danog/ir
1
0
Fork 0
mirror of https://github.com/danog/ir.git synced 2025-01-22 05:31:32 +01:00
Code Issues Packages Projects Releases Wiki Activity

Improve GCM to schedule floating nodes that depends only on constants

Browse Source 
(e.g. COPY(CONST)) to the last common ancestor.

Previously these nodes went to the first block.
This commit is contained in:
Dmitry Stogov 2023-02-15 15:18:42 +03:00
parent 1d7ab16c2a
commit d07a2db592
1 changed files with 74 additions and 3 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

77
ir_gcm.c
View File

@ -11,11 +11,12 @@
#include "ir.h"
#include "ir_private.h"
static void ir_gcm_schedule_early(ir_ctx *ctx, uint32_t *_blocks, ir_ref ref)
static void ir_gcm_schedule_early(ir_ctx *ctx, uint32_t *_blocks, ir_ref ref, ir_list *queue_rest)
{
ir_ref n, *p, input;
ir_insn *insn;
uint32_t dom_depth, b;
bool reschedule_late = 1;
insn = &ctx->ir_base[ref];
@ -30,15 +31,29 @@ static void ir_gcm_schedule_early(ir_ctx *ctx, uint32_t *_blocks, ir_ref ref)
input = *p;
if (input > 0) {
if (_blocks[input] == 0) {
ir_gcm_schedule_early(ctx, _blocks, input);
ir_gcm_schedule_early(ctx, _blocks, input, queue_rest);
}
b = _blocks[input];
if (dom_depth < ctx->cfg_blocks[b].dom_depth) {
dom_depth = ctx->cfg_blocks[b].dom_depth;
_blocks[ref] = b;
}
reschedule_late = 0;
}
}
if (UNEXPECTED(reschedule_late)) {
/* Floating nodes that doesn't sepend on other nodes
* (e.g. only on constants), has to be scheduled to the
* last common ancestor. Otherwise they always goes to the
* first block.
*
* TODO:
* It's possible to reuse ir_gcm_schedule_late() and move
* these nodes out of the loops, but then we mgiht need
* to rematerialize them at proper place(s).
*/
ir_list_push_unchecked(queue_rest, ref);
}
}
/* Last Common Ancestor */
@ -119,6 +134,51 @@ static void ir_gcm_schedule_late(ir_ctx *ctx, uint32_t *_blocks, ir_bitset visit
}
}
static void ir_gcm_schedule_rest(ir_ctx *ctx, uint32_t *_blocks, ir_bitset visited, ir_ref ref)
{
ir_ref n, *p, use;
ir_insn *insn;
ir_bitset_incl(visited, ref);
n = ctx->use_lists[ref].count;
if (n) {
uint32_t lca, b;
insn = &ctx->ir_base[ref];
IR_ASSERT(insn->op != IR_PARAM && insn->op != IR_VAR);
IR_ASSERT(insn->op != IR_PHI && insn->op != IR_PI);
lca = 0;
for (p = &ctx->use_edges[ctx->use_lists[ref].refs]; n > 0; p++, n--) {
use = *p;
b = _blocks[use];
if (!b) {
continue;
} else if (!ir_bitset_in(visited, use)) {
ir_gcm_schedule_late(ctx, _blocks, visited, use);
b = _blocks[use];
IR_ASSERT(b != 0);
}
insn = &ctx->ir_base[use];
if (insn->op == IR_PHI) {
ir_ref *p = insn->ops + 2; /* PHI data inputs */
ir_ref *q = ctx->ir_base[insn->op1].ops + 1; /* MERGE inputs */
while (*p != ref) {
p++;
q++;
}
b = _blocks[*q];
IR_ASSERT(b);
}
lca = !lca ? b : ir_gcm_find_lca(ctx, lca, b);
}
IR_ASSERT(lca != 0 && "No Common Antecessor");
b = lca;
_blocks[ref] = b;
}
}
int ir_gcm(ir_ctx *ctx)
{
ir_ref k, n, *p, ref;
@ -126,6 +186,7 @@ int ir_gcm(ir_ctx *ctx)
ir_block *bb;
ir_list queue_early;
ir_list queue_late;
ir_list queue_rest;
uint32_t *_blocks, b;
ir_insn *insn, *use_insn;
ir_use_list *use_list;
@ -238,6 +299,8 @@ int ir_gcm(ir_ctx *ctx)
}
}
ir_list_init(&queue_rest, ctx->insns_count);
n = ir_list_len(&queue_early);
while (n > 0) {
n--;
@ -247,7 +310,7 @@ int ir_gcm(ir_ctx *ctx)
for (p = insn->ops + 2; k > 0; p++, k--) {
ref = *p;
if (ref > 0 && _blocks[ref] == 0) {
ir_gcm_schedule_early(ctx, _blocks, ref);
ir_gcm_schedule_early(ctx, _blocks, ref, &queue_rest);
}
}
}
@ -275,9 +338,17 @@ int ir_gcm(ir_ctx *ctx)
}
}
n = ir_list_len(&queue_rest);
while (n > 0) {
n--;
ref = ir_list_at(&queue_rest, n);
ir_gcm_schedule_rest(ctx, _blocks, visited, ref);
}
ir_mem_free(visited);
ir_list_free(&queue_early);
ir_list_free(&queue_late);
ir_list_free(&queue_rest);
#ifdef IR_DEBUG
if (ctx->flags & IR_DEBUG_GCM) {