[Mesa-dev] [PATCH 3/3] glsl_to_tgsi: Use mesa register allocator
Tom Stellard
thomas.stellard at amd.com
Fri Jan 6 11:28:14 PST 2012
On Fri, 2012-01-06 at 20:15 +0100, Vincent Lejeune wrote:
> ---
> src/mesa/state_tracker/st_glsl_to_tgsi.cpp | 156 +++++++++++++--------------
> 1 files changed, 75 insertions(+), 81 deletions(-)
>
> diff --git a/src/mesa/state_tracker/st_glsl_to_tgsi.cpp b/src/mesa/state_tracker/st_glsl_to_tgsi.cpp
> index e0013f1..2f752d6 100644
> --- a/src/mesa/state_tracker/st_glsl_to_tgsi.cpp
> +++ b/src/mesa/state_tracker/st_glsl_to_tgsi.cpp
> @@ -55,6 +55,7 @@ extern "C" {
> #include "program/program.h"
> #include "program/prog_parameter.h"
> #include "program/sampler.h"
> +#include "program/register_allocate.h"
>
> #include "pipe/p_compiler.h"
> #include "pipe/p_context.h"
> @@ -834,7 +835,6 @@ public:
> void remove_output_reads(gl_register_file type);
> void simplify_cmp(void);
>
> - void rename_temp_register(int index, int new_index);
> int get_first_temp_read(int index);
> int get_first_temp_write(int index);
> int get_last_temp_read(int index);
> @@ -843,8 +843,7 @@ public:
> void copy_propagate(void);
> void eliminate_dead_code(void);
> int eliminate_dead_code_advanced(void);
> - void merge_registers(void);
> - void renumber_registers(void);
> + void regalloc(void);
> void renumber_temp_regs(unsigned*);
>
> void *mem_ctx;
> @@ -3419,27 +3418,6 @@ glsl_to_tgsi_visitor::simplify_cmp(void)
> delete [] tempWrites;
> }
>
> -/* Replaces all references to a temporary register index with another index. */
> -void
> -glsl_to_tgsi_visitor::rename_temp_register(int index, int new_index)
> -{
> - foreach_iter(exec_list_iterator, iter, this->instructions) {
> - glsl_to_tgsi_instruction *inst = (glsl_to_tgsi_instruction *)iter.get();
> - unsigned j;
> -
> - for (j=0; j < num_inst_src_regs(inst->op); j++) {
> - if (inst->src[j].file == PROGRAM_TEMPORARY &&
> - inst->src[j].index == index) {
> - inst->src[j].index = new_index;
> - }
> - }
> -
> - if (inst->dst.file == PROGRAM_TEMPORARY && inst->dst.index == index) {
> - inst->dst.index = new_index;
> - }
> - }
> -}
> -
> int
> glsl_to_tgsi_visitor::get_first_temp_read(int index)
> {
> @@ -3943,73 +3921,92 @@ glsl_to_tgsi_visitor::eliminate_dead_code_advanced(void)
> return removed;
> }
>
> -/* Merges temporary registers together where possible to reduce the number of
> +struct interval {
> + int first_line;
> + int last_line;
> +};
> +
> +static bool
> +overlap(const interval &a, const interval &b)
> +{
> + return MIN2(a.last_line, b.last_line) > MAX2(a.first_line, b.first_line);
> +}
> +
> +/* Merges temporary registers together where possible to reduce the number of
> * registers needed to run a program.
