fdo-mirrors/mesa
1
0
Fork 0
mirror of https://gitlab.freedesktop.org/mesa/mesa.git synced 2026-01-31 17:50:35 +01:00
Code Issues Projects Releases Packages Wiki Activity Actions 3

ir3/legalize: normalize nop state at block start

Browse source 
Now that we have the block's final cycle value available in its state,
we don't have to subtract it at the end of a block anymore, but we can
do it at the beginning when merging it into its successor state. This
will save us one iteration over all its ready slots.

Signed-off-by: Job Noorman <job@noorman.info>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/34108>
This commit is contained in:
Job Noorman 2025-05-09 11:26:05 +02:00 • committed by Marge Bot
parent 03ee7c7c0f
commit f7002802c3
1 changed files with 11 additions and 24 deletions
Show stats Download patch file Download diff file Expand all files Collapse all files

35
src/freedreno/ir3/ir3_legalize.c
View file

@ -370,19 +370,24 @@ legalize_block(struct ir3_legalize_ctx *ctx, struct ir3_block *block)
state->needs_ss_for_const |= pstate->needs_ss_for_const;
state->needs_sy_for_const |= pstate->needs_sy_for_const;
/* Our nop state is the max of the predecessor blocks */
/* Our nop state is the max of the predecessor blocks. The predecessor nop
* state contains the cycle offset from the start of its block when each
* register becomes ready. But successor blocks need the cycle offset from
* their start, which is the predecessor's block's end. Translate the
* cycle offset.
*/
for (unsigned i = 0; i < ARRAY_SIZE(state->pred_ready); i++)
state->pred_ready[i] = MAX2(state->pred_ready[i],
pstate->pred_ready[i]);
MAX2(pstate->pred_ready[i], pstate->cycle) - pstate->cycle);
for (unsigned i = 0; i < ARRAY_SIZE(state->alu_nop.full_ready); i++) {
state->alu_nop.full_ready[i] = MAX2(state->alu_nop.full_ready[i],
pstate->alu_nop.full_ready[i]);
MAX2(pstate->alu_nop.full_ready[i], pstate->cycle) - pstate->cycle);
state->alu_nop.half_ready[i] = MAX2(state->alu_nop.half_ready[i],
pstate->alu_nop.half_ready[i]);
MAX2(pstate->alu_nop.half_ready[i], pstate->cycle) - pstate->cycle);
state->non_alu_nop.full_ready[i] = MAX2(state->non_alu_nop.full_ready[i],
pstate->non_alu_nop.full_ready[i]);
MAX2(pstate->non_alu_nop.full_ready[i], pstate->cycle) - pstate->cycle);
state->non_alu_nop.half_ready[i] = MAX2(state->non_alu_nop.half_ready[i],
pstate->non_alu_nop.half_ready[i]);
MAX2(pstate->non_alu_nop.half_ready[i], pstate->cycle) - pstate->cycle);
}
}
@ -786,24 +791,6 @@ legalize_block(struct ir3_legalize_ctx *ctx, struct ir3_block *block)
list_add(&baryf->node, &block->instr_list);
}
/* Currently our nop state contains the cycle offset from the start of this
* block when each register becomes ready. But successor blocks need the
* cycle offset from their start, which is this block's end. Translate the
* cycle offset.
*/
for (unsigned i = 0; i < ARRAY_SIZE(state->pred_ready); i++)
state->pred_ready[i] = MAX2(state->pred_ready[i], state->cycle) - state->cycle;
for (unsigned i = 0; i < ARRAY_SIZE(state->alu_nop.full_ready); i++) {
state->alu_nop.full_ready[i] =
MAX2(state->alu_nop.full_ready[i], state->cycle) - state->cycle;
state->alu_nop.half_ready[i] =
MAX2(state->alu_nop.half_ready[i], state->cycle) - state->cycle;
state->non_alu_nop.full_ready[i] =
MAX2(state->non_alu_nop.full_ready[i], state->cycle) - state->cycle;
state->non_alu_nop.half_ready[i] =
MAX2(state->non_alu_nop.half_ready[i], state->cycle) - state->cycle;
}
bd->valid = true;
if (memcmp(&prev_state, state, sizeof(*state))) {