vc4: Enable the nir_opt_algebraic_late() pass.

Upcoming changes to sub optimization will make this pass required.  Over
the course of that series, we see uniforms +.46%, instructions -.24%
(seems like a fine tradeoff -- uniforms are 1/2 the size of instructions
as far as cache occupancy)

Reviewed-by: Daniel Schürmann <daniel@schuermann.dev>
Reviewed-by: Connor Abbott <cwabbott0@gmail.com>

diff --git a/src/gallium/drivers/vc4/vc4_program.c b/src/gallium/drivers/vc4/vc4_program.c
index e5f7aa31b0ebbffe1c6500ded626f4bceaf151a1..7d88444c6738962215fe16f9e47bf22b31d3f402 100644 (file)
--- a/src/gallium/drivers/vc4/vc4_program.c
+++ b/src/gallium/drivers/vc4/vc4_program.c
@@ -2335,6 +2335,21 @@ vc4_shader_ntq(struct vc4_context *vc4, enum qstage stage,
 
         vc4_optimize_nir(c->s);
 
+        /* Do late algebraic optimization to turn add(a, neg(b)) back into
+         * subs, then the mandatory cleanup after algebraic.  Note that it may
+         * produce fnegs, and if so then we need to keep running to squash
+         * fneg(fneg(a)).
+         */
+        bool more_late_algebraic = true;
+        while (more_late_algebraic) {
+                more_late_algebraic = false;
+                NIR_PASS(more_late_algebraic, c->s, nir_opt_algebraic_late);
+                NIR_PASS_V(c->s, nir_opt_constant_folding);
+                NIR_PASS_V(c->s, nir_copy_prop);
+                NIR_PASS_V(c->s, nir_opt_dce);
+                NIR_PASS_V(c->s, nir_opt_cse);
+        }
+
         NIR_PASS_V(c->s, nir_lower_bool_to_int32);
 
         NIR_PASS_V(c->s, nir_convert_from_ssa, true);