diff options
author | Eric Anholt <[email protected]> | 2012-06-05 11:23:09 -0700 |
---|---|---|
committer | Eric Anholt <[email protected]> | 2013-10-10 15:54:15 -0700 |
commit | 398656d97e8e554623f43eca31dc6a0cbf22979d (patch) | |
tree | e9de04e315fa546cd4e158aa5cc1aff5f95e33fb /src/mesa | |
parent | 097bf101c33ee14e4e163523dc4e4e0fbca9f051 (diff) |
i965/fs: Track live variable ranges on a per-channel level.
This is the information we'll actually use to replace the
virtual_grf_start[]/end[] arrays.
No change in shader-db.
v2 (Kenneth Graunke): Rebase; minor comment updates.
Reviewed-by: Kenneth Graunke <[email protected]>
Diffstat (limited to 'src/mesa')
-rw-r--r-- | src/mesa/drivers/dri/i965/brw_fs_live_variables.cpp | 150 | ||||
-rw-r--r-- | src/mesa/drivers/dri/i965/brw_fs_live_variables.h | 9 |
2 files changed, 85 insertions, 74 deletions
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 57b4d249250..a07565cecd8 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. @@ -56,6 +58,41 @@ fs_live_variables::setup_one_read(bblock_t *block, fs_inst *inst, { int var = var_from_vgrf[reg.reg] + reg.reg_offset; + /* 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. + */ + int end_ip = ip; + if (v->dispatch_width == 16 && (reg.smear != -1 || + (v->pixel_x.reg == reg.reg || + v->pixel_y.reg == reg.reg))) { + 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. @@ -70,6 +107,9 @@ fs_live_variables::setup_one_write(bblock_t *block, fs_inst *inst, { int var = var_from_vgrf[reg.reg] + reg.reg_offset; + 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). */ @@ -181,6 +221,29 @@ fs_live_variables::compute_live_variables() } } +/** + * 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() +{ + for (int b = 0; b < cfg->num_blocks; b++) { + for (int i = 0; i < num_vars; i++) { + if (BITSET_TEST(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 (BITSET_TEST(bd[b].liveout, i)) { + start[i] = MIN2(start[i], cfg->blocks[b]->end_ip); + end[i] = MAX2(end[i], cfg->blocks[b]->end_ip); + } + + } + } +} + fs_live_variables::fs_live_variables(fs_visitor *v, cfg_t *cfg) : v(v), cfg(cfg) { @@ -201,6 +264,13 @@ fs_live_variables::fs_live_variables(fs_visitor *v, cfg_t *cfg) } } + 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; + } + bd = rzalloc_array(mem_ctx, struct block_data, cfg->num_blocks); bitset_words = BITSET_WORDS(num_vars); @@ -213,6 +283,7 @@ fs_live_variables::fs_live_variables(fs_visitor *v, cfg_t *cfg) setup_def_use(); compute_live_variables(); + compute_start_end(); } fs_live_variables::~fs_live_variables() @@ -220,8 +291,6 @@ fs_live_variables::~fs_live_variables() ralloc_free(mem_ctx); } -#define MAX_INSTRUCTION (1 << 30) - void fs_visitor::invalidate_live_intervals() { @@ -253,81 +322,14 @@ fs_visitor::calculate_live_intervals() 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; - - 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 < livevars.num_vars; i++) { + int vgrf = livevars.vgrf_from_var[i]; + start[vgrf] = MIN2(start[vgrf], livevars.start[i]); + end[vgrf] = MAX2(end[vgrf], livevars.end[i]); } this->live_intervals_valid = true; diff --git a/src/mesa/drivers/dri/i965/brw_fs_live_variables.h b/src/mesa/drivers/dri/i965/brw_fs_live_variables.h index a2abde24390..694ad9edc69 100644 --- a/src/mesa/drivers/dri/i965/brw_fs_live_variables.h +++ b/src/mesa/drivers/dri/i965/brw_fs_live_variables.h @@ -62,6 +62,7 @@ public: void setup_one_read(bblock_t *block, fs_inst *inst, int ip, fs_reg reg); void setup_one_write(bblock_t *block, fs_inst *inst, int ip, fs_reg reg); void compute_live_variables(); + void compute_start_end(); fs_visitor *v; cfg_t *cfg; @@ -82,6 +83,14 @@ public: int num_vgrfs; int bitset_words; + /** @{ + * Final computed live ranges for each var (each component of each virtual + * GRF). + */ + int *start; + int *end; + /** @} */ + /** Per-basic-block information on live variables */ struct block_data *bd; }; |