ir3: use opt_shrink_vectors
Some checks are pending
macOS-CI / macOS-CI (dri) (push) Waiting to run
macOS-CI / macOS-CI (xlib) (push) Waiting to run

This is useful to remove unused components from IO loads. This is not
only helpful for reducing the size of memory loads, but also for reducing
register pressure (as we need smaller vector registers).

Totals from 55567 (33.76% of 164575) affected shaders:
MaxWaves: 665780 -> 666690 (+0.14%); split: +0.15%, -0.01%
Instrs: 30850106 -> 30825516 (-0.08%); split: -0.33%, +0.25%
CodeSize: 62502952 -> 62336580 (-0.27%); split: -0.42%, +0.15%
NOPs: 5468972 -> 5463654 (-0.10%); split: -1.39%, +1.30%
MOVs: 1078607 -> 1060627 (-1.67%); split: -3.48%, +1.81%
Full: 807907 -> 805134 (-0.34%); split: -0.49%, +0.15%
(ss): 755846 -> 747963 (-1.04%); split: -2.51%, +1.46%
(sy): 367032 -> 363947 (-0.84%); split: -1.88%, +1.04%
(ss)-stall: 2907874 -> 2900183 (-0.26%); split: -1.97%, +1.71%
(sy)-stall: 10812082 -> 10599944 (-1.96%); split: -3.43%, +1.47%
STPs: 23062 -> 22980 (-0.36%)
LDPs: 35076 -> 32286 (-7.95%)
Preamble Instrs: 6668422 -> 6612037 (-0.85%); split: -1.35%, +0.51%
Early Preamble: 36055 -> 36169 (+0.32%); split: +0.32%, -0.01%
Cat0: 6015654 -> 6009409 (-0.10%); split: -1.29%, +1.18%
Cat1: 1631110 -> 1610537 (-1.26%); split: -2.54%, +1.28%
Cat2: 11783599 -> 11783634 (+0.00%); split: -0.02%, +0.02%
Cat3: 8198147 -> 8198526 (+0.00%); split: -0.00%, +0.01%
Cat5: 968517 -> 967899 (-0.06%); split: -0.06%, +0.00%
Cat6: 349200 -> 351795 (+0.74%); split: -0.01%, +0.75%
Cat7: 1009072 -> 1008909 (-0.02%); split: -0.58%, +0.56%

Signed-off-by: Job Noorman <jnoorman@igalia.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/34600>
This commit is contained in:
Job Noorman 2025-04-18 09:38:38 +02:00 committed by Marge Bot
parent f269c7b3b5
commit e7c6037d12

View file

@ -363,6 +363,9 @@ ir3_optimize_loop(struct ir3_compiler *compiler,
progress |= OPT(s, nir_lower_bit_size, ir3_lower_bit_size, NULL);
progress |= OPT(s, nir_opt_constant_folding);
/* Remove unused components from IO loads. */
progress |= OPT(s, nir_opt_shrink_vectors, true);
const nir_opt_offsets_options offset_options = {
/* How large an offset we can encode in the instr's immediate field.
*/