vc4: Track the last block we emitted at the top level.

This will be used for delaying our VPM reads (which must be unconditional)
until just before they're used.

src/gallium/drivers/vc4/vc4_program.c

@@ -1956,11 +1956,12 @@ ntq_emit_if(struct vc4_compile *c, nir_if *if_stmt)
         qir_link_blocks(c->cur_block, after_block);
 
         qir_set_emit_block(c, after_block);
-        if (was_top_level)
+        if (was_top_level) {
                 c->execute = c->undef;
-        else
+                c->last_top_block = c->cur_block;
+        } else {
                 ntq_activate_execute_for_block(c);
-
+        }
 }
 
 static void
@@ -2084,10 +2085,12 @@ ntq_emit_loop(struct vc4_compile *c, nir_loop *loop)
         qir_link_blocks(c->cur_block, c->loop_break_block);
 
         qir_set_emit_block(c, c->loop_break_block);
-        if (was_top_level)
+        if (was_top_level) {
                 c->execute = c->undef;
-        else
+                c->last_top_block = c->cur_block;
+        } else {
                 ntq_activate_execute_for_block(c);
+        }
 
         c->loop_break_block = save_loop_break_block;
         c->loop_cont_block = save_loop_cont_block;

src/gallium/drivers/vc4/vc4_qir.c

@@ -646,6 +646,7 @@ qir_compile_init(void)
 
         list_inithead(&c->blocks);
         qir_set_emit_block(c, qir_new_block(c));
+        c->last_top_block = c->cur_block;
 
         c->output_position_index = -1;
         c->output_color_index = -1;

src/gallium/drivers/vc4/vc4_qir.h

@@ -506,6 +506,7 @@ struct vc4_compile {
         struct qblock *cur_block;
         struct qblock *loop_cont_block;
         struct qblock *loop_break_block;
+        struct qblock *last_top_block;
 
         struct list_head qpu_inst_list;