> - *
> - * Produces optimal code only after copy propagation and dead code elimination
> + *
> + * Produces optimal code only after copy propagation and dead code elimination
> * have been run. */
> void
> -glsl_to_tgsi_visitor::merge_registers(void)
> +glsl_to_tgsi_visitor::regalloc(void)
> {
> - int *last_reads = rzalloc_array(mem_ctx, int, this->next_temp);
> - int *first_writes = rzalloc_array(mem_ctx, int, this->next_temp);
> + unsigned total_temps = store.temp_amount();
> + unsigned first_non_array_temp = store.temp_array_amount() + 1;
> + unsigned allocable_regs = total_temps - first_non_array_temp;
> + struct interval *live_interval = rzalloc_array(mem_ctx,struct interval,allocable_regs);
> int i, j;
> -
> +
> /* Read the indices of the last read and first write to each temp register
> - * into an array so that we don't have to traverse the instruction list as
> + * into an array so that we don't have to traverse the instruction list as
> * much. */
> - for (i=0; i < this->next_temp; i++) {
> - last_reads[i] = get_last_temp_read(i);
> - first_writes[i] = get_first_temp_write(i);
> + for (i=0; i < allocable_regs; i++) {
> + live_interval[i].last_line = get_last_temp_read(i + first_non_array_temp);
> + live_interval[i].first_line = get_first_temp_write(i + first_non_array_temp);
> }
> -
> - /* Start looking for registers with non-overlapping usages that can be
> - * merged together. */
> - for (i=0; i < this->next_temp; i++) {
> - /* Don't touch unused registers. */
> - if (last_reads[i] < 0 || first_writes[i] < 0) continue;
> -
> - for (j=0; j < this->next_temp; j++) {
> - /* Don't touch unused registers. */
> - if (last_reads[j] < 0 || first_writes[j] < 0) continue;
> -
> - /* We can merge the two registers if the first write to j is after or
> - * in the same instruction as the last read from i. Note that the
> - * register at index i will always be used earlier or at the same time
> - * as the register at index j. */
> - if (first_writes[i] <= first_writes[j] &&
> - last_reads[i] <= first_writes[j])
> - {
> - rename_temp_register(j, i); /* Replace all references to j with i.*/
> -
> - /* Update the first_writes and last_reads arrays with the new
> - * values for the merged register index, and mark the newly unused
> - * register index as such. */
> - last_reads[i] = last_reads[j];
> - first_writes[j] = -1;
> - last_reads[j] = -1;
> +
> + struct ra_regs *regs = ra_alloc_reg_set(MAX_TEMPS);
> + unsigned temp_class = ra_alloc_reg_class(regs);
> +
> + for (i = 0; i < + 32; i++) {
> + ra_class_add_reg(regs,temp_class,i);
> + }
> +
i < + 32
This looks like a typo.
> + ra_set_finalize(regs);
> +
> + struct ra_graph *graph = ra_alloc_interference_graph(regs,total_temps - first_non_array_temp + 1);
> +
> + for (i = 0; i < first_non_array_temp; i++) {
> + ra_set_node_class(graph,i,temp_class);
> + }
> +
> +
> + for (i = 0; i < allocable_regs; i++) {
> + for (j = i + 1; j < allocable_regs; j++) {
> + if (overlap(live_interval[i],live_interval[j]))
> + ra_add_node_interference(graph,i,j);
> + }
> + }
> +
> + if (!ra_allocate_no_spills(graph)) {
> + printf("UNABLE TO ALLOCATE REGS\n");
> + }
> +
> +
> + foreach_iter(exec_list_iterator, iter, this->instructions) {
> + glsl_to_tgsi_instruction *inst = (glsl_to_tgsi_instruction *)iter.get();
> + unsigned j;
> +
> + for (j=0; j < num_inst_src_regs(inst->op); j++) {
> + if (inst->src[j].file == PROGRAM_TEMPORARY) {
> + unsigned index = inst->src[j].index;
> + if (index >= first_non_array_temp)
> + inst->src[j].index = ra_get_node_reg(graph, index - first_non_array_temp) + first_non_array_temp;
> }
> }
> +
> + if (inst->dst.file == PROGRAM_TEMPORARY) {
> + unsigned index = inst->dst.index;
> + if (index >= first_non_array_temp)
> + inst->dst.index = ra_get_node_reg(graph, index - first_non_array_temp) + first_non_array_temp;
> + }
> }
> -
> - ralloc_free(last_reads);
> - ralloc_free(first_writes);
> +
> + ralloc_free(graph);
> + ralloc_free(regs);
> }
>
> -/* Reassign indices to temporary registers by reusing unused indices created
> - * by optimization passes. */
> -void
> -glsl_to_tgsi_visitor::renumber_registers(void)
> void glsl_to_tgsi_visitor::renumber_temp_regs(unsigned *reindex_table)
> {
> - int i = 0;
> - int new_index = 0;
> -
> - for (i=0; i < this->next_temp; i++) {
> - if (get_first_temp_read(i) < 0) continue;
> - if (i != new_index)
> - rename_temp_register(i, new_index);
> - new_index++;
> foreach_iter(exec_list_iterator, iter, this->instructions) {
> glsl_to_tgsi_instruction *inst = (glsl_to_tgsi_instruction *)iter.get();
> unsigned j;
> @@ -4026,8 +4023,6 @@ void glsl_to_tgsi_visitor::renumber_temp_regs(unsigned *reindex_table)
> inst->dst.index = reindex_table[index];
> }
> }
> -
> - this->next_temp = new_index;
> }
>
> /**
> @@ -5286,10 +5281,9 @@ get_mesa_program(struct gl_context *ctx,
> */
> if (!v->indirect_addr_temps) {
> v->eliminate_dead_code();
> - v->merge_registers();
> - v->renumber_registers();
> }
> -
> + v->regalloc();
> +
> /* Write the END instruction. */
> v->emit(NULL, TGSI_OPCODE_END);
>
More information about the mesa-dev
mailing list