freedreno/ir3: Add intrinsics that map to LDLW/STLW

[mesa.git] / src / freedreno / ir3 / ir3_context.c

diff --git a/src/freedreno/ir3/ir3_context.c b/src/freedreno/ir3/ir3_context.c
index 99997427ec30db90399962919397ad6d8626bbb5..c389f750bd57596f06b4b09ed6a3972b238f9180 100644 (file)
--- a/src/freedreno/ir3/ir3_context.c
+++ b/src/freedreno/ir3/ir3_context.c
@@ -71,19 +71,31 @@ ir3_context_init(struct ir3_compiler *compiler,
         * creating duplicate variants..
         */
 
-       if (ir3_key_lowers_nir(&so->key)) {
-               nir_shader *s = nir_shader_clone(ctx, so->shader->nir);
-               ctx->s = ir3_optimize_nir(so->shader, s, &so->key);
-       } else {
-               /* fast-path for shader key that lowers nothing in NIR: */
-               ctx->s = nir_shader_clone(ctx, so->shader->nir);
-       }
+       ctx->s = nir_shader_clone(ctx, so->shader->nir);
+       if (ir3_key_lowers_nir(&so->key))
+               ir3_optimize_nir(so->shader, ctx->s, &so->key);
 
        /* this needs to be the last pass run, so do this here instead of
         * in ir3_optimize_nir():
         */
        NIR_PASS_V(ctx->s, nir_lower_bool_to_int32);
        NIR_PASS_V(ctx->s, nir_lower_locals_to_regs);
+
+       /* We want to lower nir_op_imul as late as possible, to catch also
+        * those generated by earlier passes (e.g, nir_lower_locals_to_regs).
+        * However, we want a final swing of a few passes to have a chance
+        * at optimizing the result.
+        */
+       bool progress = false;
+       NIR_PASS(progress, ctx->s, ir3_nir_lower_imul);
+       if (progress) {
+               NIR_PASS_V(ctx->s, nir_opt_algebraic);
+               NIR_PASS_V(ctx->s, nir_opt_copy_prop_vars);
+               NIR_PASS_V(ctx->s, nir_opt_dead_write_vars);
+               NIR_PASS_V(ctx->s, nir_opt_dce);
+               NIR_PASS_V(ctx->s, nir_opt_constant_folding);
+       }
+
        NIR_PASS_V(ctx->s, nir_convert_from_ssa, true);
 
        if (ir3_shader_debug & IR3_DBG_DISASM) {
@@ -518,8 +530,11 @@ ir3_create_array_store(struct ir3_context *ctx, struct ir3_array *arr, int n,
 
        /* if not relative store, don't create an extra mov, since that
         * ends up being difficult for cp to remove.
+        *
+        * Also, don't skip the mov if the src is meta (like fanout/split),
+        * since that creates a situation that RA can't really handle properly.
         */
-       if (!address) {
+       if (!address && !is_meta(src)) {
                dst = src->regs[0];
 
                src->barrier_class |= IR3_BARRIER_ARRAY_W;