start = ralloc_array(mem_ctx, int, num_vgrfs); + virtual_grf_end = ralloc_array(mem_ctx, int, num_vgrfs); for (int i = 0; i < num_vgrfs; i++) { - start[i] = MAX_INSTRUCTION; - end[i] = -1; + virtual_grf_start[i] = MAX_INSTRUCTION; + virtual_grf_end[i] = -1; } - /* Start by setting up the intervals with no knowledge of control - * flow. - */ - int ip = 0; - foreach_list(node, &this->instructions) { - fs_inst *inst = (fs_inst *)node; - - for (unsigned int i = 0; i < 3; i++) { - if (inst->src[i].file == GRF) { - int reg = inst->src[i].reg; - 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; + this->live_intervals = new(mem_ctx) fs_live_variables(this, cfg); - start[reg] = MIN2(start[reg], ip); - end[reg] = MAX2(end[reg], ip); - } - - ip++; - } - - /* Now, extend those intervals using our analysis of control flow. */ - cfg_t cfg(this); - fs_live_variables livevars(this, &cfg); - - for (int b = 0; b < cfg.num_blocks; b++) { - for (int i = 0; i < livevars.num_vars; i++) { - int vgrf = livevars.vgrf_from_var[i]; - if (BITSET_TEST(livevars.bd[b].livein, i)) { - start[vgrf] = MIN2(start[vgrf], cfg.blocks[b]->start_ip); - end[vgrf] = MAX2(end[vgrf], cfg.blocks[b]->start_ip); - } - - if (BITSET_TEST(livevars.bd[b].liveout, i)) { - start[vgrf] = MIN2(start[vgrf], cfg.blocks[b]->end_ip); - end[vgrf] = MAX2(end[vgrf], cfg.blocks[b]->end_ip); - } - } + /* Merge the per-component live ranges to whole VGRF live ranges. */ + for (int i = 0; i < live_intervals->num_vars; i++) { + int vgrf = live_intervals->vgrf_from_var[i]; + virtual_grf_start[vgrf] = MIN2(virtual_grf_start[vgrf], + live_intervals->start[i]); + virtual_grf_end[vgrf] = MAX2(virtual_grf_end[vgrf], + live_intervals->end[i]); } +} - this->live_intervals_valid = true; +bool +fs_live_variables::vars_interfere(int a, int b) +{ + return !(end[b] <= start[a] || + end[a] <= start[b]); } bool