fix: flush frozen state for regular blocks

Author: HailToDodongo

src/rdpq/rdpq.c

@@ -614,6 +614,9 @@ extern inline void __rdpq_tracking_state_reset(rdpq_tracking_t *state);
 /** @brief Autosync engine: mark certain resources as in use */
 extern inline void __rdpq_autosync_use(uint32_t res);
 
+/** @brief Flush stale frozen-block DMEM state before a block CALL is enqueued */
+extern inline void __rdpq_block_run_prepare(rdpq_block_t *block);
+
 /** 
  * @brief Autosync engine: mark certain resources as being changed.
  * 

src/rdpq/rdpq_internal.h

@@ -234,6 +234,19 @@ inline void __rdpq_autosync_use(uint32_t res)
         __rdpq_frozen_flush_pending_mode();
     }
 }
+
+/**
+ * @brief Notify the rdpq engine that a rspq block is about to run, before its CALL command is enqueued.
+ *        This intern *may* flush out pending DMEM states form a frozen blocks
+ */
+inline void __rdpq_block_run_prepare(rdpq_block_t *block)
+{
+    if (__builtin_expect(
+      __rdpq_frozen_dmem_pending && !block->frozen && !rdpq_block_state.frozen, 0)
+    ) {
+        __rdpq_frozen_sync_dmem(RDPQ_WRITE_READS_RDP_STATE);
+    }
+}
 void __rdpq_autosync_change(uint32_t res);
 
 void __rdpq_write8(uint32_t cmd_id, uint32_t arg0, uint32_t arg1);

src/rspq/rspq.c

@@ -1430,6 +1430,10 @@ void rspq_block_run(rspq_block_t *block)
 
     assertf(block->nesting_level < block->min_ph_level, "Block nesting level overlaps with used placeholders");
 
+    // flush any pending DMEM state out (e.g. a frozen block ran before we now submit a normal block)
+    if (block->rdp_block)
+      __rdpq_block_run_prepare(block->rdp_block);
+
     // Write the CALL op. The second argument is the nesting level
     // which is used as stack slot in the RSP to save the current
     // pointer position.